key: cord-0021100-mr2cgsu3 authors: nan title: Book of Abstracts ESMRMB 2021 Online 38th Annual Scientific Meeting 7–9 October 2021 date: 2021-09-18 journal: MAGMA DOI: 10.1007/s10334-021-00947-8 sha: 0914132673dfad527a86d527f0455be0bb75783b doc_id: 21100 cord_uid: mr2cgsu3 nan S1.O1. DW-and DCE-MRI and super-resolution ultrasound imaging for monitoring breast tumour response to radiotherapy S1.O3. Investigation of left-atrial flows using a 3D radial based self-gated respiratory motion corrected 4D Flow MRI sequence Introduction: Atrial fibrillation (AF) is a strong risk factor for embolic stroke likely due to prothrombotic alterations in left atrium (LA) blood flow. Four-dimensional velocity mapping MRI (4D Flow) allows quantification of LA and pulmonary veins (PV) flows. However, in the context of AF, standard 4D Flow leads to quantification errors from the inherent averaging of MRI [1] . Our goal was to develop a respiratory motion corrected free-running 3D radial based 4D Fow sequence and to investigate the AF related variability of LA blood flows. Methods: We implemented a free-running interleaved 3D radial velocity mapping sequence on a 1.5 T Philips Ingenia (Philips, Best, The Netherlands), based on a spiral phyllotaxis pattern [2] . Imaging was performed in 6 paroxysmal AF patients, 2 permanent AF, and 5 age matched volunteers as follows: TE/TR 3.5/7.5 ms, FA = 6°, VENC = 70-96 cm/s, 2.2-2.8 mm isotropic voxel for an acquisition time of 8-12 min. The respiratory self-gated signal was obtained from the k-space center as previously proposed [3] , and used to bin the data in 5 respiratory phases. The 3D translation motion of the heart between end-expiration and all other phases was estimated and used to correct the phase of the k-space data. Finally, the corrected data was binned into cardiac phases using the ECG signal and a fixed temporal resolution. This strategy reduces the velocity errors related to averaging over variable RR durations. Images were reconstructed offline using a compressed sensing algorithm implemented in Matlab (The Mathworks, Inc, Natick, MA). 4D Flow data analysis was performed with PIE Medical (The Netherlands). Results: Qualitatively, respiratory motion correction improved the sharpness of the PVs and modified slightly the flow pattern in the LA (Fig. 1) . Streamline visualization demonstrated significant differences in flow patterns, specifically a physiological vortex for healthy subjects, and absence of vortex and lower overall LA velocities for AF patients (Fig. 2 ). Discussion: We present here preliminary results demonstrating the feasibility of intra-atrial flow assessment using a 3D radial based 4D Flow MRI with respiratory motion correction. The initial results will be completed by an ongoing quantitative analysis of the LA flows (vorticity, vortex size). Particular interest will be paid to the impact of respiratory motion correction on these quantitative descriptors. LA flows are related to the entire cardiac circulation. Thus an added advantage of the 3D radial 4D Flow is the capability to measure the hemodynamic parameters of the entire heart in the same amount of time as a standard Cartesian 4D Flow covering only the LA, albeit with stronger under-sampling. Funding: ANR-18-CE19-0025-01 References: 1. Markl M, et al. (2016) Int J Cardiovasc Imaging 32:807-815. 2. Ma LE, et al. (2020) S1 .O4. High temporal-resolution dynamic MRI for the assessment of brown adipose tissue metabolism during mild-cold exposure in young healthy adults Introduction: Brown adipose tissue (BAT) is considered as a potential therapeutic target against cardiometabolic diseases 1 . Activated BAT combusts intracellular fatty acids leading to a reduction in fat fraction (FF) 2 . Both cold exposure and pharmacological stimuli can activate BAT, but the short-term dynamics of BAT activation are not fully explored. Therefore, we aimed to develop a MRI protocol for assessing supraclavicular BAT (scBAT) during cold exposure, with a high temporal resolution and also including non-rigid image registration to reduce variability due to respiratory motion. Methods: Five healthy volunteers (25.0 ± 3.0 years; BMI 21.8 ± 2.4 kg/m 2 ) underwent a standardized cooling procedure for BAT activation using a water-circulating blanket inside the MRI suite. After 10 min at thermoneutrality (32°C), the subject was moved into the scanner: after a further 10 min at thermoneutrality the temperature was set to 18°C for one hour. Subjects reported thermal perceptions at thermoneutrality and every 15 min during cooling using a thermal perception scale ranging from thermoneutral (1) to extreme cold (10) . Images were acquired at 3 T using a 16-channel array coil and a 3D gradient-echo 12-point Dixon sequence (TR = 12 ms, TE 1 = 1.12 ms, DTE = 0.87 ms, FA = 3°, 2.1 mm isotropic resolution, FOV of 400 9 229 9 134 mm3, a breath-hold duration of 16 s, mDIXON QUANT scanner reconstructions). The acquisition time per scan was 1.03 min to yield a total of 70 scans. For analysis, we co-registered the first-echo magnitude images of each dynamic to the first thermoneutral scan (reference scan) using Elastix 3 (Fig. 1 ). The FF of the scBAT depot was obtained by coarsely delineating the depot on the reference scan, and voxels below 30% FF were excluded in the reference scan and in every registered dynamic ( Fig. 1) 4 . The FF of the ROIs in the trapezius muscle was used as a reference. FF differences were calculated by subtracting the average FF of each dynamic from the average FF of the 10 thermoneutral scans. To determine the variability, we fitted a linear function through all scBAT FF dynamics and computed the mean squared error (MSE) of residuals. Results: All subjects were able to adhere to the protocol. Reported thermal scores and MSE values are provided in Table 1 . Averaged FF differences are shown for scBAT and skeletal muscle in Fig. 2A , B. We found a gradual scBAT FF decrease during cooling, whereas no changes were observed in muscle. We found an overall MSE of 0.09 ± 0.02% along the scBAT FF dynamics (Table 1) . Introduction: Magnetic resonance fingerprinting (MRF) is a novel MR data acquisition paradigm that enables simultaneous estimation of multiple quantitative MRI (qMRI) parameters, such as T1 and T2 relaxation times. Moreover, MRF is claimed to provide higher robustness to noise and be less prone to acquisition errors compared to traditional qMRI sequences 1 . However, the repeatability of MRF based T1 and T2 maps still needs to be determined for various clinical applications, such as qMRI of the knee cartilage. The aim of this study was to assess the intra-scanner variability of a MRF sequence 2 as a fast multiparametric quantitative imaging technique for assessment of femoral and tibial knee articular cartilage (AC) sub-regions. Methods: Nine Nickel-doped agarose gel phantoms 3 with T1 (300-1000 ms) and T2 (25-100 ms) within AC tissue range (Fig. 1 ) and four volunteer knee-subjects (two females, age = 30 ± 5 years) underwent 3 T-MRI (Siemens MAGNETOM Vida, 18-channel Tx/ Rx knee coil) using a recently proposed 2D-MRF sequence (TR/ TE = 7.5/3.5 ms, FA = 60°, 0.6 9 0.6 9 3.0 mm 3 , 6 radial spokes each consisting of 1000 time points) 2 . To assess the intra-scanner repeatability of the MRF sequence, the measurements in phantom and in vivo were repeated three times, each after repositioning. Femoral and tibial cartilage segmentation into 16 sub-regions ( Fig. 2A ) was performed manually for single slices from both lateral and medial knee compartments. Mean values of T1 and T2 relaxation times were calculated for all phantoms and sub-regions. Intra-scanner repeatability was quantified using the root mean square coefficient of variation (RMS-CV). Results: The results demonstrated excellent repeatability of MRF in phantom with RMS-CVs of 0.7% and 1.8% for T1 and T2 relaxation times, respectively. For in vivo analysis, the average intra-scanner RMS-CVs for lateral and medial compartments were * 3% and 2% for T1, and * 8% and * 8% for T2, respectively (Fig. 3 ). S1.O6. Paramagnetic contrast enhancement in MRI imaging of liver using an original hepatotropic high-affinity complex-2-(2-carboxymethyl-(4-hexa-decyloxyphenylcarbamoyl-methyl))-aminoethyl-(4-hexadecyloxyphenyl-carbamoyl methyl)-aminoacetic acid with Manganese(II)-GDOF-Mn-DTPA Introduction: Design of new organ-specific paramagnetic contrast agents for routine employment in clinical MRI studies is highly necessary because of very little number of such agents currently in use, in particular in liver studies. Purpose: There was studied the dependence of specific contrast properties of the paramagnetic complex 2-(2-carboxymethyl-(4-hexadecyloxyphenyl-carbamoyl-methyl))-aminoethyl-(4-hexadecyl-oxyphenyl-carbamoyl methyl)-aminoacetic acid ( Fig. 1 ) labelled with manganese-GDOF-Mn-DTPA (Fig. 2 )-in magnetic resonance imaging from the administered dose of the drug. The GDOF-Mn-DTPA was obtained and purified by original synthesis technology in the N.M.Kizhner research Center of the National Research Tomsk Polytechnic University. The dynamics of changes in the liver contrast ratio over time at different dosages of the contrast compound GDOF-Mn-DTPA was evaluated, and changes in the T1 relaxation time of liver and kidney tissue of laboratory animals (Wistar rats, more than 300 g body weight) at different dosages of GDOF-Mn-DTPA were calculated. Results: Visual analysis of contrast-enhanced MRI scans with GDOF-Mn-DTPA already at a dose of 0.025 mmol/kg reliably visualized the accumulation of paramagnet in the liver (Fig. 3) , while further concentration of the drug in the bile ducts of animals was noted, with the actual absence of visually detectable kidney contrast. When evaluating the T1 relaxation time for the liver and kidneys, a persistent decrease in the T1 time for liver tissue was obtained for doses of 0.1, 0.05 and 0.025 mmol/kg, in particular for 0.025 mmol/ kg from the initial 760 (747-755) ms to 488 (474-505) ms (p \ 0.02). On the contrary, the obtained values of T1 relaxation time for kidney tissue showed no significant accumulation of the paramagnetic contrast compound GDOF-Mn-DTPA to the renal parenchyma at a dosage of 0.025 mmol/kg or lower. GDOF-Mn-DTPA showed a high degree of hepatoselectivity, with a pronounced reduced excretion through the kidneys. The GDOF-Mn-DTPA complex is a stable compound with a high degree of selective contrast of the hepatic parenchyma, with minimal or no renal excretion, a reliable basis for a hepatoselective contrast agent for imaging and functional studies of the liver with MRI and clinical use in the near future. S2.O1. Combined 23 Na MRI and 1 H-MRSI compared to depth electrode recordings in focal epilepsy *M. Azilinon 1, 2, 3 , J. Scholly 2, 4 , W. Zaaraoui 1,2 , P. Vioult 1,2 , T. Roussel 1, 2 , M. El-Mendili 1,2 , F. Bartolomei 3, 4 , J. P. Ranjeva 1,2 , V. Jirsa 3 , M. Guye 1, 2 Na MRI and 1 H-MR Spectroscopic Imaging ( 1 H-MRSI) and compared changes probed by this multimodal approach to invasive electrophysiological recordings (using stereotactic-EEG (SEEG)), looking for neuroglial alterations deciphering EZ from the rest of the brain. Methods: We performed a multi-echo density adapted 3D projection reconstruction pulse sequence at 7 T ( 23 Na MRI) and a 3D-EPSI acquisition at 3 T ( 1 H-MRSI) in 21 patients with SEEG and 25 healthy controls. After data processing described in [1] , we obtained 23 Na concentrations and T 2 * short and long components of the signal. After data processing of 3D-EPSI described in [2] we obtained NAA, Cho and tCr 3D maps. All measures were extracted from spherical ROIs around bipolar contact of SEEG electrodes and normalized against controls using Z-score. We performed a group comparison for each imaging metrics across all bipolar SEEG-based ROIs categories (see Fig. 1 ) between patients and healthy controls, using a bootstrap t-test. Then we performed a cost-sensitive logistic regression to predict EZ using backward stepwise selection of MRI measures as predictors. Results: Group comparison showed significant increase of the short signal fraction (f) and T 2 * short in EZ, and of the long fraction concentration (Na LF ) outside the EZ (Fig. 1 ). Short fraction concentration (Na SF ) and total concentration, were overall higher in patients NAA was significantly lower in all zones in patients, particularly in EZ compared to other zones. The stepwise regression procedure did not identify any feature for the 1 H-MRSI and the 23 Na-MRI model. However, for the mixed model, T 2 * short , Na LF , Na SF and NAA were identified and were also used for the shuffled model. The mixed model exhibited a significantly better AUC and MCC than the shuffled one (Fig. 2 ). Discussion: The association of Na SF and T 2 * short increases with NAA decrease leads to the hypothesis of both energetic failure and cell viability accompanying ionic homeostasis, reflecting hyperexcitability. Conversely, increased Na LF in regions outside the EZ is more likely linked to structural changes as described by morphometry or diffusion MRI in the literature or vascular contribution. This unique combination is a useful tool to study the complex neuroglial changes in epilepsy and for presurgical evaluation. 1. Ridley, B. et al. Distribution of brain sodium long and short relaxation times and concentrations: a multi-echo ultra-high field 23 Na MRI study. Sci. Rep. 8, 4357 (2018) 2. Lecocq, A. et al. Whole-brain quantitative mapping of metabolites using short echo three-dimensional proton MRSI: 3D-1 H-MRSI Covering the Whole Brain. J. Magn. Reson. Imaging 42, 280-289 (2015) S2.O2. Longitudinal rs-fMRI and graph theoretical analysis reveal brain network changes in the GAERS rat model of absence epilepsy S2.O3. Correction of the haemodynamic artefact in fluorescence recordings using multimodal BOLD fMRI measurements Methods: AAVs encoding the ratiometric sensors Twitch-2B 2 or Laconic 3 (FRET sensors, consisting of two fluorophores: donor and acceptor) were injected in the forelimb region (S1Fl) of Fischer rats (n = 4, n = 3) . After at least 4 weeks, fluorescence recordings and fMRI were simultaneously performed under medetomidine sedation upon electrical paw stimulation (paradigm: 5 s on, 25 s off, 1 ms pulses, 1.5 mA, 9 Hz) . For fMRI single shot GE-EPI measurements (TR = 1 s) were performed at 9.4 T using a 2 cm surface coil. For fluorescence recordings the sensor expressing region was illuminated (458 nm) using an optical fiber (U: 400 lm). The fluorescence light was separated into donor and acceptor signals. Data were analyzed using MATLAB: The BOLD response of an S1Fl covering ROI was extracted using a voxel-wise U-test. 4 The ratio of the fluorescence signals was calculated (Twitch: acceptor/donor 2 , Laconic: donor/acceptor 3 ). The absorption term e -bÁBOLD was fitted to the fluorescence signals by adjusting the parameter b. For the correction, fluorescence signals were divided by this term. Data were checked for a response to the stimulation using a U-test. Results: The haemodynamic artefact was corrected in fluorescence recordings of 4 Laconic (Fig. 2 ) and 6 Twitch (Fig. 3) datasets. Before correction, stimulation-and rest-periods of all time courses (donor, acceptor and ratio) showed significant differences (p \ 0.01). After correction, no stimulus response was detected with Laconic. With Twitch we detected significant activation (p \ 0.05) in donor and ratio signals in 5 datasets. Discussion: Laconic is a lactate sensor and Twitch is a calcium sensor. At a stimulation duration of 5 s we do not expect a lactate but a calcium response. Our corrected data are consistent with this expectation. Thus, the presented correction method removes the artefact and leaves the true response. So far, light-based correction methods were presented 1 . For ratiometric sensors the calculation of the ratio of donor and acceptor signals represents such a method. However, a stimulus response was detected in the Laconic-ratio. Thus, the artefact is not completely eliminated by the light-based method, while the fMRI-based method presented in this work eliminates the haemodynamic artefact. S2.O4. Relationship between blood pressure and cerebrovascular reactivity Methods: 7 healthy volunteers had 10 MRI sessions (3 T Siemens PrismaFit) 1-week apart, at the same time of day. A BH task 5 was administered at the end of each session (ME-fMRI data, TR = 1.5 s, 5 echoes, 2.4 9 2.4 9 3 mm 3 ). CO 2 levels were measured using a nasal cannula with gas analyzer. A T1-w MP2RAGE was also collected in each session. Diastolic and systolic blood pressure were acquired before each MRI session with the subject in a supine position after a few minutes of rest, once for each arm. Then, each measurement was averaged between arms and pulse pressure (PP) and mean arterial pressure (MAP) were derived for each session. Pulse was also recorded. CVR maps were obtained from the optimally combined signal (optcom) following the steps in 6 , normalized to the MNI152 template (2.5 mm isotropic) and smoothed (5 mm FWHM) . A linear mixed effect model was set up using 3dLMEr 7 , considering the effects of MAP, PP, and pulse on CVR. As group differences were observed between females and males, the interaction with sex was also considered, resulting in the following model: The same model was run using vascular lag maps as dependent variable. The results of both models were FDR corrected (q \ 0.05). Results: MAP, PP, and sex showed significant local relations to CVR (see Fig. 1 ). MAP showed a positive impact on CVR in multiple regions of the cerebellum and cortical gray matter, except regions in the frontal lobe. PP showed a positive impact on CVR in the left insula, the right auditory cortex, the middle temporal gyrus and the medial superior frontal gyrus. At the group level, females showed a lower CVR response compared to males across most of the gray matter, in agreement with previous literature 8 . No significant relation was found for interactions, nor an impact of any variable on haemodynamic lag. Discussion and conclusion: Measuring blood pressure before collecting CVR data with fMRI should be considered, especially if differences in spatial distributions of CVR are of interest. S2.O5. Changes in intracranial compliance in patients with long-term ventriculomegaly and normotensive hydrocephalus according to PC-MRI data Introduction and purpose: To study the changes in intracranial compliance in patients with normotensive hydrocephalus and ventriculomegaly according to phase contrast magnetic resonance imaging. Materials and methods: A control group (15 people without signs of pathology) and a group of patients (10 people with clinical and radiological signs of normotensive hydrocephalus, 10 people with asymptomatic ventriculomegaly) were formed. On a Philips ''Ingenia'' 3.0 T using phase-contrast MRI and the Q-Flow method, the volume-velocity characteristics of hemo-and CSF dynamics were studied at the some intracranial levels. An assessment of arterial inflow, venous outflow, intracranial compliance index and the CSF dynamics in the control and patients groups was made. The volume flow was assessed for the following structures: brain aqueduct, foramen magnum, internal carotid and basilar arteries, superior sagittal and straight sinuses. Comparison of the measured characteristics between groups of patients and controls was carried out using nonparametric methods of analysis. Results: In the groups of patients with normotensive hydrocephalus, there was a 1.5 times decrease in intracranial compliance (p \ 0.01), a 1.5 times decrease in venous outflow in the straight and superior sagittal sinuses (p \ 0.05), an increase in the stroke volume of CSF flow at the level of the brain aqueduct by 3 times (p \ 0.01), with a change in the prevailing direction of flow to the caudo-cranial direction. In patients with ventriculomegaly, there was a 1.5 times decrease in the stroke volume of the cerebrospinal fluid at the level of the foramen magnum (p \ 0.05). At the same time, an increase in venous outflow at the level of the straight sinus is determined by 1.2 times (p \ 0.05). There was a tendency towards a decrease in the index of intracranial compliance. Discussion and conclusion: Thus, in patients with asymptomatic long-term ventriculomegaly, there is an increase in venous outflow in the absence of significant changes in the parameters of CSF dynamics as one of the compensatory mechanisms. However, such patients are at the group of risk, since there is a tendency to impairment of the functional intracranial relationship. In patients with normotensive hydrocephalus, a significant change in the compliance of the intracranial space and a decrease in venous outflow are determined, which indicates a violation of cerebrospinal fluid dynamics and changes at the level of the brain parenchyma (decreased plasticity of the brain substance). We thanks the Russian Science Foundation (project No. 19-75-20093) . S2.O6. White matter GABA 1 and GABA-levels: multivendor study For the first time GABA ? and GABA-were quantified and compared between different MRI scanners and hemispheres. The main finding of the study is that both GABA ? and GABA-concentrations are equal in the right and left hemispheres of the normal brain. This fact allows using of contralateral GABA concentration for ischemic and stroke evaluation in the situation of the one-sided lesion. Uncorrected GABA ? /tCr ratios depend on the Manufacturer in a similar way that was obtained previously in GM-riched right sensorimotor cortex [1] . Thus, comparison of GABA ? concentrations between different vendors should be at least accompanied by editing efficiency and MM co-editing corrections or using a Multi-vendor standardized MEGA-PRESS pulse sequence [2] . Both uncorrected and corrected GABA-values did not show significant changes depending on the vendor. The difference in MM contamination might not have an influence on signal intensity and should not be corrected. Moreover, since the Siemens sequence produced the least amount of MM contamination its influence on the GABA ? signal might be strong. The second reason for no difference between Siemens and Philips GABA-levels is a large spread of GABA-values. Paris-Saclay University, Gif-sur-Yvette, FR Introduction: Energy metabolism plays a crux role in brain function and its chronic deficits have been associated with many neurodegenerative diseases 1 . In vivo phosphorus-31 ( 31 P) NMR spectroscopy allows the estimation of ATPase and creatine kinase (CK) activity using saturation transfer (ST) approaches 2 . However, long acquisition times (TA) limit its use to a few regions of interest (ROI) . Here, we propose and evaluate an interleaved frequency selective 3 (FS) 3D imaging protocol to quantify the c-adenosine triphosphate (cATP) and phosphocreatine (PCr) concentrations and estimate the CK reaction flux (V CK ) in various cortical ROI in a TA compatible with clinical applications. Methods: 3 Healthy volunteers (2 M/1F, 26 ± 3 years) were scanned on a whole-body 7 T MRI (Siemens Healthineers, Erlangen, DE) using a 1 H/ 31 P 1 Tx/ 8 Rx phased array volume antenna 4 . Four PCr and cATP images were acquired using a FS SPGR sequence 5 combined to a cATP ST module (Fig. 1) . Anatomical references were also acquired. To control for the effective FA applied for excitation and saturation, as TA forbids acquisition of individual B 1 ? maps, a template in vivo B1 ? map was calculated from a cohort apart (3 M/ 3F, 29 ± 6 y.o., Fig. 2 ). Images were corrected for B 1 ? inhomogeneities following registration to the template space and segmented using nipype 7 . Five ROIs were defined from the Harvard-Oxford Atlas (Fig. 3) and both signals were averaged over each ROI. Calibration of concentrations were performed using two external references (PBS at 25/50 mM) 8 , differential and relaxation effects 9 being accounted for, using experimental or literature 10 data. To estimate the individual kinetic constant of the CK reaction (PCr ? ADP $ cATP ? Cr) (and then V CK = [PCr]k f ) from the 31 P images, magnetizations for the different saturation values were simulated using the Bloch-McConnell Eqs. 9, 11 . To compensate for the Rician noise, its average value was subtracted, estimated from an acquisition with a FA = 0°. Then, k f was determined using the ''least squares minimization'' algorithm. Results/discussion: Figure 3 shows our results. Both concentrations, k f and V CK values are all consistent to the literature 2 for ROIs 1, 4, 5. Our method would benefit from the addition of saturation bands for the muscles. Lastly, we managed to set-up a sensitive and time-efficient 3D dynamic 31 P MRI protocol exhibiting a relatively high spatial resolution (theoric 12.5 mm iso). Future work will include the application of our method to Alzheimer''s Diseases patients. This work received financial support from Leducq Foundation (large equipment ERPT program, NEUROVASC7T project). CEA. 1 Zhu et al. Front. Aging Neurosci. 2018 . 2 Zhu et al. Neuroim. 2012. 3 Rink et al., MRM, 2015. 4 Avdievich. Appl. Magn. Reson. 2011. 5 Coste et al., ISMRM 2018. 6 Boada et al., MRM. 1997. 7 Gorgolewski et al., Front. Neuroinform. 2011. 8 Soher et al., MRM, 1996. 9 Lei et al., MRM, 2003 . 10 Ren et al., NMR in biomed, 2015 Bottomley et al., MRM, 2002. S2.O8. Joint multi-field T 1 quantification for fast field-cycling imaging identifies ischaemic stroke at magnetic field strength below 20 mT Introduction: Fast Field-Cycling (FFC) imaging 1 is a novel modality that measures variations of T 1 relaxation with the magnetic field strength, exploiting novel biomarkers based on T 1 dispersion that are invisible to other imaging modalities. Here we report the results of the PUFFINS study that aimed at characterising acute and sub-acute strokes with FFC imaging. We also propose a purpose-made reconstruction method 2 that exploits the high spatial redundancy of FFC images data by using total generalised variation (TGV 3 ) regularisation over all images, combined with a mono-exponential model of magnetisation behaviour to provide T 1 maps directly. Methods: 30 patients gave informed consent to participate to the PUFFINS study from 02/2018 to 02/2020 (ethics approved by NoS-REC, number 16/NS/0136) and were scanned by FFC-imaging within 24-96 h of presentation using a field-cycled inversion recovery spin echo sequence with three to six evolution fields typically ranging from 0.2 T to 0.2 mT, 2 to 4 mm in-plane resolution, 290 mm FOV, 10 mm slice thickness, 16 to 24 ms TE and no averaging. Total scan duration was approximately 40 min. Multi-field T 1 quantification was first performed using a standard curve fitting approach and later using the joint TGV regularisation algorithm after validation on simulated test datasets. The position of the stroke found from FFC images was compared with CT and MRI images for validation. Results: The TGV-based method showed excellent noise filtering while maintaining sharp image features, clearly outperforming the reference approach on simulated images (Fig. 1) . Patient data also showed significant improvements in visual appearance over all fields (Fig. 2) and exhibited visible contrast at 20 mT and below, with lower fields showing the most pronounced contrast with brain tissues with a maximum at the lowest field used. The infarct region measured by FFC corresponded with CT and 3 T MRI images. The in-vivo T 1 dispersion profiles obtained from the T 1 images showed profiles for the stroke area that differed from other brain tissues in the low-field regime (Fig. 3) , likely due to modifications of water diffusion due to cell damage. Discussion: This is the first-ever in-vivo measurement of T 1 dispersion of stroke, which was made possible by the joint TGV regularisation method. This processing tool is available online 4 . The results demonstrated the potential for low fields in this clinical field but also the limitations of T 1 contrast above 20 mT. We hope these results will help guiding the design process of dedicated lowfield devices and we are currently starting the second stage of the PUFFINS study to explore potential FFC biomarkers for the differentiation of stroke types and mimics. References: 1. Broche, L. M. et al. Sci. Rep. 9, 10402 (2019 ). 2. Bödenler, M. et al. Magn. Reson. Med. In print, 3. Knoll, F. et al. Magn. Reson. Med. 65, 480-491 (2011 ). 4. Maier, O. et al. J. Open Source Softw. 5, 2727 . High-resolution imaging at low-fields: how far can we go? Introduction: The development of superconducting magnets with higher and higher field strength has been driven by the promise of exploring unprecedented details and anatomical richness with MRI, especially in the field of neuroscience. Following this line of thought, it was recently shown that 0.44 mm isotropic images can be achieved with a 3D MPRAGE in a 40 min scan at 7 T offering a SNR of about 50 and 20 for brain tissue and CSF, respectively (1) . Contrary in this work, we explore brain resolution limits on a low-cost clinical lowfield system. Methods: Imaging was performed on a 0.55 T low-field MR-system (MAGNETOM Free.Max, Siemens Healthineers) with a gradient system offering 26 T/m at 45 T/m/s and using a 12-channel head coil. To explore resolution boundaries for brain MRI, we used a recently proposed radial bSSFP sequence, termed bSTAR (2) , offering highest sampling efficiency (g). For a TR = 3.9 ms, a hard RF pulse = 600 us, a bandwidth = 631 Hz/pixel, for a field-of-view = 256 9 256 9 256 mm 3 , g = 0.81, and scanning was completed within 9:45 min using 150,000 half-radial projections. For brain tissue at this field strength, T1/T2 * 7-8 (3) , and scanning was performed a flip angle = 40°. This setting offers a nominal isotropic resolution of 0.66 mm that can be subsequently regridded to lower resolutions to explore SNR. The reconstruction was performed using FISTA (4) . Results: Exemplary bSTAR images with 0.66 mm isotropic resolution are shown in Fig. 1 , offering an SNR of 25 and 63 for white matter and CSF, respectively. Especially fluid-like structures, such as the inner-ear (cf. Fig. 2) , benefit from the exceptional bSSFP signal properties. Generally, with decreasing resolution, SNR is increased as expected; see Table 1 . Overall, sub-millimeter brain MRI with bSTAR can be performed in less than 5 min. Discussion: Among all sequences, bSSFP not only offers the highest SNR per unit time (5) , but also its steady state and thus its signal increases with decreasing field strength. Thus, the combination of maximum SNR with highest sequence efficiency, as proffered by bSTAR, can be used to yield an estimate of the maximum achievable resolution. Here, we have shown that for clinically tolerable scan times of less than about 10 min, this limit is far beyond the typical desire for sub-millimeter voxel sizes. Moreover, at low fields, bSSFP imaging does not suffer from banding-artifacts and can be combined with different magnetization preparation schemes to modify its intrinsic T2/T1 contrast. In conclusion, especially at low fields, bSSFP-based imaging surpasses expected resolution limits thus offering excellent prospects for broad clinical use and applicability. An open source triggered CEST module for Bruker systems for reliable CEST MRI with efficient motion artifact mitigation Discussion: It was demonstrated, how severe motion artifacts bias CEST MRI in animals. With the proposed open source module, this issue was significantly reduced. The modular source code requires only minimal manual programming to include it into an existing RO. We hope the proposed open source module is useful for other research groups, and will facilitate improved CEST experiments in animal models. References: 1 Zaiss, JMR, https://doi.org /10.1016/j.jmr.2018.11 .002 2 Zaiss, NMRB, https://doi.org/10.1002/nbm.3879 3 Guivel-Scharen, JMR, https://doi.org /10.1006/jmre.1998.1440 4 Schuenke, MRM, https://doi.org /10.1002/mrm.26133 S3.O2. Addressing inadvertent MT effects in 7 T brain T1 mapping by simple pulse length scaling Introduction: T 1 mapping by the variable flip angle method (VFA) is known to be affected by inadvertent magnetization transfer (MT) effects from the excitation radiofrequency pulse. These effects can be moderated by ''controlled saturation magnetization transfer'' (CSMT) 1 or more simply by varying the pulse duration such that B 1 ? RMS is kept constant between acquisitions 2 . We show that also at 7 T, incidental MT can be addressed by scaling the excitation pulse length to maintain constant B 1 ? Methods: Bloch simulations were used to examine the effect of pulse length scaling on the bound pool. A healthy experienced volunteer was scanned on a 7 T MR system (MAGNETOM Terra, Siemens Healthineers, Erlangen, Germany) using a multi-echo 3D gradient echo sequence (RF-and gradientspoiled) at an isotropic resolution of 0.8 mm (TR 25 ms, 8 equally spaced echoes TE 2.8.. 18.9 ms, FA 8 and 25 degrees) . The image pairs for T 1 mapping with variable flip angles were acquired three times with different on-resonance excitation pulses of sinc window shape and lengths: (a) equal (both 560 ls) 3 , (b) scaled linearly for constant B 1 ? peak (180, 560 ls) 4 and (c) scaled quadratically for constant B 1 ? RMS (140, 1368 ls) 2 . B 1 ? was mapped by the SE/STE EPI method 5 . Quantitative maps were calculated with the hMRI-toolbox 6 within the SPM12 framework (https://www.fil.ion.ucl.ac.uk/spm/software/ spm12/) in MATLAB (Mathworks, Natick, USA). The same toolbox was used to compute maps of tissue probability. Results: Simulations ( Fig. 1 ) demonstrate the expected variations in bound-pool saturation and steady-state signal in the three regimes (equal pulse lengths, constant B 1 ? peak , constant B 1 ? RMS ). In vivo results (Fig. 2 ) reveal that constant B 1 ? RMS scaling leads to tighter clustering of grey and white matter T 1 values, suggesting lower spatial bias. An example slice shown in Fig. 3 also shows lower spatial bias and a sharper white matter/grey matter interface. 3.5 ms, whereas the MT-weighted volume (S MT ) was acquired a T RF of 200 ls yielding TE1/TE2/TR = 0.13/2.7/2.9 ms. Scan time was 3:52 min for the MT-weighted and 4:40 min for the non-MTweighted scan. Magnetization transfer ratio (MTR) images in percentage units (pu) were calculated from MTR = 100 9 (S nonMT -S MT )/S nonMT . Results: As expected, bSSFP imaging is essentially free of banding artifacts at low field and MT-effects are strongly modulated by the RF pulse duration (Fig. 1 ). Corresponding MTR images are shown in Fig. 1 -C. For the proposed setting, average MTR values for WM and GM are 33 pu and 26.7 pu (see ROI in Fig. 1 ). : Surprisingly, the proposed method provides enough signal-to-noise to yield submillimeter isotropic whole brain MTR images at 0.55 T in a clinically acceptable scan time. This is particularly due to the high sequence efficiency proffered by the half-radial dual-echo approach in combination with the increased T2/T1 ratio with decreasing field strength offering increased bSSFP signal. Moreover, any possible residual off-resonance-related MTR variations are successfully mitigated at low field. Future work will explore different RF pulse prolongation settings to maximize MT sensitivity. In conclusion, bSSFP-based MT-sensitized imaging offers excellent prospects for broad clinical translation and application in low fields. References: 1. Bieri O. & Scheffler K. MRM 2007; 58:511-518. 2. Amann M. et al. Neuroimage. 2015; 108:87-94. 3. Bauman G. & Bieri O. MRM. 2020; 84:237-246. S3.O4. Sub-millisecond 2D MRI of the vocal fold oscillation using single point imaging with rapid encodin (SPIRE) Introduction: The position of the vocal folds in the larynx limits optical examinations to a top down view in laryngoscopy. MR does provide the free choice of imaging plane, however, the rapid oscillation of 100 Hz and higher is too fast for conventional acquisitions. We have shown that the use of short PE gradients enables sub-millisecond temporal resolution when a one-dimensional periodic motion is aligned with PE direction [1] . Here, we show how extending PE to the second in-plane direction allows for dynamic imaging of the vocal folds oscillation in the coronal view with sub-millisecond temporal resolution. Methods: The proposed method uses single point imaging [2] to encode k-space and each pair of PE gradient moments is applied as fast as possible, such that encoding time t PE decreases towards k-space center (Fig. 1 ). To achieve a spatial resolution below 1 mm, a rectangular FOV (60 mm 9 41 mm) and a matrix size of 64 9 44 were chosen, resulting a maximum gradient duration of t PE = 660 ls. Two transverse saturation bands were applied in SI direction to avoid oversampling artefacts (Fig. 2) . Unwanted motion of the larynx was measured by acquiring interleaved projection navigators in both inplane directions. Saturation bands and navigators were applied about every 100 ms. With a = 10°, TE = 1.2 ms, TR = 2.45 ms and 24 repetitions, the total acquisition time was about 3 min. The volunteer was asked to phonate at 160 Hz and rebreathe at will. An electroglottogramm (EGG) was acquired during the MR-measurement with two MR-safe electrodes [3] , to obtain the oscillation phase of the vocal folds during each phase encoding gradient. With this information, the acquired MR-data are gated in one of 10 k-spaces. Prior to reconstruction, motion was estimated using the navigator data by calculating the phase only cross correlation to a reference line. A spatial shift is then corrected using the Fourier shift theorem by multiplying a constant phase to the k-space point. Dynamic images are reconstructed with a total variation constraint using BART [4] . Results: About 28% of the acquired data is rejected, mainly because no phonation was detected by the EGG. During all accepted acquisitions the mean oscillation frequency was 165.13.4 Hz. The individual reconstructed frames are shown in Fig. 3 , each with a temporal resolution of 606 ls, as well as the mean EGG signal during each frame. Consistent with the low EGG signal, the vocal folds are opened in the first two images. In frames 3-7 the vocal folds are closed and an upwards motion can be seen after which both sides separate again. Discussion: For the first time, dynamic 2D MR images of the oscillating vocal folds are presented. The dynamic image information correlates well with the oscillation phases from the EGG-signal, and the open and closed phase can be clearly distinguished in the SPIRE images. References: [1] https://doi.org/10.1002/mrm.27982. [2] https://doi.org/10.1016/0378-4363(85)90087-7. [3] https://doi.org/10.1002/mrm.26037. [4] https://doi.org/10.5281/zenodo.592960. S3.O5. QSM streaking artifact reduction using non-uniform Fourier transformation Methods: To adjust the sampling pattern, we drew frequency space coordinates for the nuFT from a probability distribution that resembles D: low values in D correspond to low probabilities of sampling these values and more samples where D -1 is well defined. The nonuniformly sampled D -1 was then thresholded and used as in TKD. The nuFT was implemented using the BART toolbox [8] . All computations were performed using MATLAB on an Intel i9-10920x (12 core) CPU with 64 Gb of RAM. We compared our method to TKD [1] , and a non-linear Total Variation regularised method (nlTV), using default settings (threshold 2/3) and open source code [6, 7] , on local 3D GRE field maps of the brain acquired at high (1 mm isotropic) resolution in vivo as described in [10] . The direct methods were 4 9 oversampled in k-space (zero padded) and the results were corrected for underestimation [2] . Results: Figures 1 and 2 show a comparison of susceptibility maps between the different methods. A difference image between the standard TKD reconstruction and proposed nuFT-based susceptibility map is presented in Fig. 3 and shows primarily streaking artifacts (e.g. see arrows). The nuFT reconstructions were computed in less than 4 min, TKD in seconds, and nlTV in 2 min. Discussion: Here, we investigated the effect of non-uniform sampling of frequency space for QSM. This nuFT approach resulted in fewer streaking artefacts without using extra explicit regularisation. It also provides detailed control and insight into the kernel (D -1 ) in Fourier space and could be combined with compressed sensing techniques [11] to further improve QSM reconstruction. [1] Shmueli, Magn. Res. Med., 2009. [2] Schweser, Magn. Res. Med.,2012. [3] Karsa, ISMRM 2019. [4] Milovic, Magn. Res. Med., 2018. [5] Lai, MICCAI 2020. [6] https://xip.uclb.com/product/mri_qsm_tkd [7] https://gitlab.com/cmilovic/FANSI-toolbox [8] https://zenodo.org/record/4570601 [10] Karsa, Magn. Res. Med, 2018. Introduction: Conventional clinical routines use qualitative magnetic resonance imaging (MRI) to assess diseased tissue. Multi-parametric mapping (MPM) of R1, R2*, PD, and MTsat is a quantitative MRI method that promises better comparability across sites and time as well as a higher sensitivity to systemic tissue changes. For clinical applications of MPM, the required sequences need to be accelerated and their accuracy be verified. Therefore, this study investigates the reproducibility and repeatability of MPM using sequences accelerated with Compressed SENSE. Methods: Five healthy subjects were scanned three times on a Philips 3 T Ingenia Elition. MPM comprised B1 mapping and gradient echo sequences with T1w: TR = 18 ms, a = 25°; PDw: TR = 18 ms, a = 4°; MTw: TR = 48 ms, a = 6°, a MT = 220°, t MT = 8 ms, f MT = 1000 Hz; with six echoes each (TE1/DTE = 2.4/2.4 ms) and 1 9 1 9 1 mm 3 resolution. In each scan session, three different imaging accelerations were used: standard SENSE (AP: 2, RL: 1.25, 20 min) and Compressed SENSE (CS) with acceleration factors CS = 4 (15:40 min) and CS = 6 (10:30 min). Quantitative parameter maps (qMaps) were calculated using the in vivo histology (hMRI) toolbox 1,2 correcting for B1 bias 3 and insufficient RF spoiling 4, 5 . Coefficient-of-variation (CoV) maps were calculated for each quantitative parameter across both scan session (''repeatability'') and accelerations within the same session (''reproducibility'' with different acceleration parameters) and compared quantitatively in wholebrain gray matter (GM) and white matter (WM). Results: R1, R2*, PD, and MTsat parameter maps appeared visually similar across different accelerations (Fig. 1A ) and scan sessions ( Fig. 2A) . Repeatability-based CoV values of qMaps acquired with the same acceleration across scan sessions were comparable for SENSE, CS = 4, and CS = 6 ( Fig. 1B) . Likewise, reproducibilitybased CoV values of qMaps acquired in the same scan session with different accelerations were comparable for scans A, B, and C (Fig. 2B) . In GM and WM, subject-mean reproducibility-based CoVs (green) were comparable to repeatability-based CoVs (blue) for R1 and PD and even slightly decreased for R2* and MTsat (Fig. 3 ). Discussion: Both reproducibility-based and repeatability-based CoV values of all qMaps agree well with those from a previous study 6 . Average R1 and PD values showed similar variability across either scan sessions or accelerations. Moreover, average R2* and MTsat values were even more stable (lower CoV) across accelerations than scan sessions. Together with the high visual similarity between qMaps acquired with different accelerations, this suggests that the highest investigated acceleration, CS with acceleration factor of 6, provides accurate MPM results. As this allows to considerably reduce scan duration, CS = 6 facilitates MPM in clinical routines. 1. Tabelow McMaster University, Department of Electrical and Computer Engineering, Hamilton, CA; 2 St. Joseph's Healthcare, Imaging Research Centre, Hamilton, CA; 3 St. Joseph's Healthcare, Firestone Institute for Respiratory Health, Hamilton, CA Introduction: Computed Tomography (CT) has been the standard imaging technique to evaluate lung structure. However, due to its ionizing radiation and lack of functional information, novel Magnetic Resonance Imaging (MRI) methods are of high interest. Alternatively, hyperpolarized noble gas MRI has shown great promise. This approach requires breathing in hyperpolarized xenon-129 (HP-129 Xe) and subsequent breath-holding during acquisition. In patients with severe disease, such as severe asthma, breath-holding is often difficult [2, 3] . Accelerated imaging using Compressive Sensing (CS) can help [4] . CS aims to recover good quality images from undersampled k-space permitting faster data acquisition. CS is now becoming routine for 1H-MRI. However, its application in HP- 129 Xe lung ventilation imaging has not been optimized. The goal of the current work is to assess CS sampling schemes and their effect on image Signal to Noise Ratio (SNR), Structural Similarity Index Measure (SSIM) and resolution (i.e. Point-Spread Function, PSF) . Methods: Ten subjects (n = 5 healthy and n = 5 severe asthma) were scanned using a GE MR750 3 T scanner, acquiring fully sampled 3D multi-slice HP- 129 Xe lung ventilation images (10 s breath-hold, 128 9 80 (PExFE) and 16 slices) [6] . Using fully sampled data, 200 masks were pseudo-randomly generated [5] each at 7 different sampling rates, 15%, 25%, 35%, 45%, 55%, 65% and 75%. The Parallel Imaging Compressive Sensing (PICS) command from the Bart toolbox [7] , with L1 wavelet optimization and 100 iterations, was employed to reconstruct undersampled data. SNR, SSIM and PSF of each were subsequently compared. Results: Example HP- 129 Xe ventilation images are shown (Fig. 1) . The ratio of the main lobe to the standard deviation of the PSF sidelobes is considered the incoherence value reported for each image. The mean and standard deviation of each SNR, SSIM and incoherence values of all 5 healthy data are reported in Table 1 ; the lowest and highest sampling ratios (15%, 75%) are depicted in Fig. 2 in the right column. The exact computation for the asthmatic data also can be found in Table 1 and the left column of Fig. 2 . Discussion: With greater undersampling, the degree of incoherence and SSIM decreased, while SNR increased (expected due to CS reconstruction noise filtering [4] ). Higher SNR in asthma data was noted and is likely the result of the same amount of HP-129 Xe gas being distributed over less lung volume due to ventilation defects, compared to healthy. Most importantly, there was a high degree of variation in results from each of the 200 masks of each undersampling rate. The greater the undersampling the greater the variation in SNR and SSIM on both healthy and asthma groups (Table 1) . Incoherence also decreases for healthy, yet not for asthma lungs. These results indicate there is an optimized rate and pattern of undersampling. This needs thorough investigation to allow optimal CS accelerated HP- 129 Xe lung ventilation imaging. Iterative RAKI with complex-valued convolution for improved image reconstruction with limited scanspecific training samples *P. Dawood 1, 2 Introduction: RAKI [1] is a scan-specific machine learning approach for k-space interpolation. It has demonstrated superior reconstruction quality in comparison to the standard parallel imaging approach GRAPPA [2] . However, RAKI may reveal severe artefacts in standard 2D imaging given a reduced amount of auto-calibration-signal (ACS) for high accelerations due to its increased parameter space. We propose iRAKI, an iterative extension of RAKI with complex-valued Convolutional Neural Network (CNN) [3, 4] to improve its reconstruction quality with only a limited number of ACS. Methods: iRAKI:An initial GRAPPA reconstruction is performed using the fully sampled central k-space region as original ACS to obtain synthetic ACS of increased size for RAKI training [5] , and to assign an increased filter size to the first hidden layer in the CNN [6] ( Fig. 1 ). Synthetic ACS is extracted from N = 100 central lines of the initial GRAPPA k-space reconstruction. Subsequent iterations follow in which the CNN weights are transferred, and further optimized using N' = 100 central lines from the RAKI reconstruction of the previous iteration as ACS (orig. ACS re-inserted at each iteration). The learning rate g is decreased by Dg after each iteration (initial value g 0 = 0.05) and we set Dg = 0.003 for R = 4 (0.004 for R = 5, R: undersampling rate). The total iteration number amounts to g 0 /Dg. Complex ReLU [7] is chosen as activation, and the mean-squarederror as cost function. The final optimized CNN interpolates the multi-coil subsampled k-spaces simultaneously, which are transformed via FFT and combined via RSS. In vivo data: 2D anatomical brain imaging with 16 coils acquired on healthy volunteers at 3 T (Magnetom Skyra, Siemens Healthineers) were used for evaluation. One set (ref. to as neuro1) was acquired using TSE with TR/TE = 500/10 ms, FOV = 220 9 193 mm 2 ,matrix size 256 9 224. Another set (ref. to as neuro2) was performed using FLASH with TR/TE = 250/2.90 ms, FOV = 230 9 230 mm 2 , matrix size 320 9 320. Results: Using only 18 ACS lines (i.e. 8% of total Phase-encoding lines), iRAKI improves RAKI by removing severe artefacts encountered in the neuro1 dataset for R = 4 ( Fig. 2 top) and by enhancing the suppression of noise-amplification regarding GRAPPA in neuro2 dataset ( Fig. 2 bottom) , also indicated by its SSIM and NMSE metrics ( Fig. 3) for R = 5, too. The iterative procedure with learning-rate decay reduces the over-all reconstruction error in iRAKI and leads to convergence towards improved data consistency for the shown imaging experiments at four and fivefold acceleration. Introduction: Motion is the main extrinsic source for imaging artifacts in MRI which can strongly deteriorate image quality and thus impair diagnostic accuracy. Numerous motion correction strategies have been proposed to mitigate these artifacts [1] [2] [3] [4] . We have recently proposed a deep learning (DL) framework 5,6 (MedGAN) to perform motion correction without prior motion knowledge. In this work we reformulate the motion correction as an image reconstruction task 7 which allows to involve data consistency during inference. The proposed MedGAN motion correction is trained and evaluated on a cohort of 350 patients for non-rigid motion correction in T2w abdominal imaging. Material and methods: We formulate motion correction as an image reconstruction problem with randomly undersampled high-frequencies. We assume that i) motion can be modelled as sparse outliers in k-space 8 and ii) non-rigid motion can be represented as a superposition of local translational displacements 9 . The motion-affected k-space y m (k) is thus a composition of the motion-free k-space y(k) and the k-space modulated by linear phase drifts along phase-encoding directions k y ,k z described by the local translations u t in a neighborhood W at a random high-frequency given by the sampling mask U (Fig. 1B) . The proposed MedGAN in Fig. 1A contains a generator consisting of six cascaded UNets with intermittent proximal gradient data consistency and receives a complex-valued (magnitude and phase) motionaffected image as input. The MedGAN is trained adversarial with a patch-based magnitude/phase discriminator and with perceptual, style transfer and mean-squared-error loss 5 . During training, variabledensity Gaussian and Poisson-Disc undersampling masks are generated with 1 9 to 10 9 acceleration. Training is conducted for 50 epochs via ADAM (learning rate scheduler 0.001, batch size 8). MedGAN is trained on 330 patients and tested on 20 patients (informed and written consent), imaged with an axial T2w BLADE: TE/ TR = 104/7560 ms, ETL = 51, GRAPPA 2x, FA = 130°, 0.9 9 0.9 9 4 mm 3 resolution. Results and discussion: Figure 2 shows two patients with mild and strong simulated motion to investigate data consistency. Figure 3 depicts two patients with real respiratory motion. Motion-corrected images can be obtained with close resemble to motion-free case. The current study has limitations: Only 2D MR images of a single sequence type were corrected, i.e. through-plane motion correction is limited. Conclusion: Retrospective motion correction can be formulated as an image reconstruction with the proposed MedGAN providing high quality and motion-free images ensuring data consistency during inference. Introduction: Functional Quantitative Susceptibility Mapping (fQSM) allows the measuring of time-varying magnetic susceptibility with potentially higher spatial specificity than conventional fMRI. The use of fQSM in combination with General Linear Model (GLM) and On/Off paradigms has been previously assessed 1, 2, 3 , yet the evaluation of potentials and limitations of this method in advanced experimental paradigms and analyses is still missing. Moreover, voxels with positive fQSM response, i.e., of the same sign as fMRI, require further investigation. To test the usefulness of fQSM and to characterize its activations, we used 7 T MRI, tonotopic mapping, and both univariate (GLM and population Receptive Field; pRF) and multivariate (Representational Similarity Analysis; RSA) analyses. Methods: We acquired complex-valued GRE-EPI datasets at 7 T (TE/TR = 21.3 ms/2.5 s, voxel size = 1.8 mm isotropic) on 9 subjects passively listening to pure tones of 10 frequency bands (100-6400 Hz, 7.5 s-blocks interleaved by 7.5/10 s-rest, 8 trials per band). fMRI and fQSM data were processed and analyzed via GLM (Fig. 1 ). Positive fQSM responses were anatomically localized via gray/white matter (GM/WM) segmentation performed on 1 mm-MP2RAGE image. Tonotopic maps were computed via pRF and the cortical representation of distinct frequency bands was evaluated via RSA and Multidimensional Scaling (MDS). Results: The fQSM tone-responsive area was 15 ± 9% of the active Primary Auditory Cortex (PAC) observed with fMRI and had higher inter-subject variability ( Fig. 2A/B ). The ratio of active voxels belonging to GM was higher (p \ 0.01) for fQSM (73 ± 8%) than fMRI (65 ± 4%). Positive fQSM responses were found in 14 ± 10% of active voxels, mainly in WM (42%) rather than GM (12%). pRF revealed that fQSM and fMRI tonotopic maps were significantly correlated (Pearson's r = 0.22 ± 0.11, p \ 0.001). Although fMRI maps were smoother, voxel tunings in fQSM were more equally represented across the whole frequency range (p \ 0.05) (Fig. 2C/D) . RSA results showed correspondence between response patterns of fMRI and fQSM (r = 0.49 ± 0.15, p \ 10 -5 ) (Fig. 3A) . In MDS analysis, trials of the same stimulus condition clustered along the first dimension for both fMRI and fQSM (Fig. 3B ). Discussion: fQSM had lower sensitivity but higher spatial specificity than fMRI. Positive fQSM responses were mainly in WM and do not hamper the interpretation of GLM results. Even though fMRI and fQSM tonotopic maps were comparable, the representation of frequency tunings was more balanced in fQSM. Despite being noisier than fMRI, fQSM activation patterns can be used to reconstruct the fine-grained topographical mapping of PAC. Overall, fQSM results are complementary to conventional fMRI. Also, its quantitative nature and higher spatial specificity support its use in longitudinal and multicentric studies and pre-surgical mapping. Introduction: Current diagnosis and monitoring of gliomas include the assessment of conventional MR imaging, and gadolinium contrast enhancement (GdCE) on T 1 weighted (T 1 w) images. However, the administration of gadolinium-based contrast agents (GBCA) is invasive and of concern due to gadolinium deposition in the brain 1 . A noninvasive alternative to GdCE is warranted to minimize these associated risks. Chemical exchange saturation transfer (CEST) is an MRI technique that can provide endogenous contrast from metabolites and proteins. Previous studies have shown that amide proton transfer (APT) and nuclear overhauser effect (NOE) CEST contrasts are increased and decreased respectively in high-grade gliomas (HGG) 2,3 . More recently, glutamate-weighted CEST (GluCEST) contrast has investigated metabolic changes in diffuse gliomas at 7T 4 . However, the relationship between GdCE, APT, NOE, and GluCEST contrasts in HGG patients remains yet to be studied. In this preliminary work, we compare APT, NOE and GluCEST contrasts with GdCE contrast in HGG patients at 7 T. Methods: We prospectively included two pre-operative patients: one 48 years old male patient and one 60 years old female patient, both with a glioblastoma. CEST scans were performed on a 7TMRI scanner, whereas anatomical scans were retrospectively obtained from clinical imaging at 3 T. Image acquisition parameters can be found in Table1. Image post-processing and analysis details are shown in Fig. 1 . Results and discussion: Figure 2 shows CEST maps and the corresponding GdCE T 1 w and T 2 -FLAIR anatomical images from both patients. Patient 1, who has a clear GdCE lesion on the post gadolinium T 1 w image, exhibits a higher APT AREX contrast (Fig. 2D ) and lower NOE AREX contrast (Fig. 2E ) in the area corresponding to the GdCE. This is in accordance with previous studies where increased APT and decreased NOE contrast were found for HGG at 7T 2,3 . On the other hand, patient 2, who did not have any GdCE, exhibits lower APT (Fig. 2I ) and NOE contrast (Fig. 2J ).This suggests that APT contrast relates to the tumor physiology portrayed by contrast enhancement. Interestingly, GluCEST contrast was increased for patient 2, who did not have any GdCE. This is in accordance with a previous study where GluCEST contrast was observed in diffuse gliomas, but contrarily to their observations of increased GluCEST contrast in the presence of GdCE 4 . In conclusion, our preliminary observations suggest a relationship between tumor enhancement, APT and GluCEST contrast in HGG. These highlight the potential of CEST to be used as an alternative imaging method to GBCAs. Future work is in progress to verify our findings in a larger group of patients, as well as to optimize image quality and data analysis to allow quantitative measurements. Methods: Hybrid EPI (HEPI, a fast acquisition technique combining GRE and SE) DSC were acquired from 3 patients with confirmed diagnosis of glioma. Dynamic images (TR/TE GRE /TE SE 1.5 s/20 ms/ 70 ms, 15 slices, 3 T, MR750, GE) were acquired during bolus injection of 7.5 ml of GBCA (Gadovist, Bayer, GE) after a preload of equal size. We imported the HEPI sequence as played out on the scanner into a Bloch based DCE simulation tool 4 that simulates CA extravasation and MR signal. A dictionary was created by simulating 5 vessels with k = [0,1.8,4 .8]*10 -3 s -1 , 30 logarithmically spaced values of R = [5, 100] lm and rCBV = [0.5,10]% 3,4 . For each dictionary atom the simulation included 400 s to cover preload and delay, followed by 100 s of the main bolus and HEPI acquisition. The dictionary was matched to the patient data by allowing a separate scaling factor for the GRE and SE to compensate for baseline (T2) differences between dictionary and in-vivo signals. The parameters obtained were compared with those estimated using the VSI method 3 . Results: The k, R and rCBV values of the best dictionary match to the in-vivo data are shown in Table 1 for 3 voxels in white matter, grey matter and tumor. Non-zero k values are obtained only for the enhancing tumor voxels. For the voxels of patient 1, Fig. 1 shows the GRE and SE time series and dictionary match. Figure 2 shows for each patient a slice of the T2 FLAIR with the corresponding parametric maps obtained by the two methods. It is observed that both techniques can differentiate normal and tumor tissues. The enhancing tumor in patient 3 is also seen as hyper-intense area in its rCBV maps while those obtained for VSI are comparatively lower, which needs further investigation. Discussion: By dictionary matching of HEPI-data a good match was found between simulated and in-vivo data. We compared the quantitative maps of R and rCBV from MRVF to conventional VSI values and found similar patterns, although also quantitative differences. The dictionary should be expanded by more densely sampling k and R values to improve the matching. First-pass DSC-HEPI MRVF has the advantage that it is faster compared to MRVF measuring asymmetric SE signals during static pre/post-contrast phases 2 . Further research will study whether permeability is essential for the dictionary for DSC with preload. Remarkable myogenic effect of ultrasound combined with microbubbles on the brain perfusion in rats Introduction: The use of ultrasound with microbubbles (US ? MB) for Blood Brain Barrier (BBB) opening is an increasingly common method in clinical and preclinical research. However, its effects on cerebral perfusion have been little investigated. The aim of this study was to evaluate changes in brain perfusion following US ? MB in rat brain using pseudo-continuous arterial spin labelling (pCASL) [1] , a non-contrast-based method, and Dynamic Susceptibility Contrast (DSC), a Gadolinium based method. Methods: The protocol included 10 female Wistar rats, 3 of them were used as control group. A catheter was placed on the rat tail vein. Prior to sonication, animals were imaged in a 7 T MRI scanner (Bruker Biospec 70/20) for phase optimization acquisitions of the pCASL sequence. Outside the magnet, the animals received an injection of 200lL of microbubbles (Sonovue) and were exposed to unfocused US using a flat transducer with an acoustic pressure of 0.3 MPa applied at the right hemisphere during 2 min. A series of pCASL acquisitions (TE = 15 ms; TR = 2600 ms; PLD = 300 ms) were then carried out every 4 min to assess Cerebral Blood Flow (CBF) after BBB opening. ADC maps were computed from the DTI Bruker protocol (TE = 20 ms; TR = 2.5 s; b * 0 and 800 s/mm 2 in 3 directions). Following the pCASL acquisitions, T1 maps were acquired, prior and after the injection of 300lL of Gd (Dotarem, 0.5 mmol/mL), using a Multi-Inversion Time sequence (FAIR-EPI, TE = 13.5 ms; TR = 10 s). Three animals from the group underwent DSC acquisition during Gd injection (TE = 10 ms; TR = 500 ms; NR = 300). CBF maps were computed using MP3 [2] . Results/discussion: Following US ? MB, CBF maps show, first, a decrease at certain regions of the exposed hemisphere ( Fig. 1 ) up to 40%, then, a progressive recovery. The volume with CBF affected by US ? MB is different from the one showing a BBB opening (Gd extravasation) albeit both volumes displayed similar CBF evolution patterns (Fig. 2) with a recovery after 50 min confirmed by the DSC-CBF maps. ADC maps between the exposed region of interest and its contralateral remain unchanged. These findings suggest that the transmural pressure increase consecutive to the US ? MB leads to BBB opening but can also produce a local myogenic response. Indeed, the myogenic response is a vasoconstriction in response to an increase in transmural pressure (i.e. the pressure across the vessel wall) [3] . Moreover, pCASL-CBF measurements in our experiment were not affected by the BBB disruption in agreement with the generalized model assuming a rapid water exchange [4] . This approach is promising to assess the local vasoconstrictive capacity implied in the cerebral autoregulation and estimate the potent cerebral vascular reserve (CVR). Further investigations using small volume focusing US transducers are needed to build CVR maps. Introduction: Arterial Spin Labeling (ASL) is a non-invasive method which can be used to measure the water exchange through the blood brain barrier (BBB) to obtain its functionality and integrity using a multi-echo ASL sequence (multi-TE) [1] . The imaging voxels consist of the intravascular and extravascular space as well as the arteries. In this abstract a new three-compartment model (3CM) is presented which introduces an intra-voxel transit time ITT between the arteries and vascular space which has been neglected so far in the twocompartment model (2CM) [2] . Methods: Theory: Compartment 1 contains the signal of the blood in the arteries before it enters the intravascular space (compartment 2) after ITT. There, the blood water exchanges with tissue water (compartment 3) via the BBB (Fig. 1) . Simulation: 3D-data were simulated for TI = 0-4000 ms and TE = 0-500 ms for different exchange times (Texch) (100-1000 ms) using the 3CM with following parameters: M0 = 1; f = 60 ml/100 g/min; alpha = 1; ATT = 500 ms; ITT = 300 ms; BL = 1800 ms; T1b = 1700 ms; T2b = 150 ms; T1tis = 1100 ms; T2tis = 80 ms. The simulated signal at eight inflow times (TI) (100, 600, 1000, 1400, 1800, 2200, 2600, 3000 ms) and six echo times (TE) (19, 57, 95, 133, 171, 209 ms) were then used to determine the quantified perfusion values and the exchange time with the 2CM. Results: Figure 2 shows the simulated data and the fit of the 3CM and 2CM for exchange time Texch = 100 ms exemplarily for TE = 50 ms and 200 ms. It can be seen that for lower TE values the 2CM can describe the signal similarly to the 3CM, but with clearly overestimated Texch. For higher TE values the signal curves cannot be reproduced properly with the 2CM. In Fig. 3a the Table shows the calculated Texch and f values for different simulations with varying Texch. It can be clearly seen that the 2CM overestimates Texch in all cases. The observed relative standard deviation to the simulation value is in the range of 220% for Texch = 100 ms and decreases with increasing Texch to 35% (Fig. 3b ). Discussion and conclusion: The signal of the arteries and ITT are not explicitly considered in the 2CM. This explains the huge difference in the calculated Texch for the simulated data. Figure 2 shows that the 2CM can model the signal curve nearly as well as the 3CM for lower TE times, but overestimates Texch up to 220%. Especially, for small Texch the 2CM would compensate the ITT missing in the model with an apparently higher Texch. As a leakage of BBB would normally be seen in a decrease of Texch [3, 4] , using the 2CM instead of 3CM (where the ITT is explicitly modelled) disguises the drop in Texch. This points out how important it is to incorporate ITT for a robust and correct detection of the BBB integrity. S4.O6. Does hypocapnia affect oxygen induced changes in brain T1 times at 3 T? Introduction: Breathing high concentration oxygen causes T 1 shortening due to the paramagnetic effect of molecular oxygen. However, during an hyperoxic challenge in healthy individuals, pCO 2 reduces resulting in contraction of vascular smooth muscle and decreased cerebral blood flow. This can reduce oxygen delivery and be a confounding factor in hyperoxia challenge MRI 1 . We sought to measure in-vivo hyperoxic DT1 times with and without control of end tidal pCO 2 (PETCO 2 ) in various brain regions. Methods: 7 healthy volunteers were scanned at 3 T (Philips Ingenia) and T 1 estimated via the variable flip angle method (3D spoiled GRE, FA = 3°/18°, TE/TR = 0.896/2.7 ms, matrix: 80 9 80, FoV: 240 9 240 mm, 37 9 3.5 mm slices, 300 dynamics, 2.95 s each). Oxygen was delivered 4 min into the acquisition using a sequential gas delivery breathing circuit with flow rates adjusted to control PETCO 2 (RespirActTM, Thornhill Research). Two runs were performed; with and without control of PETCO 2 . Images were corrected for motion and segmented into Grey Matter (GM), White Matter (WM) and Cerebrospinal Fluid (CSF) using SPM12 2 . Lobar segmentation was performed by co-registering the MNI atlas 3 . Lateral ventricles (LV) were identified manually. DT1 times were defined as difference between hyperoxic (6:15-8:45 min) and initial normoxic period (2:30 min) and normalised to DPETO 2 . Results: 2 acquisitions were excluded due to technical issues. Figure 1 shows an example T 1 time series and gas trace. Table 1 displays baseline T 1 and mean DT1. All changes, except PETCO 2 control LV CSF, are significant at p \ 0.05 level. Non-lateral ventricle CSF (non-LVCSF) shows the greatest DT 1 with and without PETCO 2 control. All GM regions show similar DT 1 (Fig. 2 ). Mean ± SD PETCO 2 was 38.3 ± 1.8 mmHg with control and 33.7 ± 5.4 mmHg without. Discussion: Although GM and WM regions showed statistically significant DT 1 with hyperoxic challenge, the magnitude was relatively small compared to non-LVCSF which shows an absolute DT 1 in the region of 140 ms. This is in keeping with previous studies that only found hyperoxic challenge DT 1 significant in non-LVCSF 4 . PETCO 2 control results in smaller DT 1 for both GM and WM. Using a two compartment model 5 we suggest that a reduction in blood volume due to uncontrolled PETCO 2 results in a relative increase in tissue volume fraction, increasing DT 1 due to the shorter T 1 of tissue compared to blood. In non-LVCSF, DT 1 is greater with PETCO 2 control which may be due to the lower blood volume in that space and partial volume effects. In conclusion, hypocapnia does affect oxygen induced brain T 1 changes at 3 T. Poole Hospital NHS Foundation Trust, Poole, GB Introduction: Cerebrovascular reactivity (CVR) reflects the mechanism of cerebral blood vessels to adjust their calibre and cerebral blood flow (CBF) in response to vasoactive stimuli (e.g., hypercapnia) 1,2 . This parameter has been shown to be impaired in several pathologies. In this work, we compared CVR and CBF parameters derived from BOLD and multi-PLD ASL, using different modelling strategies and focusing on temporal features. Methods: 10 healthy subjects (5 M, 20.4 ± 0.8 years old) were studied on a 3 T Siemens Prisma scanner, with a multi-PLD pCASL sequence (2D multi-slice GE-EPI, TR/TE 4100/14 ms, PLDs 250/500/750/1000/1250/1500 ms), acquired during normocapnia/hypercapnia (* 7 min each), and a pCASL acquisition with a double-excitation (DEXI) readout during 5/4 interleaved blocks of air/hypercapnia (* 9 min, TR/TE1/TE2 6100/14/35 ms), from where BOLD data was extracted. Hypercapnia was achieved using the RespirAct Gen 3 (Thornhill Medical, DPETCO 2 * 8 mmHg). All data was analysed using FSL and MATLAB. ASL data was modelled using a standard kinetic model (BASIL) and adding a gamma-shaped dispersion kernel 3 (M disp ) and/or intravascular arterial component 4 (M art ). BOLD model fitting was performed using a GLM approach (box-car regressor convolved with gamma function, std 6 s, mean lag 15 s), yielding CVR BOLDWBopt (CVR = DCBF/DPETCO 2 ). Additional GLMs with time-shifted regressors (-10:30 s, in steps of 1 s) were performed in order to account for different temporal BOLD dynamics across the brain, selecting the shift that provides the best fit for each voxel (Delay BOLDVOXopt , CVR BOLDVOXopt ). Statistical testing included voxelwise group analysis (FSL's randomise), paired T-tests and linear regression fitting (Matlab fitlm function). Results: When comparing BOLD optimisation strategies, we observed significant differences across all regions tested (Fig. 1 ). Voxelwise analysis showed significant differences across the brain, in particular where delays were larger (e.g., WM, occipital area). Significant correlations were only observed when comparing multi-PLD ASL CVR (Fig. 2 ) and BOLD CVR with delay optimisation in GM (Fig. 3) . No significant correlations were observed for bolus arrival time and/or Delay BOLDVOXopt . Discussion: Our results suggest that BOLD delay-optimisation can influence CVR quantification in specific brain regions, improving BOLD CVR correlation with multi-PLD ASL CVR in GM, regardless of the ASL modelling strategy used. Additionally, there is a trend towards a significant correlation between baseline CBF and BOLD CVR. Further work is still required in order to fully understand the physiological meaning of CVR temporal dynamics. Acknowledgements: EPSRC grants EP/G004277/1, EP/K025716/1, EP/S021507/1. 1. Liu, P. et al., Neuroimage 187, 104-115(2019 ). 2. Pinto, J. et al. Frontiers in Physiology 11 1711 ). 3. Chappell, M. A. et al. Magn. Reson. Med. 69, 563-70(2013 ). 4. Chappell, M. A. et al. Magn. Reson. Med. 63, 1357 -65(2010 Introduction: The popularity of resting-state (RS) functional (f)MRI is rising in mouse models with still open questions on the effect of anesthetics on brain state and neuronal network dynamics. Recently, the combination of low-dose Isoflurane (ISO) and Medetomidine (MED) has become an advocated anesthetic regimen in rats 1 . However, uncertainty resides over how fast relevant parameters adapt. Therefore, we investigated the brain state and -networks of mice during the transition from a single anesthetic (ISO) to a combined anesthetic regimen (ISO/MED) using a multimodal approach with RS-fMRI and calcium recordings. Methods: 12 female C57/BL6J mice were used in this study to investigate brain states under different anesthetic conditions. MRI measurements were conducted on a 9.4 T Bruker Biospec 94/20 small animal scanner using a CryoProbe (Bruker Biospin). Functional MR acquisitions were recorded with a GE-EPI sequence (TR/TE:1000/ 18 ms, 0.5 mm slice thick., FOV 28 9 26, matrix 80 9 80). The first RS-fMRI was acquired under 1% ISO anesthesia, subsequently, MED infusion (0.1 mg/kg, 0.2 mg/kg*h) was started and ISO reduced to 0.2%. 3 further RS-fMRI scans were acquired following MED bolus. Results: Acute calcium transients occur at \ 0.5 Hz (1A, green) under ISO in S1HL. Shortly after MED infusion their frequency increases towards 0.5-1.0 Hz (1A,B) and become constant after 35 min (1C). A spectral analysis of spontaneous BOLD activity in S1HL consistently shows decreased low-frequency events throughout the change of anesthetic regimen (1D). RS-network analysis show constant network efficiency and clustering throughout different anesthetic conditions (2A,B). Connectivity strength and coherence within functional groups was also not affected (2C,D). Further, analysis of functional connectivity (FC) shows significantly altered components of ISO and ISO/MED 45 min with ISO/MED25 min, respectively ( Fig. 4) , thereby indicating that the transitional period between anesthetic conditions is especially vulnerable to changes in FC. Discussion: Calcium and RS-fMRI recordings clearly show a brain state with slow-wave activity under ISO (UP-DOWN transitions) that rapidly changes to a persistent state upon switching the anesthetic regimen to ISO/MED in the S1HL. These changes were paralleled by alterations in the pattern of FC. Specifically, the coupling between the sensory cortex, association cortex, and the limbic system changed during the transition period. Therefore, we recommend a 40-min waiting period after switching from a single anesthetic regimen with ISO to combined ISO/MED anesthesia when studying the somatosensory cortex with RS-fMRI in mice. Test-retest reliability of resting-state correlations between the default mode network fMRI activity and EEG features Introduction: Simultaneous EEG-fMRI acquisitions are used to exploit the highly complementary characteristics of the two modalities. The default mode network (DMN) is robustly identified in resting-state fMRI studies, while its electrophysiological correlates remain incompletely understood. Previous work on the topic focused on finding patterns of correlation between the fMRI DMN and EEG features that were consistent between subjects [1] , but reliability studies assessing the consistency of these correlations within subjects are lacking. Here we investigate the test-retest reliability of restingstate correlations between the activity of the DMN measured by fMRI and multiple EEG features. Methods: EEG-fMRI data was acquired from 23 healthy subjects, on a Siemens 3 T MR scanner, during 3 sessions of 10 min eyes-closed rest [2] with GRE-EPI (TR/TE = 2000/50 ms, 3 mm isotropic). EEG was acquired with a 62-channels MR-compatible system (BrainAmpMR). The fMRI DMN was identified by group-level ICA, followed by template matching [3] . Its activity time-series was extracted as the fMRI signal of interest. From the EEG time-frequency spectrum, the following features were derived: i) Univariate features: root mean squared frequency (RMSF), total power (TP) and band-specific power; and ii) Functional connectivity (FC) features: a dynamic FC network was built for each frequency band by estimating the imaginary part of coherency (ICoh) between each channel pair, and the weighted node degree (WND) was then computed for each channel. Bands considered: d (1) (2) (3) (4) , h (4-8 Hz), a (8) (9) (10) (11) (12) . The resulting EEG feature time-series were convolved with a set of double-gamma hemodynamic response functions (HRF), with overshoot delays of 10, 8, 6, 5, 4 and 2 s. Pearson''s correlation was computed between the time-series of each EEG feature and the fMRI DMN time-series. The test-retest reliability of the correlations was evaluated with the intraclass correlation coefficient (ICC) [4] , which estimates measurement consistency across the 3 sessions. Results: The estimated ICC values are presented in Figs. 2 and 3. EEG-DMN correlations were significantly more reliable for univariate than for FC features. Correlations were also significantly less reliable for alpha than for the remaining frequency bands, and less reliable for the temporal than the remaining lobes. Regarding delays, reliability was highest at 4 s. In evaluating the test-retest reliability of the resting-state correlations between the fMRI DMN and EEG features, we found significant effects with respect to the EEG metric, frequency band, channel lobe and delay. Despite global mean values of reliability being poor (ICC \ 0.04), several EEG features showed fair reliability (0.41 \ ICC \ -0.59) [5] . Percuros B.V., Leiden, NL Introduction: In recent years low field MRI has received much attention due to its low cost and portability [1] [2] [3] , as well as the possibility of its use in unconventional locations without an RF shielded room. However, this means that it is necessary to cancel any electromagnetic interference (EMI) created by equipment close to the scanner. Many different methods to reduce EMI for MRI and NMR have been explored in the past [4] [5] [6] [7] . In this work, we have used an MRinactive RF coil, which detects only the EMI. A noise cancellation system has been implemented on a Halbach-based low field MRI system operating at 2.15 MHz 1 . Images of a leg phantom have been acquired with and without EMI. Methods: A solenoid was used as the transmit coil ( Fig. 1 ). Two receive saddle coils (Rx coils) with orthogonal magnetic fields were designed and simulated by using a sum of sinusoidal function over the desired cylinder. The outer saddle (EMI detector) coil with a 17.3 cm diameter and the inner saddle coil with a 15.6 cm diameter were positioned such that B 1 aligned and perpendicular with B 0 , respectively. Both Rx coils have the same length of 20.5 cm (Fig. 1) . Because of the close proximity, coupling between Tx and Rx coils is strong, detuning circuitry was used on each coil for this purpose. The 90°rotation between the two saddle coils resulted in intrinsic decoupling. The two receive coils were connected to a 180°power combiner with 0.2 dB insertion loss, 0.5 s phase unbalance, and -25 dB inner-channel isolation. A battery-operated drill motor was used as a broadband EMI source. Body providing a path by which EMI is transmitted into the imaging field-of-view, images were acquired with a leg phantom consists of water doped with 1.5 g/l NaCl and 1.3 g/l H2KO4P with 90 cm length and 11 cm diameter. A turbo spin-echo sequence was used. Results: The simulated B 1 fields of the two saddles are shown in Fig. 2 : the magnetic field vectors of both Rx coils are almost orthogonal over the cylindrical region of interest of 10 9 16 cm. Reflection coefficients (S 11 ) for each of Rx coils and coupling between two Rx coils were lower than -22 dB, which are the same as measurement data with * 1% error. Fig. 3 shows image data with and without EMI. The results in Fig. 3c -e show * 80% reduction in EMI. Conclusion: Significant reduction in EMI has been achieved using the new design which does not require external sensors. This method of EMI reduction results in an intrinsically reduced SNR compared to an image acquired without EMI, due to the Johnson noise added from the EMI coil. EMI reduction is also affected by the slightly different sensitivities of the two receiver coils which have similar but not identical diameters. If two independent receive channels could be used, then a weighting function could be introduced to correct for this, and improve the EMI. Introduction: Receive coils play an essential role in MR imaging as the initial step in the receiving chain to pick up the RF-signal. They are usually tuned to the Larmor frequency by discrete capacitors placed on the circumference of the coil. An alternative to this is the use of Split-Coils consisting of partially overlapping microstrip transmission lines [1, 2] to eliminate the need for rigid solder joints along the coil thus providing flexibility for non-flat surfaces. At the same time unconventional manufacturing methods such as printing or cutting plotters become practical, along with the possibility of integrating coils into fabrics. In this work, the applicability of the design will be further investigated to determine whether its adjustment by overlap can be used for various clinical field strengths (0.55-3 T) . Methods: A Split-Coil is simulated by full-wave method (CST, Dassault Systèmes). Figure 1 shows a top view as well as a cross section of the coil with the overlaps u and u port , center radius b = 67.5 mm and trace width w = 5 mm. A dielectric is chosen with a height of h = 0.1 mm and rel. permittivity of e r = 3. The coil is loaded with tissue imitating medium (rel. permittivity e r = 80, conductivity r = 0.46 S/m) on one side at a distance of 5 mm. The front and back side traces are symmetrical. For validation, an exemplary coil was built and compared with simulations based on its input impedance considering de-embedding. A matching capacitor C M is connected in series to compensate the imaginary part of the input impedance. This is done discretely during post-processing to focus on the coil tuning but can be realized in practice by additional overlapping traces to form C M . Results: The validation measurement, shown in Fig. 2 , demonstrates high agreement with less than 2 MHz deviation in the inductive range between the extrema. The coil shown in Fig. 1 can be tuned to the specified target frequencies by varying only its overlaps. Figure 3 shows the S 11 curves for the coils, as well as their B 1 ? distributions. Higher quality factors and more homogenous field strengths are seen with lower frequency. The required overlaps for tuning and the matching capacitance are listed in Table 1 . Discussion: This work demonstrated frequency adjustment by overlap of Split-Coils. The simulations were based on validation measurements with real coils. The coils can be tuned with uniform geometry only by varying the two overlaps to field strengths of 0.55, 1.5 and 3 T. Simulations of the B 1 ? -fields show typical sensitivity profiles. This increases the range of application for Split-Coils to clinically relevant field strengths and can be expanded upon by varying the size of the coil, trace or dielectric for further optimization. [1] which means minimizing coil losses (and maximizing the Q factor) is vital in order to maximize SNR at low field. A high Q brings challenges too, the coil bandwidth is given by f 0 /Q which for low field can be much less than the imaging bandwidth. Commonly a resistor is used to lower the Q factor [2, 3] which reduces SNR (SNR µ HQ). Previous works has shown that an impedance-mismatched preamplifier can reduce the coupling between array coil elements without introducing additional losses [4] . In this work we show that a high impedance preamplifier can overcome the bandwidth limitations of high Q coils at low field without sacrificing SNR. Methods: Two 15 cm long, 15 cm diameter solenoids with 45 turns of copper wire solenoidal RF coils were constructed. The high Q conFiguration was constructed using 1.5 mm diameter copper wire, the ''standard'' imaging coil used 0.8 mm diameter wire and an 8X resistor. The Q factors of the high Q and standard coils, measured as 2f 0 /(Dx3dB) of the s 11 parameter, were 452 and 179 giving a bandwidth of 9.5 kHz and 24 kHz respectively. The preamplifier is a two stage preamplifier based on a On Semiconductor 2SK3557 JFET with 30 dB gain with the output impedance matched to 50 X (see Fig. 2 ). The coil is impedance matched to 50 X and connected directly to the gate of the JFET which has a very high (* 1.2 kX) gate to drain impedance. An active T/R switch based on the Analog Devices HMC545AETR was used to interface the coil to the RF power amplifier and preamplifier. Data were acquired using a Halbach based MRI scanner operating at 2.15 MHz [3] , using a 3D turbo spin echo sequence on a 15 9 15 9 30 cm cylindrical phantom with the following parameters: Field of view: 200 9 200 9 300 mm, resolution: 2 9 2 9 5 mm, TR/TE: 200 ms/15 ms, echo train length = 6, acquisition bandwidth: 30 kHz, scan duration: 3 min 20 s. Results: Figure 2 shows the gain and input impedance of the preamplifier. The noise Figure of the preamplifier is around 3 dB. Figure 3 shows 4 images acquired with the two coils connected to a 50 X preamplifier and the high input impedance preamplifier. Signal loss away from the center is reduced for the high Q with the high impedance preamp while maintaining the SNR benefit over the standard coil connected to the same preamp. Discussion: In this work we have shown that a high input impedance preamplifier can be used to overcome the low bandwidth associated with high Q factor coils at low field. Improvements to the currently high noise Figure of the preamp and further increasing the Q factor of the coil can yield significantly higher SNR compared to current approaches of coil construction at low field. Impedance mismatched preamplifier have also shown promise at reducing inter-element coupling in low field coil arrays [5] and will be investigated in future work. Introduction: In the past 20 years metamaterial (MM) science has allowed the design of novel RF coils suitable to improve MRI applications [1] [2] [3] . Moreover, the use of high-performance dielectric materials has been widely used in the context of EPR, NMR, and MRI with several functions, including dielectric resonators (DRs) and RF shimming pads [4] [5] . Only recently the two paths converged with DRs proposed as an add-on RF hardware tool, useful to mimic a negative permeability MM and enhance SNR/SAR of single or multiple channel RF surface coils [6] [7] [8] [9] . Here, we show that a smallsize high-permittivity homogeneous dielectric sphere, selected to operate at a specific Mie resonance, can accurately mimic the EM field outside the sphere produced by magnetic localized surface plasmons, generally supported by a negative permeability sphere of the same diameter. Methods: To confirm the analytical result, valid for the equivalence of lossless MM and dielectric, full-wave EM calculations at 3 T (127.7 MHz) were done taking as a model a low-losses [tand = Im(e DR )/Re(e DR ) = 5 9 10 -3 ] DR sphere (diameter of 6.8 cm; tuned at the first resonant mode, L = 1) having a permittivity of e DR % 1200 ? i48 (l DR = 1) positioned in front of a standard circular RF coil (8.4 cm diameter) , this being at 2 mm from a large cylindrical sample (diameter and length 25 cm; e SAMPLE % 64 ? i100; l SAMPLE = 1). Results: Our combined DR sphere, RF coil, and sample EM model show a strong mutual coupling with an observed maximum SNR enhancement of about 2.7, and a reduction of the peak SAR10g of about 22%, within 2 mm from the sample surface. Higher-order modes of the DR sphere, having the same e DR , will give a better improvement (L = 3: SNR = 3.3, diameter DR 12.4 cm), although with a reduced penetration depth within the sample. Our theoretical model shows that the L = 1 DR sphere is equivalent to a MM sphere having l MM % -1.98 ? i0.40, e MM = 1. Discussion: These numerical results at 3 T, seem to anticipate useful applications of DRs for UHF MRI applications. It is worth noting that there are practical advantages in using a DR sphere instead of a negative permeability MM one, thanks to the availability of low losses dielectrics, such as ceramics based on lead zirconate titanate [5] . Our approach holds at extremely near-field scales since it is not based on an effective medium theory and the considered structure does not exhibit spatial subwavelength inhomogeneities as for MM based on macroscopic arrays of LC resonators. [1] Wiltshire MCK, et al., Science 291, 849 (2001) . [2] Freire MJ, et al., Appl Phys Lett 93, 231108 (2008) . [3] Lezhennikova K, et al., Phys Rev Appl 13, 064004 (2020) . [4] Webb AG, Conc Magn Reson B 38, 148 (2011) . [5] Rupprecht S, et al., Magn Reson Med 79, 2842 . [6] Rizza C, et al., Phys Rev Appl 12, 044023 (2019) . [7] Rizza C, et al., Phys Rev Appl 14, 034040 (2020) . [8] Rizza C, et al., J Phys D Appl Phys 54, 165108 (2021) . [9] Rizza C, et al., WO 2021 /014354 A1 (2021 Introduction: With up to 130 dB sound pressure level (SPL), acoustic noise is one of the main sources of patient discomfort in MRI [1] . In the current commercial systems, the SPL of a scan noise p(t) is estimated by using linear time-invariant models, as a function of the gradient input derivative g'(t). In frequency domain, this is expressed via the transfer function H(f) = P(f) / G'(f) for each gradient coil [2] [3] . In this work, the model and the limits of the linearity assumptions are revisited. Methods: Pulse sequences using triangular gradient pulses were used on a 3 T system (Philips Ingenia) in four different modes: X, Y, Z single gradient inputs and simultaneous triple coil input. Data was averaged over 5 pulses and includes 21 ramp up times t sl = 0.10, 0.12, … 0.50 ms and 27 amplitudes g A = 1, 2, … 27 mT/m with a repetition time TR = 0.5 s. All noise signals were recorded at 44.1 kHz sampling rate using an optical fibre microphone (Phonoptics) and filtered to 0.3-6 kHz. To study the linearity of the scaling, root mean square (RMS) sound pressure was calculated for all g A and t sl in the four modes. The correlation coefficient with the best linear fit was then estimated for all t sl . To study the linearity of superposition, noise power from multiple gradient coils (4th mode) was compared to the individual gradient coils. Results: Signal scaling plots in Fig. 1 (t sl = 0. 1, 0.3, 0.5 ms) show that RMS sound pressure response is fairly linear at low g A (below 15 mT/m) but becomes noticeably saturated as the system slew rate limit is approached. The system deviates from the linear model when rampup times below 0.2 ms are used, resulting in the overestimation of the sequence SPL (Fig. 2 ). Minor deviations in the noise estimation are observed when gradients are played in superposition (Fig. 3 ). An overall increasing overestimation compared to the measured noise P meas is observed for increasing g A and at certain ramp-up times (Fig. 3a) . In contrast, when the transfer function H(f) is estimated from simultaneously played gradients as an input, minimal difference in SPL is observed (Fig. 3b ). Discussion: The linear time-invariant model shows substantial overestimation at high gradient amplitudes and high slew rates for single gradients. Notably, the Z gradient has an overall higher correlation coefficient compared to X and Y, which could be attributed to different coil geometry. Additional overestimation occurs as gradients are played out simultaneously, showing that single gradients may also induce a substantial force in the neighbouring coils. More advanced models may help to better characterize the realistic noise burden and aid the development of silent MRI sequences. Tesoro Imaging S.L., Valencia, ES Introduction/methods: Peripheral Nerve Stimulation (PNS) constrains the clinical performance of Magnetic Resonance and Particle Imaging (MRI and MPI) systems. Extensive magneto-stimulation studies have been carried out recently in the field of MPI [1] , where typical operation frequencies range from single to tens of kilo-hertz. PNS literature is scarce for MRI in this regime, which can be relevant to low-field dedicated MRI setups [2] , and where the resonant character of MPI coils prevents studies of broad-band excitation pulses. At a more fundamental level, stimulation models known to be accurate for magnetic excursion times above hundreds of microseconds [3] have been less thoroughly tested for shorter timescales. Some authors even question the validity of these models at extreme frequencies [4] . All in all, there is a general lack of PNS data relevant to medical imaging applications exploiting rapid magnetic field dynamics. We have constructed an apparatus designed for PNS threshold determination on a subject's limb ( Fig. 1) , which allows for fast narrow and broad-band excitation pulses (excursion times \ 40 us), and can be conFigured for different spatial magnetic field strength distributions. Results/discussion: From measurements on 51 volunteers ( Fig. 2 left and middle), we observe that PNS limits coincide for sinusoidal (biphasic narrow-band) and triangular (biphasic broad-band) excitations, and are slightly lower for trapezoidal (monophasic broad-band) pulses (not shown in the Figure) . The observed dependence on pulse frequency/rise-time is compatible with traditional stimulation models where nervous responses are characterized by a rheobase and a chronaxie [3] . In a second set of experiments ( Fig. 2 right) , we have confirmed thresholds increase significantly as trains transition from tens to a few pulses. By changing the polarity of the coils in our setup, we also looked at the influence of the spatial distribution of magnetic field strength on PNS effects. We find that thresholds are higher in an approximately linearly inhomogeneous field (relevant to MRI) than in a rather homogeneous distribution (as in MPI). Finally, in Fig. 3 we report the results of testing the influence of the arm position on which nerve fibers are triggered, as well as the strength of their response. As shown in the Figure, the location and intensity of the stimulation depends strongly on the exact subject position and field conFiguration. Outlook: Given the large intersubject variability of PNS sensitivity, we propose employing a versatile low-cost system (such as presented here) for fast offline determination of a subject's limits prior to medical scanning, and then using this information to boost clinical imaging while preserving the patient's safety. S6 .O1. Fast T1 and T2* mapping of tissues with very short transverse relaxation times and application for tissue segmentation of the knee 1 demonstrated that a bivariate histogram of the longitudinal (T1) and effective transverse relaxation times (T2*) allows to segment different knee tissues within a total acquisition time (TA) of 22-63 min, using a triple echo method for T2* mapping and either variable flip angle or variable repetition time method for T1 mapping 1 . We provide a fast approach for high-resolution, isotropic T2* and T1 mapping within 9 min by using a double echo sequence 2 for T2* mapping and a mixed approach relying on two flip angles (FA) and two repetition times (TR) for T1 mapping. Exemplarily, this approach has been applied to segment various tissues types of the knee. M: Measurements were performed on a 3 T MR-scanner (MAGNE-TOM Skyra, Siemens Healthcare, Erlangen, Germany) using a 15-ch. transmit/receive knee coil and a prototypical 3D spoiled gradient echo UTE sequence with stack of spirals readout. Two measurements with an isotropic spatial resolution of 0.8 mm were performed (M1: TE1/ TE2 = 0.04/4.92 ms, TR1 = 9.92 ms, FA1 = 16°, TA = 6 min; M2: TE1 = 0.04 ms, TR2 = 4.45 ms, FA2 = 3°, TA = 3 min). The T2* map was calculated based on the two echoes obtained in M1. A T1 look-up- Table ( LUT) to allocate the ratio of two gradient echo signals with identical TE has been computed using discrete 1 ms steps ranging from 1 to 2500 ms. T1 was obtained by applying the LUT to the ratio of the first echo of M1 and M2 (Fig. 2) . The proposed T1 and T2* mapping was compared with the inversion recovery method and multi-echo gradient-echo method, respectively, by measuring plasticine and solutions with different agar and Gadolinium concentrations (Fig. 1) . The T1 and T2* maps were converted into a bivariate histogram to identify tissue specific T1/T2* clusters by using previous literature values2 based on which the T1 and T2* thresholds determined various tissues. R: Workflow, images and T1/T2* maps of a healthy volunteer are shown in Fig. 2 . The bivariate histogram, presented in Fig. 3 , reveals tissue specific clusters. Applying the histogram-based classification allows differentiation of bone marrow and patellar tendons ( Fig. 3B -D). D: T1 and T2* values of the knee tissue as well as pixel count peaks of the bivariate histogram obtained in a clinically appropriate TA of 9 min are in line with findings by Krämer et al. 1 . The 3D-visualization of bone marrow and the patellar tendons shows segmentation of these tissues opening the door to quantitative lesion characterization in bone or tendons. We intend to apply our approach to a larger cohort (controls and patients) to optimize segmentation thresholds for the knee and validate the outcome with manual segmentations. Our approach might also be applied for other joints such as the ankle, elbow or shoulder. Hospices Civils de Lyon, Department of Radiology, Lyon, FR Introduction: Neoadjuvant chemotherapy (NAC) can be proposed to patients with locally advanced breast cancer to reduce tumors sizes before surgery and improve breast conservation rates. Predicting the response to NAC using preoperative MRI could improve the delivery of a personalized therapy. Radiomic models, based on features extracted from lesions, have shown promises in predicting pathological complete response (pCR) but as radiomic features (RF) values depend on each scanner, models are weakly transferable. This work intends to develop a robust multi-scanner radiomic model to predict pCR to NAC. Methods: A retrospective study included 103 breast cancer patients, of wich 49 responders, scanned before NAC using T2 and T1weighted DCE modalities. Twenty-five were acquired on a GE Optima MR450w (1.5 T) machine with an 8-channel coil and 78 on a Siemens MAGNETOM Aera (1.5 T) machine with an 18-channel coil (n = 19) or with a 16-channel Sentinelle coil (n = 59). Based on [1] , 3D images were corrected for bias field, spatially resampled using B-spline interpolation and z-score normalized with mean values and standard deviation computed in the breast outside the tumors manually segmented. A total of 2010 RF were extracted from each lesion using Pyradiomics. For robustness, RF were harmonized between the 3 MR coils using ComBat [2] . For 30 patients, lesions were segmented by two radiologists and two-way intraclass correlation coefficients (ICC) calculated between RF extracted from each radiologist's region. RF selection included a three-step procedure: 1) RF with an ICC [ 0.8; 2) lower bound of AUC of ROC curve [ 0.5; 3) Boruta algorithm after the z-score normalization of RF. Random forest models were then trained and evaluated using 10 repetitions of leave-one-out cross-validation on 3 experiments: T1-based RF (E1), T2-based RF (E2), T1 plus T2-based RF (E3). Performances of the models were evaluated using the Youden Index (Y = sensitivity ? specificity-1) to predict pCR. Finally, a permutation test was conducted 100 times, by randomly assigning a response label to each patient while respecting the proportion of responders to non-responders in each scanner. Results: Regarding RF selection, 946 T1-RF and 884 T2-RF displayed an ICC [ 0.8 between the two radiologists' segmentations. The harmonization by ComBat led to a greater number of RF selected by univariate analysis (Fig. 1 ) and better performances for each experiment (Fig. 2) . The best predictive model (E3) yielded a Youden Y = 0.49 ± 0.02 (Fig. 2) , outperforming any predictive models trained on reshuffled response labels in permutation tests (Fig. 3 , p \ 0.001). Discussion: ICC selection spotted features robust to slight variations in segmentation shapes while harmonization by ComBat enabled models to be robust to the different scanners. The combined pipeline ensured the construction of a robust and transferable model to predict pCR. Introduction: The trade-off between spatial and temporal resolution in dynamic MRI is a hindrance for MR-guided interventions, which require high temporal resolution while visualizing details. Deep learning based super-resolution (SR) has shown promising results in dealing with this trade-off [1] . Nevertheless, the available temporal information of dynamic MRI has not been exploited in this prior work. The potential of improving the reconstruction quality of dynamic MRIs by incorporating the temporal information has been demonstrated recently [2] [3] [4] . This work extends the previous work by ameliorating the previous model [2] a the dual-channel (static ? dynamic images) super-resolution approach, termed DDoS (Dynamic Dual-channel of SuperRes). Methods: An artificial dynamic dataset was generated by applying random elastic deformations [5] to the publicly available CHAOS dataset (T1 in-and opposed phase) [6] . Then, a low resolution dataset was simulated by performing in-plane undersampling [7, 8] taking only the centre of the k-space of each slice. A modified U-Net model [1] was trained with dual-channel input, consisting of the low-resolution image of the current time point (LR_TP n ) and the high resolution image of the previous time point (HR_Tp n-1 ), see Fig. 1 . This training strategy tries to let the network learn the spatio-temporal relationship over time points, with the help of perceptual loss [9] using a perceptual loss network [10] , and was minimised using Adam optimiser for 18 epochs. 3D patches of the volumes were created for training with a patch size of 24 3 , and with a stride of 6 for the slice dimension and 12 for the rest. For testing, a stride of 3 was used for all dimensions with the same patch size. A 3D abdominal dynamic MRI data was acquired at 3 T Siemens Magnetom Skyra [GRE, T1w Flash 3D, TR: 2.23 ms, TE: 10.93 ms, voxel size: 1.09 9 1.09 9 4.0 mm] and was also undersampled in the same manner and were used for testing. Results and discussion: The highest undersampling (only 6.25% of centre k-space per slice) explored in the former work [1] , was applied to evaluate the improvements with the proposed approach. Figure 2 shows the comparison of low-resolution input (for the lowest resolution examined), SR result after subject specific fine-tuning [1] , DDoS results and its corresponding ground-truth images. The average SSIM value while comparing the results against the ground-truth improved from 0.949 ± 0.003 (for SR result after fine-tuning [1] ) to 0.979 ± 0.004 and the average PSNR value improved from 34.647 ± 0.240 to 39.411 ± 0.536. One can observe qualitatively that the results of the DDoS model are more similar to the groundtruth compared to the result of SR after fine-tuning. In conclusion, this research illustrates that by incorporating the temporal information in a super-resolution approach [1] using the DDoS model can improve the reconstruction quality, and may be extended for application during real-time interventions due to the fast inference speed. Noise removal in line-scanning fMRI 4 Utrecht University, Utrecht, NL; 5 University Medical Center Utrecht, Department of Radiology, Utrecht, NL Introduction: By sacrificing spatial coverage of the brain outside the region of interest, line-scanning (LS) fMRI allows extremely high spatial and temporal resolution 1 , resulting in a more accurate characterization of the HRF across cortical layers. The small voxel size and very short TR (* 100 ms) lead to substantial amounts of thermal noise and the small area being scanned increases the sensitivity of LS to motion. In this work, we present a LS denoising technique based on Noise reduction with Distribution Corrected (NORDIC) PCA 2,3 . Methods: Four healthy volunteers were scanned at 7 T (Philips) using a 32-channel receive head coil (Nova Medical). The acquisition was based on a modified 2D ME-GRE sequence, with line resolution 250 lm, TR = 105 ms, TE1/DTE = 6/8 ms, 3200 time points, array size 720 and line thickness 2.5 mm. Fat suppression was performed using SPIR. Two OVS pulses (7.76 ms) were used to suppress the signal outside the line of interest in the visual cortex (Fig. 1) . The phase-encoding gradient perpendicular to the line was turned off 4 . -One run of functional data was acquired for each volunteer with block design visual task, consisting of a 20 Hz flickering checkerboard (10 s ON/OFF, n = 16). Offline reconstruction using Matlab and Gyrotools included a denoising step based on NORDIC prior to coil combination. A singular value decomposition (SVD) of the data was submitted to hard thresholding to eliminate all components indistinguishable from zero-mean gaussian noise 2 , followed by a tSNR and coil sensitivity-weighted sum of squares coil combination. Reconstruction with and without denoising were compared (tSNR, t-values and PSC) within a 10-voxel ROI for 4 thresholds (20, 50, 70 or 90% of singular values were set to zero for noise removal). Results: Figure 2 shows LS-fMRI data for the original (panel a) and denoised data, after the removal of 50% (panel b) and 90% (panel c) of the SVD components. Figure 2d shows the resulting timecourses averaged over the 10-voxel ROI. Figure 3 shows the tSNR and PSC for the different thresholds for an example subject and the distribution of t-values across the line for all four subjects. Overall, more voxels with higher t-values were found in all volunteers. Discussion: Our results reveal that NORDIC PCA improves the quality of LS-fMRI data, making its application more feasible. The signal outside the brain was visibly reduced for both thresholds, while the BOLD response to the visual task was maintained in the time series (Fig. 2) . By removing the thermal noise, NORDIC facilitates the removal of cardiac and respiratory components 3 , thus being valuable for physiological noise removal. These outcomes can be further enhanced by assessing an ideal numerical threshold and by optimizing the line placement during planning 5 . Introduction: Dual flip angle R1 mapping-collecting two FLASH volumes with different flip angles (a1 and a2) and (potentially) different repetition times (TR1 and TR2)-is a workhorse of proton density (PD) and R1 mapping in the in vivo human brain [1] . An efficient method for computing R1 and PD from the two volumes, implemented in the hMRI toolbox (hmri.info), uses a rational approximation of the Ernst equation derived under the assumption of short TR and small a [1, 2] . However, the assumption of small a can break down at 7 T, where B1 inhomogeneities are large [3] . Here we present and validate a modified rational approximation that is valid even for large a. Methods: The derivation is presented in Fig. 1 . First, we make a halfangle tangent substitution into the Ernst equation (Eq. (1)) to simplify the trigonometric functions of a without approximation (Eq. (4)) [4] . We then make the assumption TR is small such that R1 TR is small. Because the Ernst equation is a rational function, we use rational Padé approximants [5] to expand around TR = 0, then solve for R1 and PD. We tested the method in simulations and on a 7 T in vivo dataset. To demonstrate that the result is non-obvious, the simulations also compared the result of simply inserting the linear expansion of exp(-R1 TR) into Eq. (4). Results and discussion: Figure 2 shows the benefits of the novel approximation for a simulated 7 T acquisition. Figure 3 shows these benefits also extend to a typical experimental dataset. Formally, our result is the specific dual-angle case of an N flip angle result suggested (but not evaluated) in [6] . The left panel of Fig. 2 shows that it is important to always evaluate the numerical accuracy of such results, as substituting linear approximations into rational functions is not guaranteed to give good results. We retained the assumption of short TR. While a closed form exact solution without approximation is possible when TR1 = TR2 [4] , this is not generally applicable. Nonlinear fitting can be used to estimate R1 and PD from Eq. (1), but will be very slow. We omitted magnetization transfer and imperfect spoiling in our analysis, which also increase with increasing flip angle [7, 8] . However, Fig. 3 implies that our method still removes a significant amount of bias. This method has been implemented into the hMRI toolbox (hmri.info) and will be included in an upcoming release. This will allow others to efficiently extract more precise R1 estimates. advanced brain segmentation tools, such as FreeSurfer, 4 have become more robust, segmentation errors still might occur. Therefore, manual quality assurance is still indispensable, which however becomes difficult and time consuming at a larger scale. Here, we present a tool for fast, explorative and interactive quality assessment optimally suited for larger cohorts, combining the visualization of a segmentation output along with the numeric outcome of the resulting dataset. Materials and methods: QCFlex was developed under Python 3.9.5 with the Python packages Numpy, Matplotlib, pandas, Pillow, qim-age2ndarray. The GUI was built with PyQT5 and can be distributed as an executable. A single table file builds the basis of variables to explore and the location of corresponding image paths. The interactive image viewer displays a precomputed high resolution 2D image of the resulting segmentations, which can be inspected in detail by zooming into the region of interest. The integrated comment feature allows the annotation of images and storage of those comments in the table file for later use. The interactable scatterplots, depicting two selectable outcome-measures (such as brain volume and gray matter volume), allow an instant overview of the data distribution. This feature facilitates the detection of outliers, which can be further inspected by clicking on the datapoint. Buttons to annotate the images with Pass/Fail further supports the user's ability to assess the overall quality of the calculations, while still being able to keep track of outliers. Features illustrated at the exhibit: The tool will be presented based on a precomputed dataset of brain segmentations processed by FreeSurfer (v.6) . Moreover, we will explain, how QC Flex can be adapted for different purposes, where imaging data is used to generate numeric outcome measures and where quality assessment can be performed by a 2D image. Introduction: In vivo monitoring of transgene expression in the brain after gene therapy remains a complex task. Chemical Exchange Saturation Transfer (CEST) is an emerging imaging modality that could be used to achieve this goal [1] . An Adeno-Associated Virus (AAV) expressing an arginine rich-protein reporter (polyARG) has been designed, with frequency-selective contrast as well as tdTomato (a fluorescent protein). To validate the use of the CEST reporter (polyARG), we need precise co-localization of the AAV with anatomical ground truth (histological slices series). In this abstract, we propose an automated method to identify and register the best histological slice candidate corresponding to a single CEST image. Methods: An AAV-PolyARG-tdTomato was injected in the left striatum of 2 wild-type mice. For each mouse, a single CEST slice (0.13 9 0.13 9 1 mm 3 resolution), as well as its corresponding anatomical MRI (A), were acquired in the striatum. Acquisition of 25 anatomical MRI sections (R) was performed around the CEST signal acquisition location (0.07 9 0.07 9 0.3 mm 3 resolution). Finally, 12 histological slices (T) were cut and digitized in the same region (0.03 9 0.03 9 0.2 mm 3 resolution). To identify the best T b candidate among T, matching A, and given intensity bias in A, we chose to first identify the best R b section in R, and use it as a reference to identify T b . After preprocessing, we explored every possible 2D-2D registrations between all R sections and A using BlockMatching [2] , with first rigid and then affine registration, increasing degrees of freedom. For each pair of registered sections, Normalized Mutual Information (NMI) [3] quantified their similarity at every registration step (rigid, affine). Finally calculating the maximum of an equal contribution of both rigid and affine NMI values enabled the identification of R b corresponding to A (Fig. 1 ). Same steps were subsequently applied to identify T b corresponding to R b . Results: Results show that one single maximum NMI value stands out in both cases for T b identification (Fig. 2) . Paired image results were visually validated by experts and are presented in Fig. 3 . The proposed method demonstrated its ability to automatically identify and register the best histological slice corresponding to the CEST image. These results enable the evaluation of the AAV co-localization between a CEST image and its ground truth in the brain. This generic method makes it possible to associate one or several ground truth histological slices to a limited brain region covered by the CEST image. Introduction: While several arterial spin labelling (ASL) processing pipelines are freely available, programming skills are usually needed to efficiently process large datasets (1), deterring clinicians from analyzing ASL data. Leaving data unanalysed reduces the speed of progress in medical research and this unavailability keeps research restricted to a select few specialized academic institutions. These hurdles are addressed by ExploreASL, an ASL processing pipeline that analyzes ASL data on an individual and population level, and is free for non-commercial purposes (2) . To enable its use for researchers without any coding knowledge, a graphical user interface (GUI) was developed in collaboration with clinical radiologists to specifically address their needs (3) . Methods: This GUI wraps around ExploreASL (4), which can be downloaded for Linux, MacOS, and Windows (3), either as Matlab code, as compiled executable or as a Docker image. The GUI is divided into four steps ( Fig. 1 ): Step 1 will create directories needed for processing and save the data as NIfTI. Step 2 allows for precise definition of study parameters (sequence type, labelling method, M0 map availability, etc.), subject exclusion, ASL sequence parameters, and options for processing methods and masks. Step 3 runs ExploreASL and its structural, ASL, and population processing modules. It can easily be parallelized to workstation capacities, and processing steps can be selected and deselected at will, as well as a data rerun. Step 4, enables the user to create graphical presentations of the processed dataset, including options for different atlases, partial volume statistics, and parameters such as the coefficient of variation (Fig. 2) . The GUI launches an offline dashboard application (Fig. 3 ). Users can select various plot types, interactively set variables of interest, subset the data, and receive hover feedback. Users may also click on data points to load in corresponding images. Results: First experience with two programming-naïve clinical radiologists showed that successful data processing could be achieved, supported by online tutorials and manuals. The GUI was considered user friendly after the online material had been studied. Data processing quality was comparable with standard processing in ExploreASL. The clinical users considered the GUI a game-changer for their scientific work. Discussion: The ExploreASL GUI empowers clinicians to process their scientific ASL data. Thereby, it may democratize and accelerate medical research. The ExploreASL GUI is a multi-institution ongoing freeware project, developed by developers and users. 1 Here we propose a statistical model to correct this bias for quantitative applications of 19 F MR. Methods: We examine 11 in vivo datasets from 5 mice partially scanned on multiple days and ex vivo datasets from 5 different mice. PFCE nanoparticles were administered daily from 5d following EAE induction. 2 3D-RARE 19 F-MRI: TR = 800 ms, TE = 4.4 ms, ETL = 40, FOV = (45 9 16 9 16) mm 3 , (140 9 40 9 40) matrix. 32 m and 80 m (in vivo) or 128 m (ex vivo) scans were used as test and Reference data. Noise level r was estimated in a background region. Images were thresholded at SNR = 3.5 and \ 3 connected signal voxel features were removed as outliers. S t and S r denote the test and Reference signal after RNBC. 3 For true, but unknown signal S * and measured signal S m , the forward model is given by Rician distribution Ri(S m ; S * , r). We model S * as drawn iid from log-normal prior distribution LN(S * , h) (Fig. 1A) . Parameters h are estimated separately for each image by maximizing the likelihood of the measured signal under the marginal distribution p rh (S m ) µ $Ri(S m ; S * , r)LN(S * ,h)dS * (Fig. 1B) . We compute the corrected SI as the posterior mean following Bayes theorem (Fig. 1C ). Estimation bias S t -S r and S c -S r is shown for single voxels and features. In an example in vivo dataset, the SI is overestimated at all signal levels by on average 0.5r ( Fig. 2A) . The correction removes this bias without altering random spread. The same holds true when averaging over signal features. No difference between smaller and larger features is noticeable (Fig. 2B) . In contrast to the original, the corrected image contains a balance of overestimated and underestimated voxels (Fig. 2C ). Averaged over all datasets the overestimation in uncorrected in vivo data is slightly larger than in the example ( Fig. 3A/C) . Again, the average bias in the corrected data is close to 0. Analysis of the ex vivo data yields similar results ( Fig. 3B /D). Discussion: Systematic overestimation in low SNR MRI of 19 F-NPs can be explained and corrected by assuming that SIs are drawn from a heavily skewed distribution, e.g. a log-normal. The effect size is below the random uncertainty for individual voxels, but overestimation in uncorrected data and effectiveness of model-based correction persist when averaging over ROIs. Thus, correction is essential for quantitative conclusions. The correction can be applied without the need to change data acquisition protocols. It remains to be seen whether similar effects occur in other low SNR MRI applications, such as quantification of fluorinated drugs. Introduction: In Quantitative MRI, parameters are estimated from weighted images that are often misaligned by patient motion. This induces errors in the quantitative maps. Routinely, an image registration step is performed prior to mapping, however, residual misalignment might remain. We propose a Recurrent Inference Machine (RIM) based method to estimate T 2 maps while ignoring small misalignments in weighted images. The method is evaluated with simulated and in-vivo data. Methods: Theory: A RIM is a neural network that learns an iterative optimization scheme to find a regularized solution to an inverse problem [1] (Fig. 1 ). Our RIM uses the gradient of the likelihood function L = (j,r|S), given the signal model S =|A exp(-s n / T 2 )|? g(r), where s n is the n th echo time, and j = (A, T 2 ) T , with T 2 the transverse relaxation time and A related to the proton density. g(r) is a Gaussian noise with standard deviation r. Datasets and evaluation: Simulated datasets are denoted D t,r d , with t the SNR and r d the standard deviation of deformations: First, motion-free weighted images are created with the procedure described in [2] , without acquisition noise. We then simulate motion through 2D translations, rotations and bicubic interpolation. Translation parameters are drawn from N(0,r d ) for each image in the set. The rotation parameter is defined as the number of pixels displaced at the border of the image, and is also sampled from N(0,r d ). These parameters are normalized to ensure zero-mean motion. Finally, noise is added, with same r as [2] . To train the RIM to ignore misalignments (mRIM), a motion-corrupted dataset was used, with r d = 0.25. To train the baseline method (aRIM), we set r d = 0. Nine datasets D t,r d were created: D 5,r d , D 25,r d , and D 50,r d , with r d = {0, 0.25, 0.5}. In each, 100 samples were generated with same ground-truth. The RMSE was used as measure of error and its distribution across the (masked) brain is presented. Additionally, the T 2 dataset from [2] , which was pre-registered but still slightly misaligned, was used to assess performance with in-vivo data. Results: Figure 2 shows the evaluation with simulated data for different SNR and displacement levels. The proposed method mRIM has lower RMSE than the baseline aRIM for all cases except for dataset D 25,0 . T 2 maps from the in-vivo dataset are shown in Fig. 3 . The mRIM produced a smoother map than aRIM, with better distinction between tissues and shorter T 2 in vessels. We presented a novel method for MR relaxometry that is able to compensate small misalignment between images. Our method reduces the estimation error compared to a mapping-only baseline. Our results suggest that, with mRIM, small discrepancies between signal model and data can be compensated by the learned prior, improving the quality of estimates. Acknowledgements: Research Grant: B-Q MINDED (EU H2020 MSCA ETN 764513). C1.O2. Accelerated multi-shot diffusion weighted imaging with joint estimation of diffusion and phase parameters Introduction: Even though multi-shot EPI has great potential for diffusion MRI (dMRI), its long acquisition time compared to singleshot EPI and its sensitivity to shot-to-shot phase inconsistencies limit its application [1] . We propose ADEPT, a framework for Accelerated Diffusion EPI with multi-contrast shoTs. ADEPT provides 1) accelerated dMRI by varying diffusion contrast settings for each shot in a multi-shot acquisition, introducing intra-scan modulation, 2) estimation of the diffusion tensor (DT) parameters directly from the k-q-space data, surpassing the intermediate image reconstruction step of conventional dMRI. To account for the phase mismatches between different shots, ADEPT estimates the phase map parameters for each shot jointly with the DT parameter maps. In a simulation study, the performance of ADEPT in terms of the RMSE is evaluated for different acceleration rates and compared to conventional two-step diffusion estimation. Methods: Intra-scan modulated k-space data is considered, where each shot of a multi-coil (n c coils), multi-shot (n s shots) EPI acquisition is acquired with a different diffusion weighting and the phase map of each shot is assumed to vary linearly (due to rigid coherent bulk motion during the application of the diffusion sensitized gradients) [2] . The diffusion tensors, the complex-valued non-DW image, and the parameters of the linear phase maps of all shots are estimated jointly from the measured k-q-space data using a least-squares estimator (see Fig. 1 ), adopting a cyclic Block Coordinate Descent approach [3] . Experiments: Monte Carlo simulations with 20 noise realizations (SNR = 20) of the multi-contrast multi-shot data were performed. A fully sampled multi-channel dataset (n c = 8) was simulated including 1 b0 and 60 DW (b value = 1.15 ms/lm 2 ) images, using ground truth (GT) DT maps estimated from a real dMRI dataset and GT phase parameters drawn from zero-mean Gaussian distributions. This dataset was retrospectively sampled to generate 4-, 8-and 12-shot datasets. The estimation performance of ADEPT was compared to that of a conventional approach in which the diffusion tensor parameters for each voxel are estimated from SENSE reconstructed images of the individual shots (SENSE-re). Results: Figure 2 shows the estimated Mean Diffusivity (MD) and Fractional Anisotropy (FA) maps for different acceleration factors. Figure 3 shows the RMSE maps of MD and FA estimated with SENSE-re and ADEPT. Discussion: Figures 2 and 3 reveal that ADEPT better preserves structural details at high acceleration rates and has a lower RMSE compared to SENSE-re. Further validation of ADEPT using in-vivo rat brain experiments is currently ongoing and we will present the corresponding results by the time of the conference. Introduction: T2* mapping using ultra-short echo time (UTE) MRI enables quantitative evaluation of knee structures with short transverse relaxation times. 1 However, acquisitions with low throughplane resolution are commonly used to obtain T2* maps within reasonable scan time, affecting the estimation accuracy due to partial volume effect. 2 In this work, a model-based resolution-enhancing reconstruction (RER) method is used to obtain isotropic, high-resolution (HR) T2* maps of the knee from UTE Spiral VIBE MRI with low through-plane resolution within a reasonable scan time. Methods: 3 volunteers were scanned on a 3 T MR scanner (Prisma Fit, Siemens Healthcare) with a 15-channel knee coil (QED) using 3 T2* mapping protocols based on the accelerated prototypical 3D UTE Spiral VIBE sequence (Table 1 ). 3 TEs were chosen close to the inphase TEs of water and fat. A longer TR was used for protocols B (Reference) and RER to include more and longer TEs. The RER protocol consisted of five acquisitions with low through-plane resolution rotated around the frequency-encoding direction (Fig. 1 ). For protocols A and B, rigid registration was performed in Elastix 4 and T2* values were estimated voxel-wise using constrained non-linear least-squares fitting of a mono-exponential T2* relaxation model. Model-based RER T2* mapping with joint inter-scan motion estimation was performed as a constrained and regularized non-linear least squares estimation ( Fig. 1) . 5, 6 Regions-of-interest (ROIs) were drawn in eight knee structures and verified by an MSK radiologist. Finally, medians and ranges of median ROI T2* values were computed. Results: Figure 2A shows representative T2* and PD maps. Protocol A displays the lowest T2* values. Protocols B and RER provide similar T2* maps. Fig. 2B displays the medians and ranges of median ROI T2* values from all subjects computed for all protocols. Median T2* values of protocols B and RER are comparable, while protocol A seems to underestimate T2* values in long T2* structures. Discussion: RER UTE T2* mapping provides T2* estimates that are comparable to T2* results obtained with direct 3D UTE T2* mapping, while requiring approximately 25% less acquisition time. RER UTE T2* mapping thus shows great promise for HR T2* mapping of knee structures within reasonable scan time. C1.O4. Outcome prediction of mild traumatic brain injury using support vector machine based on longitudinal MR diffusion imaging from CENTER-TBI Introduction: Mild traumatic brain injury (mTBI) accounts for over 85% of head injuries, affecting millions worldwide each year, with a third showing persistent functional impairment for months1. We investigated the applicability of Support Vector Machines (SVMs) for prediction of Glasgow Outcome Scale score [GOSE]-based outcome via multicenter longitudinal diffusion MRI. Methods: Fifty-nine CENTER-TBI2 mTBI patients (15-70 years old, Glasgow Coma Scale (GCS): 13-15) with acute (MR1: 0-4 days post-injury) and subacute (MR2: 9-36 days post-injury) MR scan were analyzed. Patients were divided into groups with good (n = 35, 6-months GOSE = 8) and poor (n = 24, GOSE \ 8) outcome. T1w and DTI scans were acquired with similar protocols at 12 centers using 3 T scanners3. Mean diffusivity (MD) and fractional anisotropy (FA) maps were coregistered to T1w and projected to a template space. FA, MD and meta-data (age, sex, GCS and injury mechanism) were simultaneously used to predict GOSE-based outcome. SVMs were trained on MR1, MR2, longitudinal MR (MR2-MR1) and/or meta-data via fourfold cross-validation. The 10% most discriminative WM voxels were identified in each fold. Voxels identified in at least 2 folds were mapped to the JHU atlas4. Relative voxel count was calculated as the percentage of discriminative voxels in the tracts relative to all voxels in the WM structure. Results: MR1's FA and MD yielded highest mean accuracy (74.4%, Table 1 ), outperforming meta-data, longitudinal and subacute MR. Discussion: The SVM analysis predicted clinical outcome based on early diffusion MRI, achieving 74.4 ± 18.8% accuracy, supporting the selection of mTBI patients for clinical follow-up. Our results suggest that combining meta-data with acute MR provides better GOSE-based outcome prediction than meta-data combined with either subacute or longitudinal MR, or meta-data only. The cross-sectional findings are coherent with reported analysis 1. The longitudinal SVM exhibits lower prediction metrics, suggesting that subtraction might not be the most efficient way to exploit longitudinal information. Regions identified by the SVMs were previously reported as structures relevant for outcome after mTBI5. This multicenter diffusion study indicates the potential of GOSE-based outcome prediction using acute MRI. Grants: B-Q MINDED (EU H2020 MSCA ETN 764513) and CEN-TER-TBI (EC 602150). 1-Richter et al., JAMA. 2021; 4(3) , e210994. 2-Maas et al., Neurosurgery. 2015; 76(1) :67-80. 3-CENTER-TBI https://www.center-tbi.eu/project/mri-study-proto cols 4-Mori et al., Neuroimage. 2008; 40(2) , 570-582. 5-Wallace et al., Neurosci & Biobehav Reviews. 2018; 92, 93-103. C1.O5. Optimal experimental design for the T2-weighted diffusion kurtosis imaging free water elimination model 3 University of Antwerp, lNEURO Research Centre of Excellence, Antwerp, BE Introduction: The two-compartment free water elimination (FWE) diffusion model [1] allows to account for partial volume effects in brain tissue voxels affected by free water contamination. The illconditionedness of the FWE model can be mitigated by explicitly incorporating in the model the T2 relaxation properties of the tissue and free water compartments [2] , T 2 tissue and T 2 fw respectively. By representing the tissue compartment signal with the diffusion kurtosis imaging (DKI) model, we obtain the T2-DKI-FWE model. Due to acquisition time constraints in the clinical practice, the choice of the acquisition scheme remains crucial. Therefore, in this work, we show the potential benefits of using a data acquisition protocol with optimized b values (b) and echo times (TE). Methods: The acquisition settings, namely b and TE, were optimized in terms of the Cramér-Rao lower bound (CRLB) variances of the model parameters [3] . During the optimization procedure, the number of uniformly distributed gradient directions was fixed to 120 while b and TE were free to assume any values in the range 0-2 ms/lm 2 and 63-121 ms, respectively. Ground truth values for the model parameters were obtained by fitting the T2-DKI-FWE model to a real dataset. Subsequently, 500 white matter voxels were selected with an estimated free water signal fraction (f) between 0.05 and 0.3. As an optimality criterion, we minimized a weighted sum of the CRLB variances of the model parameters assuming Gaussian distributed data and, in our experiments, T 2 fw was fixed to 1573 ms. Results: In Fig. 1 , a conventional acquisition scheme with fixed b-TE combinations is compared to the scheme resulting from the optimization procedure as well as to a more practical clustered version of the optimized scheme. Figure 2 shows that, compared to the conventional scheme, the clustered optimized scheme leads to a higher attainable precision for the diagonal elements of the diffusion tensor and for all kurtosis tensor elements. However, the conventional scheme results in a lower CRLB of T 2 tissue . Finally, we assessed the sample variance of fractional anisotropy (FA), mean diffusivity (MD) and mean kurtosis (MK) in single-voxel experiments for different signal-to-noise ratio (SNR) values and Gaussian distributed data: for all 3 metrics, extracted using an unconstrained non-linear least squares estimator, the clustered optimized scheme outperformed the conventional one (Fig. 3 ). Discussion: This study points out that, given a fixed acquisition time, a higher precision of the DKI parameter estimates and derived metrics can be achieved by using CRLB-optimized b-TE combinations. Lower SNR ranges can be addressed by introducing constraints in the fitting procedure. Introduction: GRASE allows acceleration of MR acquisitions by acquiring gradient echoes next to a spin echo (SE) 1 . However, the standard GRASE acquisition and reconstruction causes some imaging artifacts. In earlier work 2 , we showed that by explicitly modeling the gradient echoes and SEs in a 3D-GRASE scan in a subspace constrained reconstruction (SCR) method, we could reconstruct images with less artifacts from highly undersampled GRASE acquisition. Figure 1 shows the difference between the conventional and proposed GRASE scans. However, with realistic B0 variations the required number of subspace components is large, limiting the acceleration factor. Here, we propose compensation of the majority of the B0 variation based on a calibration region, allowing a reduction of the number of components in the SCR and hence higher acceleration factor. We compare the performance of the proposed compensation technique to B0-inhomogeneity correction in conventional GRASE scans. Methods: A fully sampled scan of ISMRM model 130 3 phantom was performed using a 3 T GE Discovery MR750 scanner with scan settings as in 4 , except for repetition time set to 1800 ms due to SAR limits. The total scan time was 5 h:23 m. In this scan, every k-space position was sampled for every echo allowing retrospective generation of both the conventional GRASE-CONV and proposed GRASE-SCR. For GRASE-CONV B0-inhomogeneity correction was applied to each frequency encoding line using the reference echo train 1 . GRASE-SCR used a pseudo-random sequence generated by the Halton sequence with a 12 9 12 fully sampled central region (FSCR). We use the FSCR to generate low-resolution images by direct reconstruction. Coil combined images S-CAL j, x where x is spatial index and j indexes echoes, were generated using coil maps drived using the ESPIRiT technique as implemented in the BART toolbox 5 . Using the phase U j,x = \S-CAL j,x , which accounts for the majority of the phase changes due to DB0, we reconstruct S-REM j,x with the SCR 2 where S-REM j,x = S j,x U j,x . As S-REM j,x has lower range of DB0 than that of S j,x , reconstruction requires less number of subspace components. Results/discussion: Figure 2a -c shows comparison between DB0 estimated from calibration region and the actual DB0. Figure 2d shows the energy in each components after singular value decomposition for the range of DB0 expected with and without compensation. Figure 3a shows the comparison between GRASE-CONV and GRASE-SCR (first SE) where the blurring effect seen in GRASE-CONV is not present in GRASE-SCR, showing it can produce better images than GRASE-CONV. Figure 3b compares results from fully sampled fast SE scan and accelerated GRASE-SCR where the difference is mostly contained in the first echo while the rest of the images are similar showing potential for acceleration. Introduction: In clinical MRI, direct acquisition at isotropic high resolution (HR) and with high signal-to-noise ratio (SNR) is often infeasible due to the long scan time required. Multi-slice super-resolution reconstruction (MS-SRR) can reduce this limitation by reconstructing an HR 3D isotropic image from a series of MS images with an anisotropy factor (AF, defined as the ratio between image slice thickness and in-plane resolution) exceeding 1 1 . The multi-slice images are conventionally acquired either with parallel orientations, shifted by different sub-pixel distances in the through-plane direction 2 or rotated around the phase encoding axis by different angles 3 . In this simulation study, both MS-SRR acquisition protocols were compared using the Bayesian Mean Squared Error 4 (BMSE) of the maximum a posteriori (MAP) estimator as performance criterion. Methods: A Gaussian Markov random field 5 prior distribution was adopted for the HR image to be reconstructed and its hyperparameters were estimated from a training dataset of realistically simulated T2-w brain images 6 ( Fig. 1) . The bias and variance components of the BMSE were calculated separately to allow evaluation of both accuracy and precision. The acquisition schemes were compared for different combinations of AF and number of acquired MS images, while keeping the scan time and the spatial resolution of the reconstructed image fixed for all protocols. An isotropic HR MS acquisition protocol (AF = 1) requiring the same scan time was additionally used as a reference. The results of the BMSE analysis were validated numerically by means of Monte Carlo simulations. Bias, standard deviation and root MSE (RMSE) maps were estimated for each protocol from 100 noise realizations. Compared to the HR reference, MS-SRR based on rotated multi-slice images shows a significant improvement in precision, especially for higher AF, at the cost of a slight increase in bias. The difference in terms of BMSE suggests that adopting the rotated acquisition scheme in an MS-SRR experiment can lead to a substantially reduced scan time compared to the shifted acquisition scheme and the reference scheme, while preserving the same MSE of the estimated SRR image. Validation of the BMSE results on real data will be subject of future work. Research Grant: B-Q MINDED (EU H2020 MSCA ETN 764513). (MS) is an autoimmune disease that affects the brain and spinal cord. Symptoms of MS are various and can be of different severity. There is currently no cure for MS, but many treatments can slow down the symptoms' progress, resulting in less disability over time. Disease progression monitoring and clinical decision making often rely on the expanded disability status scale (EDSS). Unfortunately, EDSS is manually assessed and suffers from poor reliability, repeatability, and high inter-rater variability. Therefore, automatic and objective EDSS scoring using MRI information could potentially help to monitor disease progression reliably and optimize treatment. This work is the first step towards the prediction of future EDSS in MS patients from MRI. Methods: We used multi-center and multi-scanner FLAIR and T1 MRI data from 316 MS patients. EDSS score was established within 89 days before or after the MRI scan for each patient. The distribution of EDSS was strongly unbalanced with 240 patients belonging to the class of EDSS B 4. Icobrain [1] was used to preprocess the images and obtain the volumetric quantification of grey matter, white matter, whole brain, lateral ventricles, T1 hypointense and FLAIR hyperintense lesions. EDSS was automatically scored with three different approaches: 1) a random forest regression (RF) model based on the brain volumes; 2) a deep learning (DL) model based on pure MRI images with the two modalities T1 and FLAIR as channels; and 3) a DL model based on the combination of pure MRI and brain volumes. The DL model used consists of 4 consecutive 3D convolutional blocks, followed by 2 fully connected layers. The performance was assessed using the mean squared error (MSE), mean absolute error (MAE) and Pearson q correlation computed on the concatenated results of the test folds in a fivefold cross-validation experiment. In addition, the accuracy and the percentage of correct binary classification were calculated considering 2 classes split according to an EDSS threshold of 4. The results of the proposed automatic EDSS estimation methods are summarised in Table 1 . Overall, these results indicate a slight improvement in the EDSS decoding performance of the DL ? biomarkers model compared to the other two methods. The percentages of patients with correct classification per method using the EDSS threshold 4 are shown in Fig. 1 . Discussion: These findings suggest that machine learning methods are, to a certain extent, able to learn relations between MRI-based brain volumes and clinical disability measured by EDSS. MRI-based EDSS score might also capture complementary information on disease activity compared to the clinically measured EDSS. Understanding such differences is a prerequisite to predicting future disability progression in MS. Reference: Introduction: Tissue water content is tightly regulated in the healthy brain, and even small changes are indicators of pathology [1] . Techniques to precisely measure water content in the human brain with MRI have been previously developed. Short, clinically relevant, 2D acquisitions for water content mapping have been proposed [2] but using non-isotropic voxels and slice gaps for whole-brain coverage. We aim to use 2D imaging while also achieving high isotropic resolution by combining data from several short scans using superresolution reconstruction (SRR) [3] . Methods: For Super-resolution reconstruction-water content (SRR-H 2 O) mapping (SRR-H 2 O), three low resolution (LR), multiple-echo spoiled gradient echo (mGRE) images were acquired in orthogonal orientations with slice thickness = 2 mm, in-plane resolution 0.75 mm 9 0.75 mm, TR = 3270 ms, FA = 15 deg. These LR images were then reconstructed to one high resolution (HR) mGRE image with isotropic resolution (0.75 mm) 3 , using a conjugate-gradient based SRR method [3] . The original method was modified and adapted here for water content mapping (Fig. 1 ). The developed technique was validated using a carrageenan phantom [4] (Fig. 2 ). -The (0.75 mm) 3 isotropic resolution H 2 O map obtained from the SRR-reconstructed data was compared to the 0.75 mm 9 0.75 mm 9 1.5 mm H 2 O map obtained using the ''long-TR'' method. The acquisition times of both the methods were approximately the same. In vivo results were obtained from a healthy volunteer (male, 26 years). All measurements were carried out using 3 T PRISMA scanner, Siemens, Erlangen, Germany. Results: Figure 2 shows 2 and SPoiled GRadient echo (SPGR) signals and simultaneously corrects for the inhomogeneities of the excitation field B 1 , whose knowledge is of cardinal importance for quantitative imaging at high magnetic field 3 . Here we introduce Echo Planar Imaging Fast Actual Nutation Imaging (EPIFANI), an EPI implementation of VAFI for simultaneous mapping of T 1 , B 1, and proton density (PD) with 2D ultrafast acquisitions. Methods: EPIFANI was developed by including a 2D EPI readout to acquire data from both an AFI and SPGR sequence ( Fig. 1 ). Voxelwise mapping of T 1 , B 1, and PD was performed by minimizing the residual sum of squares of the signal intensities and the model signal for AFI and SPGR 1 . Simulations to assess the effects on parameter estimation of noise, T 2 * induced signal decay, and misalignment among even and odd k-space lines were performed. Data were acquired on an MR Solutions 4.7 T preclinical scanner in a single shot with flip angle FA = 60°, n = TR 2 /TR 1 = 1000 ms/200 ms, matrix size = 128 9 128, FOV = 40 9 40 mm 2 , slice thickness = 1 mm-SPGR acquisitions with FA = [15,25,35]°, TR = 200 ms-NEX = 3, and an optimized RF phase increment with high spoiling gradient areas. Steady-state was reached by the use of a preparation pulse 4 . A homogeneous agar gelatin phantom and a mouse head were scanned with a quadrature coil with 38 mm ID. A 14 points Inversion Recovery (TR = 10 s, TI [ [0.05,4] s) was performed on a gelatine phantom to compute reference T 1 values. • Simulations: maps of T 1 and B 1 reliably report the original input values, while PD maps show signal intensity variations provided by both T 2 * decay and lines misalignment in k-space (see Fig. 2 ). For T 2 * \ 41 ms median inaccuracies in T 1 greater than 10% arise. This suggests the use of multi-shot acquisitions for very short T 2 * s. • Acquisitions: with a total acquisition time of 51 s for a single slice image EPIFANI provided mean T 1 values (standard error, SE) of 2.056 (0.02) s with respect to 2.054 (0.001) s (IR) for the gelatine phantom. For in vivo data, an example of AFI image and T 1 map is reported in Fig. 3 . Median values for EPIFANI for cortical grey matter T 1 values are in line with literature 5 . Discussion: EPIFANI allows the acquisition of ultrafast and accurate T 1 and B 1 maps. Future developments will focus on the investigation of multi-shot EPIFANI, the implementation of distortion correction methods, and the estimation of their impact on parametric maps. University of Leeds, School of Medicine, Leeds, GB Introduction: MRI biophysical models are formulated based on certain assumptions, derived from biological information of the underlying tissue, to interpret MRI signals. Hence, there is a need for the numerical phantoms to validate these assumptions. Moreover, numerical phantoms can be used to analyse the sensitivity of MRI measurements to changes in tissue microstructure and the use of different pulse sequences [1] . The aim of this study is to develop a numerical phantom for myocardial microstructure, which incorporates the native probability distribution functions (PDFs) of cardiomyocyte (CM) shape parameters, regional curvature, ventricular twist and collagen into the phantom. Methods The proposed method receives the PDFs of CMs volume (V), length (L), major-axis (A), and minor-axis (B) PDFs, reported in literature [2] , as the inputs (Fig. 1a ). Using the PDFs of the A and B of the cross-section of the CMs, 2D ellipses are then packed [3] (Fig. 1b ) and subsequently transformed into polygons to mimic sheetlet crosssections (Fig. 1c) . Next, the 3D structure of a small section of a sheetlet is formed based on the length and volume PDFs of CMs, as illustrated in Fig. 1d . In subsequent steps, several such sheetlet sections are combined vertically to generate a complete sheetlet ( Fig. 1 .e). Finally, several sheetlets are concatenated horizontally and transformed by applying rotation, twisting, and bending transformations, shown in Fig. 2a -d, to generate a realistic myocardial wall. Finally, a cube is extracted from this wall to mimic a voxel of MRI (Fig. 2e) . A two-step statistical test evaluates how inputs PDFs are preserved following the phantom generation. First, using a virtual morphometry, output PDFs of 12,000 virtual CMs are measured. The Shapiro-Wilk test [4] is used to test the normality of the input and output PDFs of V, L, A, and B. Finally, the significance of the difference between the input and output PDFs is tested using a t-test [5] for parameters that follow a normal distribution, and a U-test [5] for those that do not follow a normal distribution. Results: The resulting p values of the test, shown in Table 1 , indicate no significant difference between the distribution of V (p = 0.76), L (p = 0.99), and A (p = 0.09) and B (p = 0.19) of the virtual CMs and the reference values reported in literature [2] . Discussion: The result demonstrates that the proposed method generates a numerical phantom for myocardial microstructure with realistic CMs. Future work will focus on validating this phantom respect similarity between packing of CMs in the virtual tissue slab and actual tissue using histology images. Undersampled spiral acquisition schemes have recently gained renewed interest 1 , as their realization is more compatible with current scanner hardware, and they lead to incoherent aliasing artifacts wellsuited to reconstructions with compressed sensing (CS). The purpose of this study is to investigate the feasibility of an efficient acquisition/ reconstruction strategy for accelerated high-resolution cardiac DTI based on the combination of a 3D variable-density stack-of-spirals trajectory and a joint CS reconstruction that incorporates group sparsity priors in both gradient and wavelet domains. Methods: Ground-truth DTI data were acquired from an ex-vivo rat heart on a 7 T preclinical Bruker scanner (BioSpec 70/20USR) using a 3D PGSE sequence, with diffusion gradients applied in 27 directions (b = 1000 s/mm 2 , RES = 100 lm 3 ). A 3D variable-density spiral sampling with golden-angle ordering of the interleaves was designed with the same specifications ( Fig. 1 ) and used to simulate spiral data from complex k-space measurements of the ground-truth data. A joint CS reconstruction method, which exploits the structural similarity among diffusion-weighted images (DWIs) acquired at different directions in both 3D total variation and 3D wavelet-transform domains, was adapted to jointly recover the non-Cartesian data 2 . Retrospectively undersampled data were obtained through randomly skipping the interleaves for each DWI to further ensure the incoherence and reconstructed using the proposed approach. For comparison, a basic disjoint CS method was also implemented. DTI metrics (MD and FA) were derived and compared with the reference. Results/discussion: Reconstructed images and the corresponding error maps of one diffusion direction with the proposed method at varying acceleration factors are depicted in Fig. 2 . Plots comparing the mean RMSE, SSIM, and signal to error ratio (SER) within the ROI (3D myocardium) over all DWIs obtained from conventional CS and the proposed method (PJGS) are also displayed. Figure 3 represents maps of FA and MD obtained from the reference and 10 9accelerated data with respective RMSE values. Conventional CS results in over-smoothed images at high undersampling rates which is reflected in the larger bias for resulting FA and MD maps while the PJGS provides images with superior quality and leads to lower errors. Introduction: To quantify T2*, multiple echoes are usually acquired with a multi-echo gradient-echo sequence using either monopolar or bipolar readout gradients. The bipolar readout gradients have a shorter echo spacing time compared to the monopolar readout gradients, enabling the collection of a larger number of echoes within the same scan time, with potentially more accurate T2* quantification. 1, 2 Nonetheless, comparison of the readout gradients for T2* quantification has not been addressed to date. In this work, the performance of the readout gradients has been investigated at 3 T. The bipolar readout gradients provide more precise T2* maps and show robustness under a relatively lower SNR condition. Methods: Experiments were performed on a 3 T MRI scanner (PRISMA, Siemens Halthineers). The monopolar and bipolar readout gradients were optimised to have the shortest echo spacing time under the given imaging parameters: FOV = 240 9 240 mm 2 , matrix = 240 9 240, slice thickness = 1 mm, # of slices = 11, TR = 1200 ms, TE1 = 3.84 ms, DTE = 2.92 ms for monopolar (32 echoes) and 1.47 ms for bipolar (64 echoes) readout gradients. The flip angle was changed from 75°to 15°in 20°decrements for phantom experiments and 76°for in vivo experiments. The parallel imaging acceleration factor (AF) was also varied from 1 to 3. Magnitude fitting was performed with a mono-exponential function, considering Rician noise distribution. 3 Results: Figures 1 and 2 show the calculated T2* maps from the monopolar and bipolar readout gradients. In Fig. 1 , each column and row displays the result for a different set of flip angle and AF combinations. Visual inspection suggests that bipolar readout gradients give less variation than monopolar readout gradients, particularly with a lower flip angle and higher AF (see the area indicated by yellow arrows). In Fig. 2 , as with the phantom results, T2* maps from the bipolar readout gradients look less noisy than those from the monopolar readout gradients. The mean and standard deviation of the calculated T2*s are presented in Fig. 3 . The results obtained with the bipolar readout gradients show a smaller standard deviation than those from the monopolar readout gradients. Discussion: An increase in the mean and standard deviation was observed for a larger AF and smaller flip angle for both the monopolar and bipolar readout gradients. The mean T2* was comparable between the two readout gradients. However, in the case of standard deviation, the bipolar readout gradients were less sensitive to the changes in the SNR. It can be seen that the bipolar readout gradients show greater robustness against noise, which was mainly due to the use of a larger number of echoes for T2* quantification. Research Grant: B-Q MINDED (EU H2020 MSCA ETN 764513). [1] . The PI-RADS guidelines state that DWI obtained using b values C 1400 s/mm 2 should be used [2] , so consistency in b values amongst centers cannot be expected in clinical practice. This may hamper performance of algorithms trained using specific b value DWIs. Moreover, previous studies concluded that computed instead of acquired DWI could be used effectively for PCa detection [3] . We investigated prostate lesion segmentation accuracy by training a segmentation network using acquired DWI (aDWI) or computed DWI (cDWI) at different b values. Methods: The dataset comprised of 414 single-center 1.5 T mpMRI scans (Siemens, AvantoFit). For each study, lesions were manually segmented and PI-RADS scored by one radiologist, out of a pool of radiologists. An additional distinction was made for PI-RADS 3 lesions, since biopsy may be appropriate on factors besides mpMRI [2] . Lesions with a PI-RADS C 3 were grouped between high (PI-RADS 4-5, biopsied), mild (PIRADS 3, biopsy requested by radiologist) and low (PIRADS 3, not biopsied). Our lesion segmentation method is based on 3D UNet architecture [4] with three input images: T2w, ADC and DWI. Models were trained and validated in a fivefold cross-validation setup with either aDWI or cDWI with b values of 800 or 1500 s/mm 2 to investigate the model robustness against varying b values. cDWI were calculated from ADC and b800 scans using a mono-exponential fit [1, 3] . Classification performance was evaluated with sensitivity (SE) and false positive rate (FPR) at a threshold of 0.5 and FROC curves. Segmentation performance was determined by the Dice Similarity Coefficient (DSC), computed for the true positive (TP) lesions. A prediction was considered TP with at least one voxel overlaps with the reference lesion and false positive (FP) otherwise. The DWI train/infer combinations and results are shown in Table 1 . Figure 1 illustrates FROC curves for each combination. Figure 2 shows an example of combination outputs. Discussion and conclusion: Best results are obtained with ab1500/ ab1500 (acquired b1500) train/infer combination (SE = 0.9/0.74/0.62 for high/mild/low), followed by ab1500/ab800. Similarity between both FROCs demonstrates robustness against the specific b value used during inference, regardless of the b value used for training. This is not the case for cb1500 (computed b1500), for which performance is lower. We therefore conclude that lesion segmentation networks are less sensitive to different measured b values, than to computed higher b values. Research Grants: B-Q MINDED (EU H2020 764513). Aim: Hybrid imaging systems play an essential role in the non-invasive characterization of oncological diseases. Therefore, stringent quality control (QC) procedures are required to ensure the proper functionality of the systems and to gain accurate quantitative results. Here, we evaluate existing QC procedures for combined PET/MRI systems in clinical routine. Materials and methods: Eight highly-experienced European imaging sites ( Fig. 1 ) were surveyed about locally implemented PET/MRI QC procedures. The survey was based on existing recommendations on QC for stand-alone PET and MRI systems. In addition, relevant vendor-specific information on QC measures was collected. Results: In total, five Siemens, two GE, and one Philips PET/MRI systems were reported. For all the systems, the centres reported performing the PET daily QC as implemented by the vendor. However, we found moderate to high variations across other tests and testing frequencies for the PET and MRI components of the systems. Aiming at addressing this significant variability across sites and vendors, we propose a consensus on QC procedures. This consensus includes for the PET component: a daily QC as suggested by the local vendor, a quarterly cross-calibration measurement including the assessment of uniformity, and a yearly image quality (IQ) test for the PET component of the systems. For the MRI component, monthly coil checks and a quarterly MR IQ test, including the assessment of signal-to-noise ratios and artefacts that could affect clinical scans and image quality. Figure 2 summarizes the reached consensus. We observed significant variations in daily QC procedures across PET/MR imaging sites. Therefore, in this work, we propose a consensus on minimum QC procedures for clinical routine. University of Nantes, CRCINA, Nantes, FR Introduction: Hybrid PET/MR imaging will reveal its full clinical potential if we revisit the acquisition paradigm towards fully quantitative multi-parametric imaging and do not replicate what is currently performed on separate PET/CT and MR systems. Admittedly, fully quantitative MR imaging requires more time than in usual clinical practice, but it also offers the opportunity to perform dynamic PET imaging in parallel for parametric PET imaging. PET has the ability to deliver fully quantitative functional information of underlying imaged processes, by use of dynamic imaging and kinetic modelling. Dynamic PET can deliver unique biomarker information complementary to quantitative MRI, which can assist in clinical applications and delivery of precision medicine. Oncological applications of PET/MRI often require imaging over the whole body for tumour staging but hybrid PET/MR scanners provide only a limited axial field of view. To increase the effective field-ofview, dynamic whole-body (DWB) protocols are used over multiple bed positions, at the cost of limitations in acquisition counts and sampling frequency. Advanced and dedicated processing tools are necessary to deal with these limitations. The objective of this work was to improve whole body multi-parametric imaging for PET/MRI applications by use of direct multi-bed dynamic reconstruction in PET. Methods: A dynamic reconstruction platform was developed within the open source reconstruction software CASToR 1 , that can support direct reconstruction of multi-bed PET data and MRI-guided regularization. The spectral analysis dynamic model 2 was used in the reconstruction process, because it does not impose any strong assumptions on the underlying kinetics and provides temporal regularisation of reconstructed data, enabling more flexibility for parametric imaging. The direct dynamic reconstruction was applied and evaluated on simulated PET FDG data and on a first in man whole-body pharmacological PET/MR study, using [ 11 C]Glyburide to evaluate the role of OATP transporters in drugs pharmacokinetics 3 . Results: K i maps for FDG and K 1 maps for Glyburide were calculated. Comparison between regular frame by frame static reconstructions followed by post reconstruction parametric modelling and dynamic reconstruction of DWB data showed good agreement with no introduction of bias on evaluated metrics. Furthermore, the use of dynamic reconstruction resulted in noticeable noise reduction of activity and parametric images. Discussion: We effectively applied DWB protocols for whole-body PET/MR parametric imaging. Findings showed that dynamic PET reconstruction is desirable in DWB parametric imaging to achieve accurate and precise quantification, while providing whole-body parametric images of comparable image noise to regular single bed dynamic protocols. Introduction: MR imaging for clinical follow-up and pre-operative work-up of patients with glioma generally involves a contrast-enhanced T1-weighted (T1ce) sequence. Although the use of contrast agents (e.g., gadolinium) greatly enhances the visibility of tumor and facilitate the delineations of viable tumor tissue, it exposes the patient to higher risk and increases the scanning time [1] . Therefore, we proposed a cross-modal distillation approach for improved segmentation of the viable tumor using only a non-enhanced T1-weighted (T1w) MRI sequence. Methods: Cross-modal knowledge and feature distillation (KD and FD) refer to the process of training a student model to achieve better performance by learning from the teacher model which has access to a privileged source of data [2] . For KD, the softened output of the teacher model was used in addition to the ground truth labels to train the student model such that the predictions by both models were matched. However, the softened outputs do not directly reflect the intermediate feature representations learned by the teacher. Therefore, the concept of KD was extended to a more generic framework called FD where the additional supervision was used to synchronize the intermediate activation maps between the teacher and student models. A deep learning segmentation model based on 3D U-Net was used to implement both teacher and student models. Given that the T1ce sequence is highly informative for segmenting the viable tumor, this sequence was used to train the teacher model, while the T1w sequence was the only data available for the student model during inference. For the evaluations, the BraTS 2018 dataset consisting of 285 multi-sequence MRI and an in-house dataset consisting of 37 subjects with LGG and HGG gliomas were used. Due to the limited size of the in-house dataset, transfer learning was applied. Results: KD/FD approaches improved the segmentation Dice scores up to 8% compared to the baseline model trained by T1w only and provided comparable segmentation performance with T1ce-based segmentations. For the clinical evaluations, segmentation results were consistent with the BraTS 2018 evaluation, especially when transfer learning was used to fine-tune the pre-trained models. Discussion: Segmentation performance of KD/FD using only T1wsequence closely approximated the performance of the model trained by T1ce-sequence. Therefore, this approach provides a valid, clinical alternative for accurate tumor core segmentation when the T1ce-sequence cannot be acquired because of scan time restrictions or safety reasons. C2.O4. Multi-parametric PET/MRI for radiation treatment planning of catients with cervical cancer Introduction: The concept of personalized medicine has brought increased awareness to the importance of inter-and intra-tumor heterogeneity for cancer treatment. Therefore, biological tumor characterization based on functional and molecular imaging might be highly valuable for radiotherapy (RT) planning to allow for an improved target volume definition and an individualized dose prescription within the tumor. The combination of PET and MRI can serve as a one-stop-shop RT planning session and provide measures of e.g., angiogenesis, vascularization and perfusion characteristics, and cellular density. The aim of this study was to prove the potential and feasibility of multi-parametric hybrid PET/MRI to non-invasively describe tumor heterogeneity prior to chemoradiotherapy. Methods: Eight patients with histologically proven primary cervical cancer were scanned 30 min after injection of 200 MBq 68 Ga-NODAGA-E[c(RGDyK)]2. PET data were acquired simultaneously with T1 and T2 weighted TSE sequences, DW-MRI, and DCE-MRI. Standardized uptake value (SUV), apparent diffusion coefficient (ADC), and pharmacokinetic maps obtained with the extended Tofts model (Ktrans, Ve, and Vep) were generated for each dataset. The primary tumor and a reference structure in the gluteal muscle were manually segmented on T2-weighted images by an experienced radiologist. The anatomical tumor delineation was used for correlation analysis between available pairwise combinations of functional and parametric maps on a voxel and a regional level. Results: A representative image dataset of functional images and parameter maps of the primary tumor is presented in Fig. 1 . MRIderived parameters and SUVmean were significantly different between tumor and reference tissue. Correlation analysis on the regional level showed a moderate voxelwise correlation between RGD-PET and Ktrans (R2 = 0.3), while the correlation between RGD-PET and ADC map was weak. However, the respective correlation coefficients varied strongly within the patient cohort. Discussion: The multi-parametric PET/MRI datasets suggests that a combined voxel-wise analysis of PET and MRI-derived parameter maps can contribute to a better characterization of tumor heterogeneity than the modalities alone. Our findings showed a tendency toward higher values of tumor perfusion in areas with more intense RGD uptake [1] . However, according to the literatures, tumor perfusion could also be simulated by FDG uptake and hypoxia, which may occur when the tumor blood flow is inadequate [2, 3] . It needs to be stated that the current results are based on small patient numbers, but they might contribute to the future design of individually adapted treatment approaches based on multiparametric functional imaging. According to the World Health Organization, cardiovascular diseases are the main cause of death in the world, showing the need for substantial contributions by the new medical technologies. In the last decades the value of the information obtained by non-invasive cardiovascular imaging (Echo, MR, SPECT, PET/CT, PET/MR) is considerably increasing in clinical tasks such as diagnosis, prognosis, and therapy monitoring. However, to make significant contributions, it is crucial to mitigate the technical issues associated with cardiovascular imaging, and then to pursue a complete characterization of the disease. Physiological and physical factors such as heart motion during the image acquisition and partial volume effects impact on the quality of the images. Hybrid PET/MR systems allow us to take advantage of the strength of each modality, and due to the simultaneity of the acquisition, to accurately integrate the information to overcome the above issues. Using FDG PET/MR images of a myocardial viability assessment protocol, different integrative studies were performed in cardiac and respiratory PET gating, PET partial volume correction based on MR derived anatomical references, and PET/MR based contractility recovery prediction after revascularization of chronic coronary total occlusion patients, seeking for parameters validation and potential clinically relevant synergistic effects. Positive correlations between PET and MR values of morphologic and functional parameters of the left ventricle were found, but with a tendency of underestimation with PET that suggested a non-interchangeability between these modalities in the morphology assessment. Myocardium focused MR image-navigators and PET data-driven methods for heart tracking along the respiratory cycle showed similar results in respiratory motion correction. Hybrid PET/ MR multiparametric assessment presented an improvement beyond the standard analysis based on FDG uptake and LGE for the prediction of contractility recovery after revascularization. However, the prediction based on PET images corrected for partial volume effects was reduced. Synergistic effects of cardiac PET/MR imaging were observed in myocardial viability assessment protocols, even though further studies are needed to evaluate the trade-off between improvement in the information and the added complexity to the complete workflow. Introduction: Selective internal radiation therapy (SIRT) requires liver registration for CT and MR images to integrate multi-modality information for precise dose calculation. This study aims at investigating the feasibility of using liver segmentations from convolutional neural network (CNN) and manual landmark segmentations to guide liver registration for CT and MR images. Methods: An in-house image-based registration method was extended by incorporating segmentations of livers (CNN-guided) or both livers and landmarks (CNN&LM-guided) for guidance. For the segmentation-guided methods, the deformable registration is initialized by an affine one based on segmentations instead of images. The deformable registration uses a combination of mutual information as the image similarity loss function (L I ) and the sum of squared differences as the segmentation similarity loss (L S ). To avoid topologyviolating deformation, the voxels are assumed to be connected via springs, which oppose distance changes. The spring rigidity determines the power of the regularization loss (L R ). Therefore, the final loss function is w I L I ? w S L S ? L R . Liver segmentations were generated by a CNN model in [1] . Twenty SIRT patients were selected for training and testing of the algorithms (see Table 1 ). Each patient has one Reference CT. Manual liver and landmark (mainly lesions) delineations were used for registration evaluation. The root mean squared distance (RMSD) of mean surface distance between livers and mass center distance between landmarks is used to evaluate each liver registration. The optimal wI and wS for all registrations were determined by finding the minimal root mean square of RMSDs (RRMSD) through grid searches for the training datasets. Results: As shown in Fig. 1 , the CNN-guided registration decreases the RRMSD by 3.4 mm (34%) compared to the image-based one. The optimal performance of the CNN&LM-guided registration was achieved when using only affine registration (w I & w S = 0). The RRMSD is decreased by 1.4 mm (21%) after landmark guidance for the CNN-guided registration. The CNN-guided affine registration outperforms the image-based registration, since CNN segmentations help align the liver contours and exclude the disturbance from other organs. The CNN-guided deformable registration degrades the affine registration. This might be caused by segmentation errors in some low-intensity lesion regions of the MR. These regions are deformed due to the surface matching of CNN segmentations. Landmark guidance shows its value of improving the CNN-guided registration. Therefore, we will develop an automatic lesion segmentation to obtain a fully automated segmentation-guided registration. This study is funded by EU Horizon 2020 research and innovation programme Marie Skłodowska-Curie grant 764458. [1] and PET/CT [2] . With rising adoption of PSMA-ligand imaging, standardized reporting frameworks and image-derived biomarkers are increasingly employed for reproducible and accurate image assessment. Nevertheless, clinical applicability is challenging in cases where manual measurements of numerous suspected lesions are required. In this context, automated image analysis methods are promising to enable replicable, timeefficient, and rigorous evaluation. We developed a deep learning method to support staging in PSMA-ligand imaging. Methods: 173 patients with confirmed prostate cancer referred to 68 Ga-PSMA-11 PET/CT were retrospectively analyzed. For each subject, an expert physician segmented regions of elevated tracer uptake, labeled them as physiologic or suspicious for prostate cancer, and assigned them an anatomical location classification. A multi-task convolutional neural network was trained to both classify elevated uptake sites as physiologic or suspicious and assign them an anatomical location, based on the combined PET/CT information. Additional training information was included from 629 patients with lymphoma or lung cancer, imaged with 18 F-FDG PET/CT and analogously annotated in a former investigation [3] , encoding the radiotracer type as network input. A hold-out test set of 52 prostate cancer patients was used exclusively to evaluate the network performance. Moreover, in the test subjects we evaluated the agreement between N and M stage assigned based on the network annotations and expert annotations according to the PROMISE miTNM standardized framework [4] . .7) for identification of suspicious uptake sites; 77% (CI: 70.0-83.4) accuracy for anatomical location classification of suspicious findings, 81% agreement for identification of regional lymph node involvement, 77% agreement for identification of metastatic stage. Discussion: The evaluated method was able to automatically identify sites of tracer uptake suspicious for prostate cancer and classify their anatomical location in good agreement with the expert assessment. Image quality characteristics may influence the network output and results will require expert supervision for use in clinical context. The investigated method is promising for enabling efficient assessment of cancer stage and tumor burden based on multimodal PSMA-ligand imaging. [2] . Upscaled production can be difficult, especially at GMP-grade, being the conventional drying method of freeze-drying a rate limiting step [3] . Methods: We reviewed the literature to assess PFC formulation and production techniques, focusing on more efficient upscaled production. We tested some of these techniques with different formulations. Results: High-pressure homogenization, ultrasonication and microfluidization techniques are the most effective methods for PFC encapsulation. While ultrasonication is often an effective and versatile option for small-scale emulsification processes, microfluidization and high-pressure homogenization are capable of producing higher volumes, showing a good reproducibility and size distributions. Stabilization techniques for the aqueous storage of PFC formulations include electrosterical stabilization or the addition of hydrophobic compounds in the majority of the cases. Drying and rehydration processes also involve structural damage risk, so parameters like glass transition temperature of the encapsulating agents or vapour pressure of the PFC need to be considered to choose the right method. Discussion: The literature and our data show that upscaling and stable formulations are possible if the certain combination of formula, operating parameters and production regime are found. Further research in drying of PFCs is needed to reduce the costs involved with production. 1) . Results: The most commonly used tracers for 19F-MRI are perfluorocarbons. Nanotechnologies are vital for using perfluorocarbons. Indeed, due to their immiscibility with organic and inorganic solvents, perfluorocarbons need to be emulsified or incorporated into polymer or lipid nanoparticles by vigorously mixing a water solvent and PFCs in the presence of a surfactant, and lastly by applying high pressure homogenization or sonication (Fig. 2) . Fluorinated biomaterial-based nanoprobes, such as fluoropolymers, copolymers and lipids are a further pioneering and promising class, characterized by tailorable properties which enable them to be ideal for many biomedical applications,such as theranostic purposes or functional imaging in response to specific changes in physiological stimuli. Moreover, they hold a high number of fluorine atoms causing a high 19 F-MRI signal, and a low cytotoxicity. In order to address the sensitivity issue of 19 F-MRI techniques, other types of nanoprobes have been designed, aimed at improving relaxation properties of PFC molecules. In this regard, the combination of a paramagnetic center (e.g. Mn 2? , Fe 3? , Gd 3? ), with a 19F nanoprobe, reduces 19 F T1 values for a rapid image acquisition and increases the signal-to-noise ratio per unit time 1 . The literature showed that PLGA-based nanoprobes are a very promising 19 F probes developed to date. PLGA nanoparticles have been employed for both small fluorinated molecule encapsulation 2 and fluorine attachment 3 in order to synthesize the 19 F copolymer. Introduction: Atherosclerosis is a vascular inflammatory disease characterized by the development of fatty plaques in the intimal layer of arteries, hence becoming the underlying cause of strokes and heart attacks. Thereby, it represents one of the most prevalent causes of patient mortality in the western world. Early diagnosis of inflammation and plaque development in atherosclerosis is yet a challenge, but it can be accomplished using nanotechnology, by developing specific targeting contrast agents to the damaged area. Among common techniques applied in the detection of cardiovascular diseases, 19F MRI presents the unique combination of quantitative non-invasive molecular imaging of biological processes and safety due to lack of ionizing radiation, thus being an optimal candidate for atherosclerosis diagnosis. Methods: As perfluorocarbons-based contrast agents, we will use several compounds with a single resonance frequency and with high local concentration of fluorine atoms for good signal to noise ratio. Envisaged designs for MRI probes will exhibit the fluorine label attached to polymeric NPs'' surface through a water soluble linker, also, we will use fluorine probes (PERFECTA, PFCE) encapsulated inside micelles or liposomes emulsions. Once generated, physico-chemical characterization of fluorinated formulations can be assessed by 1H and 19F-NMR, TEM, DLS, whilst cell viability and nanotoxicity will be evaluated by hemolysis, MTT and ROS assays. Contrast agents targeting can be achieved using different atherosclerosis biomarkers attached to the fluorinatd probes, in order to obtain specific imaging of thrombi. Afterwards, preclinical experimentation will be conducted by using 19F MRI on high magnetic field (7 T) to evaluate probes biodistribution and in vivo fate. Animal models will be developed by ApoE knockout mice fed with high-fat diet to effectively generate atherosclerotic plaques in the main arteries. Discussion: 19F-MRI represents a promising platform for early and sensitive diagnosis of atherosclerosis. However, the correct stabilization of fluorine-based formulations in aqueous environments is key for their real application in clinics. Future work in this field will involve strategies to overcome this challenge and develop groundbreaking contrast agents. 3 , and perfluoro-octylbromide (PFOB) 4 . Recently we showed that it is possible to acquire rapid images of perfluorocarbons on a very low-field system 5 , taking advantage of the very small spectral dispersion of even multi-peak compounds, the short T 1 relaxation time, and the ability to increase the effective T 2 in scalar coupled systems using very short interpulse CPMG sequences. Here we extend this work to design sequences which can be used to differentiate between molecules with different scalar coupling behaviour, as well as ones with no scalar coupling. Methods: The system operating at 50mT has been described in detail previously 5 , as has the solenoid coil which can image both 1 H and 19 F via operation of a mechanical switch. PFOB, fomblin and PCE were studied ( Fig. 1) . Each sample was placed in a 1.5 mL cylindrical tube. The T 1 values were measured via a spectroscopic inversion-recovery sequence, and the T 2 as a function of inter-pulse delay in a CPMG sequence. The first sequence is a spin-echo with minimum echo time, meaning that essentially a fluorine-density image is obtained. The second sequence is also a spin-echo with a long echo time, which produces very low SNR from PFOB and fomblin samples. The third sequence uses a CPMG preparation module with very short interpulse delay, which results in high signal from PFOB and PCE, but low signal from fomblin due to their different behaviour regarding scalar coupling rephasing. By combining the signal intensities from the three sequences, images can be produced primarily from each of the three species. Results: The measured T 1 values were 530 ms (PCE), 720 ms (PFOB), and 85 ms (fomblin). Figure 2 shows the apparent T 2 values as a function of inter-pulse delay in a CPMG sequence. The strong dependence for PFOB is shown, along with a weaker one for fomblin, and no dependence for PCE. Figure 3a -c shows images from three sequences. Figure 3b essentially produces an edited image, with signal from only the PCE, due to the short apparent T 2 value of the other two compounds when a long interpulse delay is used. Figure 3c shows the effect of the CPMG preparation module, resulting in almost complete suppression of the fomblin signal as desired. Figure 3d , e show the results of subtraction of the relevant images to produce images corresponding to fomblin and PFOB only. Conclusion: A simple editing scheme has been designed to differentiate between various perfluorocarbons based upon their scalar coupling properties. The scheme is currently based on spin-echo sequences, but will be extended to more time-efficient turbo-spin echo sequences. Introduction: The lymphatic system is pivotal in whole-body fluid homeostasis and immunity. Lymphatic failure causes fluid accumulation within tissue (lymphoedema), with primary lymphoedema (PL) originating from an inherited defect. Magnetic Resonance lymphangiography (MRL) offers improved spatio-temporal resolution and three-dimensional (3D) depiction of lymphatic vessels (LVs), compared to the 2D images of the main lymphatic trunks produced with routine lymphoscintigraphy. We investigated whether co-registration and baseline-subtraction of dynamic contrast-enhanced MRL (DCE-MRL) improved LV depiction and diagnostic potential. Methods: Four PL and four healthy volunteers (HV) legs were imaged with 3D T2 weighted (turbo spin echo, TR/TEeff = 2800/ 565 ms, voxel = 2 9 2 9 3 mm 3 ), and pre and post-contrast 3D T1 weighted (spoiled gradient echo, TR/TE/FA = 3.7/1.6 ms/12o, voxel = 1 mm 3 ) sequences on a 3.0 T Philips dual Tx Achieva system and 16-element torso coil. Contrast was delivered intradermally to the inter-digital spaces of the feet, after which C 30 min DCE imaging occurred. Field of view and imaging time varied, however T1 volumes (* ankle to knee) were typically acquired in 3 -4 min. DCE-MRL datasets were registered to the first post-contrast phase (PC1), using an affine or deformable approach (NiftyReg, Modat et al., 2010) , or left unregistered. Default parameters for affine registration, and bending energy (a penalty term promoting smooth deformation) = 10% and grid spacing = 8 voxels for elastic registration, were used as previously reported (Borri, 2018) . Subtracting PC1 from all subsequent dynamics after registration was also performed. in HVs previously not depicting LVs (Fig. 3) . A single reviewer rated the three pipelines and ranked deformable registration as producing subtraction images of highest quality in 3 PL and 2 HV cases. However, in cases with substantial limb motion (rotation and translation) registration reduced LV conspicuity (1 PL and 2 HV). Baseline-subtracted uptake curves were also assessed as being smoother post-registration (Fig. 2 ). Discussion: Baseline subtraction of DCE-MRL images can increase LV conspicuity in both HV and PL participants and may be improved with prior registration, however substantial whole-limb movement may not be sufficiently compensated. . The MR images were acquired with a 1.5 T Ingenia (Philips) through a GRE sequence with slice thickness: 3 mm, matrix: 120 9 120, ET/RT: 3.42/5.28 ms and a flip angle: 12°. A standard protocol was performed following the medical order (3 patients with a cerebrovascular accident -CVA-, 1 with vestibular schwannoma -VS-and 3 with arteriovenous malformation -AVM-), including the acquisition of 4DFlow. Two specialist neuro-radiologists segmented the lesions and traced at least two ROIs in each segment to study the vascular flow (Fig. 1) . The blood dynamics differences between healthy (HT) and pathological tissue were analyzed taking into account the hemodynamic response function (HRF). Using a Python algorithm (Qt, VTK and SimpleITK), the visualization was integrated into the 3DSlicer platform (v. 4.11) through vectorial DICOM-RGB encoded images. Results: A change in the amplitude and behavior of the HRF were evidenced in 2 subjects with CVA, 2 patients with AVM and the subject with VS (Fig. 2) . The RGB encoding allowed us to verify the correct directionality of the vascular flow in all HT ROIs, generating turbulent areas with homogeneous speed in the studied pathologies (Fig. 3) . The 4DFlow is a technique that generates additional information to the anatomical vascular sequences, quantifying information that can be used to determine new standardizations in the evaluation of vascular pathology. Implementing numerical methods like second order Runge Kutta, the pressure exerted by the vascular walls can be quantified through the analysis of temporal changes of the blood velocity, which can be used to assess diseases associated with the elastic properties of the vascular bundle. University of Lisbon, Lisbon, PT Introduction: Multiparametric MRI (mp-MRI) holds promise to characterize lesions, and predict patient outcomes. However, localize/ stage the lesions in mp-MRI is a time consuming process. Automatic segmentation would reduce analysis times and would potentiate radiomics analysis. Here we explore the U-Net architecture for lesion segmentation with mp-MRI data and evaluate the impact of mp-MRI data. Methods: Data Description: Data from 28 patients with biopsy-confirmed tumor were included. Data Acquisition: Data were acquired in a 3 T MR equipment (Siemens Healthineers). MR sequence were: (1) Axial 2D-T2wTSE, (2) DWI b = 50, 1000, 2000), ADC images were calculated within the equipment workflow. Manual Segmentation: Prostate and index lesions were manually segmented by a senior radiologist. Segmentations were performed on T2w and ADC images. Algorithms: Both images (T2w and ADC) were normalized using the Z-score prior to segmentation. We used the U-Net architecture (https://github.com/zhixuhao/unet) where the residual blocks and the multiple-input capabilities were implemented. Models: Two approaches were evaluated: 1-step) segmenting the index-lesion directly from T2w data and 2-steps) use a cropped prostate and seg ment the index-lesion using ADC and/or T2w data to evaluate the impact of each input. Training/testing: For prostate segmentation, a training/test split of 80/20 was used. For lesion segmentation, fourfold cross-validation was considered. The optimization metric was the DSC for the validation set. Hausdorff Distance (HD) and Mean Square Distance (MSD) were also evaluated. Results: Automatic prostate segmentation resulted in a DSC of 0.88, HD of 16.5 mm and MSD of 2.1 mm in the validation set. Figure 1 presents a sample slice for prostate segmentation. Regarding the lesion segmentation, using the 1-step, 2-step and 2-step mp-MRI approaches, the average DSC in the validation set were: 0.51, 0.71 and 0.76, respectively. Performing the 2-step resulted in an improvement of 39% when compared to the 1-step approach. By introducing the mp-MRI information, we obtained an improvement of 10%. Table 1 summarizes the results and Fig. 2 presents a segmentation result with the proposed mp-MRI model. The model for prostate region segmentation achieved a DSC of 0.88 which is agrees with the state-of-the-art References. The 2-step obtainedbetter results in terms of DSC when compared with 1-step, which highlights the importance restrict the segmentation region. The T2w index-lesion segmentation improves with the introduction of ADC information in the network, suggesting the potential of mp-MRI. Our work is limited by the factthat we consider the segmentation of a single expert and limited database. Conclusions: A model to segment prostate index-lesions in mp-MRI was proposed using U-Net. The initial results suggest that mp-MRI information improves index-lesion segmentation and that 2-step approaches improve lesion segmentation results. [1] . However, at present, a non-invasive MRI technique for assessing liver overload by T2* values has become widespread [2] . Despite the advantages of this non-invasive technique, its use implies a correlation between MRI data and biopsy data. Since this correlation can depend on many scan parameters and other factors, it should be performed at every medical center wishing to use this technique reliably. We propose a standardization method that avoids the long and complicated process of constructing a conversion formula between T2* values and iron concentration (LIC). For this, we use the MR-compatible phantom and the previously obtained formula for recalculating the T2 * values in the LIC [3] . Materials and methods: The phantom was created on the basis of complex iron oxide nanoparticles, which were obtained in our laboratory in Dmitri Rogachev Pediatric Hematology Center as a result of the Elmore reaction [4] : The phantom consisted of 28 tubes and covered the entire range of T2* values found in children with liver iron overload ( Fig. 1) . To check the repeatability of the results in our center, the phantom was scanned 6 times on Philips Achieva/3 T and Signa GE/1.5 T in Rogachev Center. To test the methodology for standardizing the production of T2* cards, the phantom was scanned at two corresponding scanners in other institutions. Results and discussion: Repeatability tests showed satisfactory results-the standard deviation was acceptable for determining the overload grade (Fig. 2) . As a result, a method was formulated for standardizing the assessment of liver iron overload using the phantom: 1) Scanning a phantom at another institution. 2) If the results are in the confidence intervals of the reference values obtained on MRIs at the Rogachev Center, then it is possible to use the formula we obtained earlier in a new institution (Fig. 3) . 3) If this match is not found, it is possible to calibrate the study parameters until the desired match is reached. 4) If this does not help, it is possible to enter the calibration factors into the conversion formula. Thus, for a non-invasive assessing of liver iron overload in these medical centers, it is possible to reliably use the previously obtained formula. It is significantly speeds up and simplifies the process of introducing this technique into clinical practice. (perfusion). This effect is modelled by a biexponential fit and characterized by a diffusion coefficient (D), a pseudodiffusion coefficient (D*) and the f-IVIM, that represents the flowing blood volume fraction. Similarly, another deviation from normal gaussian distribution is obtained when high b values are used, and a Kurtosis model is also used to increase sensitivity to identify tissue features of intracellular water behavior (perfusion). Another feature associated with prostate lesions analysis is its stiffness. The elastography (MRE) have shown significant differences between prostate carcinoma (PCa) and healthy tissue.The aim of this study is to correlate the results obtained with IVIM/KURTOSIS versus MRE. Methodology: The study was approved by the local Institutional Review Board; magnetic resonance (MR) images were acquired in a 3 T scanner MR/PET of General Electric (GE) for 16 patients (12 patients with PCa and 4 with prostatitis); these included diffusionweighted images (DWI), that were obtained with a FOCUS sequence with 15 b values (from 0 to 3000 s/mm 2 , Fig. 1 ), and MRE-SE (frequency = 60 Hz, amplitude = 70). Two specialist radiologists traced ROIs in healthy (HT) and pathologic tissue. IVIM/KURTOSIS were processed using a Python algorithm (QT, VTK and SimpleITK), while MRE were processed with AW VolumeShare 5 (GE). The triexponential fitting was implemented with the Scipy Python library. Two specialist neuro-radiologists segmented the lesions and traced at least two ROIs in each segment to study the vascular flow ( Fig. 1) . Results: Patients with PCA showed a decrease in the D (m = -34.81) and f (m = -71.49) coefficients, while the k coefficient and the stiffness increased (m = 23.93) compared with the HT (Fig. 2) . A correlation between the coefficient values in the ROIs of HT and PCA was observed (Fig. 3) . Conclusión: While the f and k coefficients followed the perfusion behavior reported in other studies, as well as the MRE, the D* coefficient did not show a proportional behavior. The curve fitting for prostatitis patients did not generate logic results due to technical failures in the adjustment. Puropose of the study: In the recent time of pandemia of viral pneumonia (VP) caused by COVID-19, we aimed to study the possibility of using chest MRI to image lung damage in this pathology both in primary detection and for follow-up control of recovery. Material and methods: MRI of the chest in T1-, T2-weighted modes (T1-w, T2-w), also with fat suppression, diffusion-weighted, STIRmodes, in the axial and frontal planes, with breath holding, or with automatic synchronization of acquisition with breathing was carriedout in 47 patients with VP, 32 of them were confirmed by PCR as COVID-19, all did have a clinic of pneumonia. The control group comprised 15 volunteers (8 non-smokers, and 7 smokers). In 18 patients, an CT study of the chest was also performed. In 8 patients, MRI of the chest was then performed again, for follow-up control of clinical recovery. There were no deaths among our patients. Results: The duration of MRI of the chest was \ 25 min in all cases (21 ± 4 min on average), and less than 10 min in the chest CT. In all cases, MRI imaging of the affected area was achieved already using a group of MRI screeninf protocols, comprised axial T1-w and axial and frontal T2-w, did last \ 12 min. In normal patients without pathology of the lungs, not smoking, the lung was visualized as a diffuse homogeneous air region with a quota of interstitial and vascular space \ 4%. In patients-smokers, lung MRI was enhanced in the dorsal parts of both lungs, disorders of airiness, interstitial exudative and fibtosis changes weren't present. In the acute phase of the disease, pulmonary ventilation disorders and interstitial exudative changes that form the morphological basis of lung damage in COVID-19 were visualized as local foci of, corresponding to the location and nature (sub-segmental, segmental, polysegmental) of the pathological focus, both T1-w and T2-w modes. MRI of the chest provided diagnosis of lung pathology in all cases, while the extent of the pathological focus on the MRI image in T2-w was 14-19% over the CT values. The correlation of the calculated volume of affected lung tissue between CT and MRI of the chest was as high as r = 0.95 (p \ 0.001). The values of the volume of the affected tissue in T1-w and T2-w did not differ from each other in the intergroup comparison and correlated as r = 0.985 (p \ 0.001). MRI in DWI mode showed a sensitivity of 81% (38/47) in detecting COVID-lung lesions. The volume of pleural effusion, clearly visible on T2-w images, in all our cases did not exceed 100 ml. In a prospective follow-up of 8 patients with COVID-19, chest MRI ptovided imaging of recovery process in all cases. Conclusion: MRI of the chest with respiratory synchronization or with breath-holding can be used for early diagnosis of inflammatory lung lesions in COVID-viral pneumonia and for subsequent follow-up control of therapy. [1] [2] [3] [4] [5] . The compression stiffening rate measured with MR elastography (MRE) has recently been used to non-invasively assess tumor solid stress in small animals [6] . The aim of our study was to assess if compression MRE using different breath hold modes as stress source [7] could be used to assess cancer invasiveness (i.e. microvascular invasion) in patients with hepatocellular carcinomas (HCC). Methods: This prospective clinical study, approved by the ethics review committee of our institution, included 41 patients with HCC (median size: 40 mm, interquartile range: [30 mm; 60 mm]; 24 HCC with microvascular invasion) who were imaged before tumor surgical resection. The MRI examinations were performed on a 3.0 T Philips Ingenia scanner. MRE was performed with the ''''eXpresso'''' sequence at a frequency of 50 Hz (n = 34) and 40 Hz (n = 7) [8] , with TR/TE: 89/6.9 ms, spatial resolution: 4.2 mm 3 , FOV: 420 9 420 mm 2 , slices: 8 and 4 temporal steps. Patients were instructed to hold their breath at end inspiration and expiration. Maps of elasticity (G'') and viscosity (G'''') were calculated by inverting the Helmholtz wave equation. In addition, T2-weighted MR images were acquired with TR/TE: 531/120 ms, spatial resolution: 0.59 9 0.59 9 4 mm 3 and FOV: 280 9 376 mm 2 . Tumor deformation (e) maps were calculated by applying a 3D affine registration on the T2 images (Matlab R2020a, Natick, MA). Stress (r) was calculated with the Hooke law, as r = 3G' inspiration .e. Finally, the compression stiffening rate was determined as the linear slope between G'' and r. Microvascular invasion was determined with histopathological examination. Results: HCC elasticity was significantly higher at end inspiration than at expiration (2.8 ± 1.1 kPa, 2.1 ± 0.9 kPa, respectively: t test p = 0.002), (Fig. 1 ). At logistic regression, the compression stiffening rate was the only MRE parameter of microvascular invasion: r partial = -0.42, p = 0.007 (Fig. 2) . The compression stiffening rate, in contrast to the visco-elasticity, showed significant diagnostic performance for detecting microvascular invasion as shown by the areas under the receiver operating characteristic curves (AUC): AUC compression stiffening = 0.75, p = 0.002 versus AUC G'' = 0.55, p = 0.6, and AUC G'''' = 0.57, p = 0.4 (Fig. 3) . Discussion: In patients with HCC, MRE compression stiffening rate had significantly higher accuracy than visco-elasticity to assess microvascular invasion. These results suggest that compression MRE during forced inspiration and expiration might be useful as a noninvasive clinical method to assess hepatocellular carcinoma invasiveness. Introduction: In patients with head and neck cancer, accurate detection of lymph node metastases is of great relevance regarding prognosis and treatment [1, 2] . USPIO-enhanced MRI is a promising imaging method to detect metastatic lymph nodes. After intravenous administration, USPIOs accumulate in cells of the immune system residing within healthy (parts of) lymph nodes [3] , attenuating MR signal intensity in T2*-weighted pulse sequences. Suspicious lymph nodes retain MR signal intensity. In the validation process of this technique it is of interest to determine whether phagocytic uptake occurs in the blood or inside the lymph nodes. The aim of the present work is therefore to investigate which blood compartments contain the USPIO particles after infusion. Methods: Blood samples were collected from five head and neck cancer patients at two different timepoints. The baseline sample was obtained prior to the slow-drip USPIO infusion, the second sample 45 min after the start of USPIO infusion. The blood was separated into six components: serum (SER), plasma (PLA), full blood (FB), peripheral blood mononuclear cells (PBMC), and granulocytes with and without red blood cells (GRA ? RB, GRA-RB). The presence of USPIOs influences the relaxivity of the blood compartments and therefore the transverse relaxation times (T2) were measured, performing the Carr Purcell Meiboom Gill (CPMG) sequence [4, 5] with 16 echo time points ranging from 1 to 100 ms, up to 2 to 5000 ms, depending on the a priori estimated T2. The relaxometry measurements were performed at a Bruker Avance 11.7 Tesla magnet, at room temperature (298 K). After fourier transform of the 16 free induction decays, the decay curves of the integrated NMR signals of water were fitted with a mono exponential, using the Bruker Topspin software. Shorter T2 relaxation times indicate the presence of USPIOs. Results: After infusion, significantly shorter T2 times were observed in full blood, serum and plasma indicating the presence of USPIOs, which seem absent in other compartments. The decay of the NMR signal of water is depicted in Fig. 1 . Averaged transverse relaxation times of the six blood compartments with and without USPIOs are presented in Table 1 . Discussion: Intravenously injected USPIOs are present in the serum and plasma compartments. These data indicate that USPIOs are not phagocytosed by the white blood cells present in the blood within 45 min after administration, but rather remain in the serum/plasma fraction. The full blood component represents the other compartments all together and therefore also shows significant difference in T2 times between blood with and without USPIOs. Introduction: Obesity is a severe disease, characterized by accumulation of fat in tissues, that induces important alterations in liver, pancreas or brain. It has grown to epidemic proportions, with over 4 million people dying every year as a result of being obese. Besides, obesity is associated with other diseases like diabetes or cancer (1). The influence of overweight in brain tumors, however, is still unclear. On these grounds, an orthotopic glioblastoma mouse model was developed in animals fed with standard diet (SD) and mice submitted to ten weeks of high-fat diet (HFD) feeding prior to intracranial glioma cells injection (2) . Methods: Two cohorts (n = 12) of C57BL6/J male adult mice with HFD (60% fat) and SD feeding were used. Animals were submitted to GL261 glioma cells intracranial injection 10 weeks after diet diversification. They were constantly supervised, and MRI studies were performed 21-24 days post-intracranial tumor cell injection. Anatomical images, magnetization transfer (MT), diffusion tensor Images (DTI), T2, and T2* parametric maps were acquired in a 7 T Bruker Biospec horizontal equipment. Parametric maps were generated using an in-house program developed with MatLab which adjusted the MRI signal to the corresponding mathematical equation. Four regions of interest were manually selected, using ImageJ: i) tumor core, ii) tumor periphery, iii) peritumoral region and iv) contralateral healthy brain. Statistical analysis was performed by twoway ANOVA tests with area and diet as factors and post-hoc multiple comparisons with Bonferroni corrections. Results: Animals submitted to HFD presented an augmented weight the day of intracranial injection, then, weight tended to decrease till endpoint in both mice cohorts. Obese mice exhibited significantly higher axial diffusivity (Dax) and fractional anisotropy (FA) (p \ 0.05 for diet effect). Slightly increased mean diffusivity (MD), T2* and magnetization transfer ratio (MTR); and decreased radial diffusivity (Drad) and T2. Discussion and conclusion: The results are consistent with the existence of inflammatory proceses, which are more remarkable in the case of the obese mice (3). These preliminary data point to remarkable differences in the assessed MRI parameters suggesting that obesity might influence in glioblastoma''s development. Introduction: Nonalcoholic fatty liver disease (NAFLD) is a chronic liver disease that can occur as nonalcoholic fatty liver (NAFL, without or with inflammation at histopathology) or steatohepatitis (NASH) which is characterized by steatosis, lobular inflammation and hepatocellular ballooning (1) . Measurements of damping ratio (2) , loss modulus (3) and frequency dispersion coefficient (4) have been proposed to assess NASH and liver inflammation. In several rheological models, damping ratio is tied to frequency dispersion, but this relationship has not been precisely assessed in vivo. Here, we evaluated the relationship between damping ratio and frequency dispersion in a mouse model and assessed the value of these mechanical parameters as potential biomarkers of NASH and liver inflammation. Methods: After Ethics Committee approval, mice fed with normal diet, high fat diet or high fat, choline-deficient diet were examined with MRE at 4, 10 and 16 weeks (n = 12/time point and group). Liver 3D MRE examinations were performed with a 7 T Bruker system (300 lm 3 , at 200 Hz, 400 Hz and 600 Hz). The damping ratio f was calculated at all frequencies, and the frequency dispersion of the shear stiffness (c G* ) was evaluated as the coefficient of a power law (5) . To investigate the relationship between f, c G* and histological features, rank correlations, logistic regression, Mann-Whitney tests and ROC analysis were used. Discussion: Our results show that the frequency dispersion coefficient and the damping ratio are determined by lobular inflammation rather than steatosis and ballooning. This may limit their value as biomarkers of NASH and indicate a larger potential for the diagnosis of inflammation. Our study suggests that the diagnostic potential of the damping ratio is not the same at all frequencies, and that the damping ratio at a single frequency may not fully recapitulate the frequency dependence of the viscoelastic parameters. References: Introduction: Compared to sonography or computed tomography, magnetic resonance imaging (MRI) offers high soft-tissue contrast and allows the measurement of physiological parameters such as temperature non-invasively, as well as multiplanar real-time images for better navigation. These aspects are attractive for controlling and monitoring minimally invasive diagnostics and therapeutics (e.g. biopsy, tumor ablation, brachytherapy). Despite the number of benefits, several disadvantages, such as the constricted space in the MR bore or the non-standardized workflows, make the procedures in clinical practice difficult and limit interventional MRI (iMRI) to relatively small case numbers [1, 2] . To counteract this, an intuitively operable, MR-compatible, ultra-light and remotely controllable micropositioning system called lRIGS has been developed for universal interventions [3] . The aim of this first study is to evaluate the performance of lRIGS in the anticipated clinical workflow for a simulated MRI guided prostate intervention. Methods: The instrument positioning unit (IPU) of lRIGS is developed mainly with 3D-printed parts in compact dimensions to achieve five degrees of freedom (DoF) by Bowden cables and a remotely controlled electrical drive unit placed inside the scanner room (SR). For validation, lRIGS is set up for an MRI guided pelvis phantom to puncture the prostate (transgluteal access) in a clinical radiology environment (see Fig. 1 ). Within the anticipated workflow (see Fig. 2 ) the system and therefore the main puncture is performed from the control room (CR). Results: After initial feed (see Fig. 3a ), the current trajectory (CT) did not correspond to the target trajectory (TT). This was followed by an iterative angle adjustment with the needle already advanced until the CT matched the TT and immediate verification by the radiologist using the fiducial marker (white) under real-time imaging. Fig. 3b shows the remote feed of the needle along the CT. Discussion: This study demonstrated the performance of the lRIGS system in a clinical workflow that provides fully remote-controlled needle puncture. Through the iterative angle adjustment, the user has to rethink from the system kinematics relative to the orientation in the slice images. In conclusion, this work has shown the possibility to adjust the instrument position in the MR bore remotely without being present inside the SR. These first results will be used for next lRIGS demonstrators, which can guide various instruments at ubiquitous application areas autonomously and more precisely. In future, the complex process of instrument navigation will be taken over by robotic systems to save time and costs as well as to reduce complications during interventions. Introduction: In cancer research, a challenging goal is to increase the amount of drug that can reach the tumor. Low intensity ultrasound (LIUS) could be an appealing technique, in virtue of their high biocompatibility and good tissue penetration. 1 The aim of this study is to investigate the ability of LIUS to enhance the tumor delivery of a liposomal formulation of doxorubicin (Doxo), using a clinically approved Gd-based MRI contrast agent (Gadoteridol), co-encapsulated in the liposomes (Gd-Lipo-Doxo), to report on the US-triggered drug release by MRI in a mouse model of ovarian cancer. Methods: Female athymic nude mice were purchased by Envigo (Milano, Italy) and were inoculated with A2780 cells (human ovarian cancer cell line). Mice tumor growth was monitored by 7 T-MRI. Mice were divided in three groups; i) not treated group ii) group treated with Gd-Lipo-Doxo, and iii) group treated with Gd-Lipo-Doxo and exposed to US stimulation. The amount of the doxorubicin and Gd in the tumour was assessed spectrofluorimetrically and by ICP-MS, respectively. Low intensity pulsed US waves were locally applied using 1 MHz piezoelectric transducer (PA, Dorset, UK) in order to induce the temporary permeabilization of the cell membrane and the drug release from the liposome system. Results/discussion: To optimize the maximal release of agents from liposomes and the permeabilization of the cell membrane (sonoporation) in vitro experiment were performed. Both Gadoteridol and doxorubicin showed a similar ability to be released from the liposomes and be internalized in tumor cells. In the in vivo experiments, the tumour LIUS stimulation resulted in a T 1 -contrast enhancement in the tumour and in the kidney calix and bladder, thus indicating the effective intratumor release of gadoteridol (and doxorubicin) from the liposomes. The tumour uptake of the Gd-agent and the drug was confirmed by confocal images (Fig. 1) and quantification. Notably, the group of mice treated with Gd-Lipo-Doxo and US-stimulated showed an arrest in the tumor growth. Interestingly, the tumour T 1 contrast enhancement measured just after the US treatment showed a positive correlation with the therapeutic outcome after three weeks. The ultrasound treatment appears a safe and efficient method to improve the chemotherapy efficacy of a liposome-encapsulated doxorubicin in a mouse model of ovarian cancer. All components of the theranostic agents (Lipo, Doxo, and gadoteridol) are already clinically approved (though not in the same system), making this approach suitable for clinical translation. Moreover, MRI seems to be an excellent imaging technique to monitor and predict the therapeutic outcome. Introduction: Magnetic resonance imaging (MRI) is becoming a valuable tool for the diagnosis 1 and treatment planning 2 of uveal melanoma patients. For radiotherapy planning it is important to know the reproducibility of target delineation. The primary aim of this study was to assess the interobserver variation in gross tumour volume (GTV) delineation of uveal melanoma (UM) on MRI. Our secondary aim was to determine the optimal MRI contrast for target volume delineation. Methods: Six observers delineated the tumours of ten patients. All MRI scans were acquired using the setup as described by Ferreira et al 2 and segmented using Big Brother software 3 . The GTV was segmented separately on the T2-weighted (T2w) and the post contrast T1-weighted (cT1w) images. The other sequences (T1w and cT1w or T2w) were only used as reference to differentiate the different structures. The volume and variance of the delineations on T2w and cT1w were compared. Results: It was observed that on cT1w flat tumour extensions were included more often compared to T2w. The average tumour volume was significantly larger when delineated on cT1w compared to the T2w (p = 0.01). The average contour variation was 0.4 mm on cT1w and 0.3 mm on T2w. Greater variability was found at the tumourretinal detachment (RD) interface and at the tumour edge (average 0.5-0.7 mm), compared to the tumour-vitreous and tumour-sclera interface (average 0.3-0.4 mm). Discussion: The difference in segmentation volume and observer variation on cT1w and T2w has been observed before 4 and can be partially explained by the interobserver window level differences found in cT1w. The average observer variation of 0.3 mm and 0.4 mm is low (* 0.5 acquisition voxel) although for small lesions like uveal melanoma it might still be of importance. Finally, even though the variation is higher on cT1w, we recommend to delineate the GTV on cT1w with T2w and T1w as reference, as small tumour extensions appear to be missed on T2w. We hypothesize that CAF diet given to rats before, during pregnancy, and lactation differently influence metabolic profiles and fat content. Females were exposed to CAF diet for four weeks before pregnancy, during pregnancy, and lactation. Control rats (C) were fed standard chow diet AIN93G. Body composition of females were analyzed. Methods: MRI experiments were carried out using a preclinical horizontal scanner operating at 9.4 T (400 MHz for protons-Agilent) equipped with a 600 mT/m gradient system. For MRI imaging a 72-mm i.d. quadrature birdcage type coil was used. During the MRI experiment animals were put at a special designed holder and anesthetized with 1.5-2% isoflurane in a 50/50 air-oxygen mixture. The temperature of the animal was kept at 37°C. Respiration of the animal was monitoring and used to synchronize MRI experiments Results: CAF diet in mothers increased food intake during pregnancy and altered body weight and fat content before, during and after pregnancy. There were significant differences in food intake in mothers in the second and third weeks of administration of CAF diet before pregnancy. We have found that during these weeks mothers from CAF group ate more compared to C group. Moreover, there was difference in food intake in the second week of pregnancy, when CAF mothers ate more compared to C. Mothers'' body weight in CAF diet group was also increased, but reached statistically significant value at 6 after CAF diet exposure. A similar trend was seen at week 7. Moreover, there were changes in percentage of fat content in animals. While at 1st week fat content was higher in C compared to CAF females, at week 13th CAF rats had higher fat content compared to controls. Discussion: We confirmed the hypothesis that mothers cafeteria diet differently influences metabolic profiles and fat content. Besides study of pups, we also examined changes in females before during pregnancy and lactation. We found that dams on CAF diet had increased food intake in the second and third week before pregnancy and in the second week of pregnancy when compared to C group. These findings are in agreement with data obtained by Sanchez-Blanco et al., which shown that female rats kept on CAF diet 22 days before pregnancy and during pregnancy had higher food intake than control. Sanchez-Blanco C, Amusquivar E, Bispo K, Herrera E. Influence of cafeteria diet and fish oil in pregnancy and lactation on pups' body weight and fatty acid profiles in rats. Eur J Nutr. 2016;55 (4) Introduction: Stroke is one of the leading causes of death and disability in the world. According the latest data (1, 2) it was suggested the diffusion characteristics of the brain change not only in the ischemic focus, but also in the intact white matter. The aim was to study diffusion characteristics changes of the intact (contralateral) hemisphere white matter at the acute and chronic period ischemic stroke vs control group using diffusion-weighted and kurtosisweighted imaging. Methods: The three groups of the patient at the age 40-70 yo was examined: control (n = 11), acute stroke (1-3 days after clinical manifestation, n = 12), chronic stroke ( [ 3-month, n = 17) . It was performed on the 3 T MR-scanner with sequences: 3D T1-TFE (axi), T2-TSE (axi), 3D FLAIR-SPIR (sag), DWI-EPI (axi). DTI sequence was additional to the routine protocol: b = 0, 1000 and 1500 s/mm 2 , TR/TE 10,500/73, matrix 96 9 94 9 25, voxel 2.33 9 2.33 9 2.33 mm, 25 slices, 13 min 55 s, 32 directions of the diffusion gradient, located mainly along the edges of the cube. Based on the this the coefficients of axial, radial, mean diffusion and fractional anisotropy of the intact hemisphere white matter were assessment with a mono-exponential (gaussian model, apparent diffusion coefficient map) and bi-exponential analysis (non-gaussian model, intravoxel incoherent motion model or diffusion kurtosis imaging) of data (3) . Results: The values of the coefficients axial, radial, mean diffusion and kurtosis for the white matter of the intact hemisphere were obtained for each of the groups (Fig. 1) . According the analysis of variance the statistically significant difference was founded for the mean diffusion coefficient for the conrol vs acute and control vs chronic stroke groups, as well as for the mean kurtosis coefficient for the control vs acute stroke groups ( Table 1 ).The statistically significant differences between the control vs stroke groups in the mean diffusion coefficient (p = 0.0009) and the mean kurtosis coefficient (0.007) in the white matter of the intact hemisphere was confirmed by the Mann-Whitney test. Discussion: It was suggested the use of the diffusion and kurtosis characteristics of the white matter as the indicator of the brain state and it's may be prognostic factors of the stroke risk. The mean kurtosis coefficient probably indicates the severity of the loss of cellular structures, for example, in degenerative disease or neuronal/axonal atrophy (4) . The results can be potentially useful both for early diagnosis and determination of the stroke risk, and for prognosis clinical outcome (5). S2.P2. Altered brain activation associated with working memory during spontaneous migraine attacks Introduction: Migraine is the second most common neurological cause for disability in young and middle-aged adults 1 . It is a cyclic disorder that consists of recurrent migraine attacks (ictal phase), and symptom-free periods in between (interictal phase) 2 . There is evidence showing cognitive disturbance in patients while undergoing the ictal phase, but controversy exists regarding other phases of the cycle 3 . We investigated working memory in a clinical sample of migraineurs using fMRI. Subjects/methods: 12 adult women with episodic migraine without aura underwent 2 sessions of fMRI (during (1) ictal and (2) interictal phases). fMRI was acquired on a 3 T Siemens Verio System, with a 12-channel RF coil using T2 weighted gradient-echo EPI sequence TR = 2000/30 ms and voxel size of 4 9 4 9 3.6 mm. N-back paradigm consisted of a block design with two conditions (0-back and 2-back) (NordicNeuroLab fMRI). The fMRI data was analyzed using FSL. Preprocessing included fieldmap distortion correction, spatialsmoothing (FWHM = 5 mm) and non linear registration to MNI space. For statistical analysis a general linear model was built including standard motion parameters as covariates. For second-level analyses we used flame1 with all sessions and per group of sessions; Session1 (S1) and Session2 (S2). For ROI analysis, a mask was extracted subtracting S1-S2 in order to obtain the percentage of BOLD signal change on the working memory contrast for all subjects on each session. SPSS was used for statistical analysis using non parametric tests and correlations. Results/discussion: Group mean maps of activation during the 2-back task, and significant differences between ictal and interictal sessions, are shown in Figs. 1 and 2 . Percentage of BOLD signal change in a left prefrontal ROI showing significantly higher activation during the ictal phase, and its correlation with task performance, are shown in Fig. 3 . Prefrontal cortex regions on the left hemisphere were significantly more activated while performing a verbal working memory task under pain condition. Learning effect of this task can be observed as a decline in activation as well as improvement in performance, as described in previous literature 4 . This is the first time that working memory is investigated using fMRI to observe test-retest effects during spontaneous migraine attacks. Introduction: The experience of violence is a stressor and is often associated with long-term effects on mental health. These effects might lead to changes in the structure and function of brain networks. The aim of the current study was to elucidate this stress-violencebrain relationship by testing structural and functional covariances among four stress-related networks: the default mode (DMN), salience (SN), dorsal attention (DAN), and frontoparietal (FPN) network. The four networks were parcellated into spatially distinct nodes, and brain data were averaged across all voxels of a node. We contrasted covariance in victims of violence (VoVs, n = 34) and matched controls without experience of violence (n = 32). Methods: Participants underwent an extended scanning protocol, with a pre (R1) and a post-task (R2) resting-state fMRI scans and an anatomical MPRAGE scan. The task consisted of stress/emotion induction paradigms. Structural connectivity was determined from covariance in gray matter mass (GMM 1 ) using VBM. Sex, age, psychiatric diagnosis, antidepressants, the number of other psychotropic drugs, and total intracranial volume (for VBM) were controlled. We used sparse inverse covariance with threefold GraphicalLassoCV 2 to estimate the partial correlation between two regions while conditioning for all other correlations in one step for each group and each modality. Results: From VBM, we identified 7 pairs of nodes that differed significantly between VoVs and controls, 6 of them showing less GMM covariance in VoVs. Within-network disturbances emerged only in the SN (3), while the FPN showed the highest proportion of inter-network differences between groups (Fig. 1) . In RS1 ( Fig. 2A , B) 5 node-to-node connections differed significantly between groups: SN-DMN (2 positive, 1 negative), DAN-DMN, and DAN-FPN (2 negative), and 1 increased intra-network connection within FPN. Proportionally, differences were largest in DMN (67%) and SN (50%). For RS2 (Fig. 2C, D) , 5 connections differed between groups. Less covariance in VoVs was found between SN Insula and DMN LP (R), while more covariance was observed for all other inter-and intranetwork connections. Differences were more frequent in SN and FPN (60%). Discussion: Results demonstrate structural and functional connectivity differences prior to and after the two mild experimental stressors in a transdiagnostic and heterogeneous sample with a history of violence compared to unaffected controls. In line with previous findings 3, 4 in normal healthy and traumatized populations, the important role of the SN, and SN-centered network interactions, may reflect dysfunctional mechanisms in VoVs that depict a risk factor for diverse mental health problems. Introduction: Temporal Lobe Epilepsy (TLE) is one of the most common epilepsies and is drug-resistant in 76.8% of cases [1, 2] , therefore needing surgery. TLE is often associated with sclerosis of the mesial temporal structures: the hippocampus, entorhinal cortex, and amygdala, which are involved in memory encoding, storage, and retrieval [1] . So, mapping of the memory network is critical for preventing surgical sequelae. Common memory paradigms require complex and costly instrumentation, preventing its more widespread use. Herein, a memory paradigm was adapted with hand grasping to collect patient feedback seamlessly. Methods: Twelve TLE patients (8 males) aged 25 to 62 were studied. As part of a pre-surgical planning protocol, all participants performed a hand grasping and a face-name memory paradigm adapted from [3] . A block design of 100 s duration was used for the hand grasping paradigm in which the subjects were asked to alternately move each hand for 10 s. The memory paradigm consisted of viewing faces unfamiliar to the patient paired with fictional first names, in a modified Novel vs. Repeated design, with a total duration of 10 min. Each face-name pair was presented for 5 s. A total of 4 runs were done, in which each run was composed of Novel face-name pairs followed by a fixation cross for rest, followed by Repeated face-name pairs. The Repeated facename pairs could have wrong names associated, to check whether the patient remembered the correct pair. The patients' answer was provided through hand grasping (left hand to wrong/ right hand to correct). Brain scans were done using a 1.5 T MRI scanner with a 12-channels head matrix coil. A volumetric T1-weighted scan was done, followed by BOLD sequences for the functional paradigms. A general linear model (GLM) was used in FSL to analyse the functional images to map hand grasping regions and regions related to memory encoding and retrieval. The higher-level group analysis included the following contrasts: EncodeRegion, regions active for encoding; and RetrieveFalse and RetrieveTrue, regions related to wrong and correct face-name correspondence, respectively. The regions observed to be related to encoding and retrieval of memory information were (see Figure) : -Encoding: bilateral temporal lobe, fusiform gyrus, and right inferior orbitofrontal cortex. -RetrieveFalse: sensorimotor cortex, procedural memory regions including basal ganglia and the cerebellum, and the culmen, fusiform gyrus, anterior cingulate, and limbic lobe. -RetrieveTrue: similar activation patterns as those of Retrieve-False, with additional activation in the occipital lobe and lingual gyrus. Discussion: This study has demonstrated that memory network mapping can be fairly easily implemented in a clinical setting using a face-name paradigm with hand grasping as a response. PET imaging with 15 O-water as the radiotracer has been the standard modality to measure CBF and CVR. However, it is impractical in most hospitals due to the requirement of an on-site cyclotron. Arterial spin labeling is a quantitative MRI technique that enables non-invasive CBF and CVR measurement and has demonstrated effectiveness in measuring CVR of normal subjects 2 . In this work, we compare the CVR of Moyamoya patients, whose vessel occlusion is a risk factor to acute stroke, measured by ASL MRI with the reference-standard 15Owater PET. Methods: Imaging data were acquired from 26 Moyamoya patients (18-64 years, 16 females) using a simultaneous 3 T PET/MRI system (GE SIGNA, Waukesha, WI, USA). All patients had unilateral or bilateral vessel occlusion at the anterior cerebral artery (ACA), middle cerebral artery (MCA), and/or posterior cerebral artery (PCA). The scanning parameters of the single post-labeling delay pseudocontinuous ASL (single-PLD PCASL), delay pseudo-continuous ASL (multi-PLD PCASL, 3 PLDs), and PET were described in our previous work 2 . PET and ASL data were acquired simultaneously at baseline and 15 min after the injection of the vasodilator acetazolamide (15 mg/kg with a maximum of 1 g). Each patient received 891 ± 71.8 MBq of 15O-water. CBF of PET was computed using the single-compartment pharmacokinetic model; CBF of ASL was computed using the general kinetic model 3, 4 . CVR was computed as the percentage of CBF change compared with baseline CBF. Flow territories (right and left ACA, MCA, and PCA) were defined based on the Harvard-Oxford cortical and subcortical structural atlases 5 . Mean CVR within the territories affected by occlusion and normal territories was computed for each subject. Paired t-tests were performed to compare the mean CVR between the affected and unaffected territories. Concordance correlation coefficient values were computed to examine the agreement between CVR measured by ASL and PET. Results: Figure 1 shows the hemodynamic maps of an example patient. Figure 2 shows that the CVR of the affected regions was significantly lower than the normal regions. Figure 3 shows that multi-PLD PCASL achieved the higher agreement with PET based on higher concordance correlation coefficients. Discussion: Multi-PLD PCASL had the highest agreement with the reference PET modality in CVR measurements. Both single and multi-PLD ASL were effective in detecting impaired CVR in Moyamoya patients. The post mortem determination of brain edema is routinely performed by forensic pathologists during autopsy by rating macroscopically visible signs [1] . An objective and equally reliable method uses the normalized cerebral weight, which is calculated by dividing the brain weight measured at autopsy by the intracranial volume segmented in computed tomography (CT) images [2] . In this work, we evaluated the sensitivity of quantitative magnetic resonance imaging (MRI) for the assessment of brain edema by correlating the values of post mortem T1, T2, T2*, fractional anisotropy (FA) and mean diffusivity (MD) with the normalized cerebral weight. Methods: In this study, 18 deceased were included. A head CT scan was performed and the MRI protocol, shown in Table 1 , was applied to the brains in-situ using a 3 T scanner (both Siemens, Siemens Healthineers, Erlangen, Germany). For every examined MRI parameter, the values in the automatically segmented cortex, white and deep gray matter were determined separately. To evaluate the agreement of the MRI parameters with the normalized cerebral weight, the correlation coefficient r and the corresponding p value p were calculated. Image and statistical analysis were performed using FSL (FMRIB Software Library, Analysis Group, [3] ) and MATLAB (The Math-Works, Inc., Natick, MA, United States). The obtained correlations and p values are given in Table 2 . The correlations between the normalized cerebral weight and T1, T2 and MD, respectively, are significant in the cortex. FA exhibits a significant correlation with the normalized cerebral weight not only in the cortex, but also in the white and deep gray matter. Discussion: Especially FA values and the cortex are suitable to detect brain edema post mortem as significant correlations were detected in all regions for FA as well as in the cortex for almost all MRI parameters. However, potential influencing factors like the post mortem interval, the age at death and the brain temperature of the deceased during MRI scanning were not evaluated in this study and need to be carefully investigated in future work. Introduction: Glioblastoma multiforme (GBM) is the most aggressive manifestation of brain tumors [1] . The prognosis and survival remain poor and tumors generally recur after standard surgery and coadjuvant radio-chemotherapy [2] . Many studies have been done with GBM models, identifying magnetic resonance imaging (MRI) parameters to be used as non-invasive surrogate biomarkers of this pathology [3] , which are able to determine the therapeutic response during the initial stages of treatment or even before. Anti-inflammatory (AI) agents have been studied as anti-tumoral drugs [4] and tested in GBM models with promising results [5] . Furthermore, the modulation of inflammation in GBM by AIs may improve responses to current treatments [6, 7] . In this study we aimed to assess the MRI tumor features evolution in GBM-bearing rats treated with a nonsteroidal AI to investigate its impact on the evolution and treatment of the disease. Methods: GBM was induced in Sprague-Dawley rats by intracranial injection of C6 cells [8] . Animals were subcutaneously (s.c). injected with the AI drug Meloxicam (2 mg/kg) or saline (control) during 15 days, starting the treatment 5 days after tumor cell injection. MRI evaluations were carried out in a 7 T system until day 27 post-cell injection (or until end-point criteria). Tumor volume was followed-up with Gd-T1W images at different time points. Multiparametric MRI studies with magnetization transfer, diffusion-weighted, T2W, T2*W and dynamic contrast enhancement images were acquired at an early and an advanced tumor stage. Parametric maps were generated with home-made software developed with Matlab. Results: So far, we have observed lower MTR% and axial diffusivity (AD) values in the tumor area of treated rats compared with control rats at the advanced tumor stage (Fig. 1) , while fractional anisotropy (FA) values were higher in the peritumoral region. Discussion: These results could suggest that treated animals presented a greater structural integrity of the tumor tissue, which means the absence of characteristic necrosis of a developing cancer, while in the peritumoral region there was not so much accumulation of cells that would occur in an proliferation process, pointing Meloxicam as potential candidate therapy against GBM. [1] . Juvenile Huntington''s disease (JHD, onset B 20 years) is even rarer and it is associated with large CAG repeats ([ 60) . The clinical features of JHD are heterogeneous, and differ from those characterizing the adult form (e.g., rigidity and seizures are more common than chorea). For this reason, it is crucial to monitor and characterize patterns of disease progression. In this work, we propose an hybrid-imaging approach for the longitudinal assessment of brain changes in JHD, using multimodal PET-MRI data fusion (i.e., merging information regarding structural, functional and metabolic alterations). Methods: A 20-year-old female with JHD underwent 3 T PET-MRI (Siemens Biograph mMR) at two timepoints (18 months follow-up). The protocol included FDG-PET as well as the simultaneous acquisition of T1-weighted and resting state functional MRI (rs-fMRI) data. Figure 1 shows the proposed analysis workflow, carried out using FSL and AFNI, applied over single-subject longitudinal data. Briefly, voxel-wise percentage changes at follow-up relative to baseline status were calculated on (i) grey matter (GM) density, (ii) FDG-PET uptake, (iii) rs-fMRI regional homogeneity (ReHo), and (iv) rs-fMRI amplitude of low frequency fluctuation (fALFF). Results: Figure 2 shows voxel-wise maps of percentage changes at follow-up in the single JHD patient. GM density decreased on average over the whole brain of 31.9 ± 21.1% (median [interquartile range] = 28. 7[14.2-47.5] ). Percentage change in FDG uptake was increased in the frontoparietal network, while widespread reduction was observed in the remaining regions (including thalamus). ReHo and fALFF followed the pattern of PET changes, also highlighting increased functional connectivity in the frontal cortex opposed to decreased connectivity in posterior regions. (1, 2) . Diffusion tensor imaging metrics were extracted from the cingulum bundle, intersecting these key regions. Regression analyses were performed to assess the associations between sgACC microglial activity and functional connectivity, structural connectivity, and HAM-D scores. Results: We found significantly increased [18F]-FEPPA uptake in the depressed patients compared to the healthy controls in the left sgACC. Controls compared to patients showed an overall increase in connectivity between the sgACC and the insula, and functional connectivity was a significant non-linear factor in explaining the microglial activity in the left sgACC. This left sgACC microglial activity was also linearly associated with structural connectivity of the cingulum bundle measured by diffusion tensor imaging. Furthermore, [18F]-FEPPA uptake predicted severity scores of HAM-D. Discussion: In this pilot observational study, we show that there are distinct differences in microglial activity between MDD patients and controls that are associated with altered functional connectivity between the sgACC and the insula and which can influence the structural connectivity of the cingulum bundle. The non-linear relationship between microglial activity and functional connectivity may indicate peaks of microglial activity that correspond with specific functional outcomes. This study, albeit preliminary, suggests that neuroinflammation relates to both functional and structural brain network function in MDD. [1] . This study aimed to use cerebral blood flow (CBF) measured with ASL and a Z-scoring approach to quantify focal brain damage and left/right symmetry in retired Canadian Football League (rCFL) players. It was hypothesized that subjects would have asymmetric CBF and regional hypoperfusion due to their history of repetitive head trauma. Methods: Seventeen rCFL players (100% male, aged 58 ± 6.15y) were scanned using a 3 T GE Discovery MR750 MRI and a 32-channel head coil. 3D T1-weighted fSPGR and 3D pseudo-continuous ASL (pCASL) scans acquired. The pCASL scans were processed using ExploreASL to perform brain segmentation, motion correction, spatial normalization to MNI152 T1 1.5 mm space, partial volume correction, and spatial covariance (CoV) CBF quantification ( Fig. 1 ) [2] . CoV was expected to be more robust with compromised cerebrovasculature and physiological differences [3] . Twelve concussion-related regions-of-interest (ROIs) were selected from the Harvard-Oxford (HO) and Hammers (HS) Atlases (Table 1 ) [4] . ROI CBF was analyzed for the left and right hemispheres separately and bilaterally. Prior to Z-scoring, ROIs that failed Shapiro-Wilk normality testing or had insufficient data were excluded. A Z-scoring approach was applied in MATLAB with the bilateral CoV value as the comparable metric. Z-score outliers that fell 2, 3 or 4 standard deviations from the bilateral means were classified as mild (1), moderate (2), or severe (3) injury burden (IB). A subject-specific IB symmetry index was calculated as: (L-R)/(0.5(L ? R)). Normality, multiple linear regression, and correlation tests were performed with RStudio. Age, type of positional head impact, and career length were used as covariates. Results: The cingulate gyrus (anterior) (IB = 6), parahippocampal gyrus (anterior) (IB = 5), and inferior temporal gyrus (posterior) (IB = 4) had the greatest IBs. IB was symmetrical (-0.02 ± 1.31) with equivalent mean IBs (left = 0.94 ± 1.20; right = 1 ± 1.41), but almost twice as many positive Z-scores (positive = 18; negative = 10). Left IB and negative Z-scores had a significant variance (p = 0.0102) and correlation (r = ? 0.695), and right IB and positive Z-scores had a significant variance (p = 0.00377) and correlation (r = ? 0.773) (Fig. 2) . Also, age was significantly correlated with total IB (r = ? 0.526). Discussion: IB was overall symmetrical, but subject-wise IB was often unilateral. In agreement with literature, the cingulate cortex may be especially vulnerable to injury [4] . We also found indications of left ROI hypoperfusion and right ROI hyperperfusion. Further studies may analyze concussion CBF in relation with DTI and fMRI. References: Purpose: To assess changes in perfusion and volume of the brain substance in clinically isolated syndrome (CIS). Materials and methods: The MR study was carried out on a MRscanner ''Ingenia'' (''Philips'') 3 Tesla. The study included 12 healthy volunteers and 6 patients with demyelinating disease of the central nervous system-with clinically isolated syndrome (CIS) and 13 patients with multiple sclerosis (MS). To assess perfusion using the method of dynamic susceptibility contrast (DSC). Quantitative and qualitative assessment of CBF and CBV in white and gray matter of different lobes of the brain. To assess morphometry, the obtained T1-WI and FLAIR images were loaded into an automated system for calculating the volume of brain structures based on the segmentation method. White matter brain (relWMV) and gray matter (relGMV) volumes were calculated relative to total intracranial volume as a percentage. Results: A moderate correlation was found between a decrease in the volume of the white matter of the brain and the severity of focal changes (coefficient of r-Spearman correlation 0.4, p B 0.05), compared with CIS in patients with relapsing-remitting multiple sclerosis (RRMS) in remission and patients with secondary-progressive multiple sclerosis (SPMS) significantly reduced by 32% and 40% relCBF (p B 0.05), compared with the CIS group in patients with RRMS in the stage of exacerbation and remission, the volumes were significantly reduced by 8% and 11% (p B 0.05). Discussion/conclusion: Based on the data obtained, it follows that the pathogenesis of CIS and MS is based on identical processeschanges in perfusion and neurodegeneration, starting at the earliest stages of the disease. Little is known about differences between small and large-vessel health across age in SCD. Cerebrovascular reserve (CVR) is a measure of hemodynamic compensation and vascular health. Our aim was to investigate the agreement between CVR from 4D flow MRI (large vessels) and arterial spin labelling (ASL) MRI (small vessels). Methods: 10 adults with SCD and 8 healthy controls were included. Images were acquired at 3 T (Philips Healthcare, Best, NL) and included 3D anatomical, 2D multi-slice single-delay pseudo-continuous ASL, and 3D time-resolved phase-contrast MRI (4D flow MRI) (0.5 mm isotropic resolution, 4 cardiac frames) sequences. Acetazolamide (ACZ) was administered intravenously after baseline images. ASL CVR was calculated as %change in grey matter cerebral blood flow (mL/100 g/min) from baseline and post-ACZ ASL. Time-averaged velocity (m/s), peak time-frame flow (mL/s) and peak timeframe lumen area (mm 2 ) perpendicular to each vessel were calculated from 4D flow MRI. 4D flow CVR (%) was calculated as %change in flow over both internal carotid arteries. Wilcoxon rank tests were used to compare groups (p \ 0.05 statistically significant). Agreement between ASL CVR and 4D flow CVR was assessed using Bland-Altman and regression analysis. Results: While velocity did not differ between groups, flow was higher in most vessels in SCD compared to controls as well as lumen area (Fig. 1) . ACZ induced an increase in flow and lumen area with smaller changes noted in SCD patients. Results from orthogonal regression and Bland-Altman in Fig. 2 . shows that there was poor agreement between ASL-CVR and 4D flow CVR particularly for higher CVR. Interestingly, there was a positive correlation with age for 4D flow CVR but not for ASL CVR (Fig. 3 ). Discussion: The discrepancy between ASL and 4D flow CVR could be related to technical errors: (i) noise propagates when calculating CVR as % change; (ii) ASL covered the whole grey matter but 4D flow was only assessed in anterior vessels; (iii) flow changes and downstream perfusion changes should have similar physiological responses but flow territory mapping would be better suited to determine the exact correspondence between flow changes in a large vessel and its specific perfusion territory. The correlation with age was only significant for 4D flow CVR. Although interesting, this could be inclusion bias since relatively healthier SCD patients are more likely to participate in research. Recent studies show dilation of the MCA to CO 2 on high resolution 7T 3 , and 4D flow MRI 4 . Here we also found small increases in area (* 2 mm 2 ), and conclude that the large conduit arteries of the circle of Willis are capable of dilation in response to ACZ. 1 Kassim, Blood (2016). 2 Ohene-Frempong, Blood (1998). University of Copenhagen, Center for Translational Neuromedicine, Copenhagen, DK Introduction: Crucial for brain functions, the glymphatic system encompasses brain metabolite and waste transport involving an exchange of CSF with interstitial fluid [1] . CSF flows into the brain via periarterial spaces facilitated by astrocytic aquaporin-4 (Aqp4) channels at the astrocyte end-feet and disperses into the neuropil [2] . MRI is the most suitable for in-vivo imaging of the whole brain, but CSF space imaging often requires tracer injection altering CSF dynamics. Herein, we present fully noninvasive high-resolution livemice CSF volumetry using 3D constructive interference steady-state (3D-CISS) [3] at 9.4 T. We investigated if 3D-CISS can disclose subtle differences in CSF volumes between Aqp4-null (KO) vs. wildtype (WT) mice, and compared imaging with 50 vs. 33 lm 3 isometric voxels. Methods: Aqp4-KO and WT littermates (n = 30; 40%F, mean ± SD 13 ± 3 (range 9-20) weeks age, 25.8 ± 3.6 g; Danish Animal Experiments Inspectorate approval) underwent 9.4 T 1 H-MRI (BioSpec 94/30USR, Bruker) using: 1) 1500 mT/m gradient coil (BFG6S, Bruker) with room-temperature volumetric Tx/Rx resonator (in. diam. = 40 mm); 2) 240 mT/m gradient coil (BGA-12S) with cryogenically-cooled surface Tx/Rx quadrature-resonator (CryoProbe). All animals were under anesthesia (Ketamine/Xylazine, 100/10 mg/kg, 11.5 ml/kg/45 min) with monitored temperature and respiration. Every 3D-CISS image was calculated as a maximum intensity projection from at least 2 realigned 3D-TrueFISP volumes, acquired in the opposite phase encoding direction (Table 1) . For every 3D-CISS volume (Fig. 1) , CSF space was segmented and the CSF/ brain volume ratio was calculated in few steps: • 1) Semiautomatic brain volume extraction in ITK-SNAP • 2) Automatic: • Brain volume box-bounding for removal of remaining residual regions • In-house CSF space separation in Matlab R2019a: • Adaptive CSF separation based on cumulative distribution of the brain intensities • Slice-wise 2D region-growing for CSF boundaries correction based on contrast calculation [4] , in all 3 planes. Obtained CSF/brain volume ratios from KO and WT animals were compared with Kruskal-Wallis ANOVA (Dunn's correction). Results: The brain volumes were significantly different between combined from 2 MRI setups KO and WT groups (Mann-Whitney: U = 64, p \ 0.05, med. KO-WT = 17.8 mm 3 ; Fig. 2C ). Automatically segmented CSF/brain volume ratios were different between KO's and WT's (Kruskal-Wallis: KS (4,30) = 8.4, p \ 0.05), but the difference was visible only between the results from CryoProbe (p = 0.036; Fig. 2D ). Discussion: 3D-CISS MRI using CryoProbe is feasible to depict subtle differences in CSF volumes between Aqp4-KO and WT mice. Despite smaller voxels, reduced averaging and longer TEs, the setup provided good differentiation without invasive tracer injection. S2.P14. Assessment of aquaporin-4 role in brain activity detected by diffusion magnetic resonance imaging Introduction: AQP4 is a transmembrane water channel highly expressed in central nervous system, regulating fluid exchange [1] by water transport between two sides of plasmatic membrane, and depending on concentration gradients of solutes [2] . Previously, we have shown that functional diffusion MRI can detect cellular swelling associated to glucose uptake [3] . On these grounds, role of APQ4 in this process was studied. Methods: Sodium salt of AQP4 inhibitor TGN-020 [4] was used to assess cerebral changes after glucose bolus administration in C57BL6/J adult male mice with TGN administration plus glucose, saline administration plus glucose, and saline administration only. Diffusion tensor imaging studies were acquired in a 7 T equipment and mean diffusivity (MD) and fractional anisotropy (FA) were measured in hypothalamus, hippocampus, and cortex regions. Temporal distribution of MRI studies and administration of treatments: basal diffusion study was acquired without any dispensed treatment. Then, inhibitor (100 lL/25 g per animal of 0.24 M TGN-020) or saline (300 lL/25 g body weight (b.w.) of saline) were administered intraperitoneally through a catheter. After 20 min, glucose (200 lL/ 25 g b.w. of 2.08 M glucose in saline) or saline were injected and subsequently diffusion study was performed. 30 min after, last diffusion study was acquired. In a parallel cohort, we acquired 1H High Resolution Magic Angle Spinning (HRMAS) spectra in a 11.7 T system. Statistical analysis: for MRI data, parametric maps are still acquaring. For HRMAS; a linear mixed model was done. Fig. 1 Results: Glucose group experienced a decreased in MD 30 min after glucose injection. However, this response seems to be partially inhibited in TGN group. Metabolic studies show a decreased glutamate concentration in cortex and hippocampus of TGN group, as compared to glucose and control groups, while GABA concentration is higher. Fig. 2 Discussion: Glucose administration induces a cerebral response detectable with DTI. Use of TGN inhibits AQP4 and partially avoids its function, noting that AQP4 is involved in cellular volume regulation. Metabolic results correspond with an alteration of the glutamate cycle due to AQP4 obstruction, that yield a net increase of this metabolite concentration. Introduction: Using MRI to non-invasively study animals over a prolonged course of disease development and treatment would eliminate the need to sacrifice similarly sized groups of animals at each given time point during a study (1) . However, how to quantify tumor growth and treatment response non-invasively poses a challenge to all tumor models. In this animal study, we examined the value of automated texture analysis and visual structure assessment of images derived from a clinical 3 T scanner to estimate the tumor burden of heterotopically implanted glioblastoma multiforme in mice. Methods: This study received ethics committee (Thüringer Landesamt für Verbraucherschutz) approval. Xenografts were established in 18 immunodeficient mice by subcutaneous injection of human glioblastoma cells into both flanks. The animals were followed up on a clinical 3 T-MR using a 25 mm Litz volume resonator and were finally sacrificed for histological work up. 3D TSE SPACE and 3D T1-weighted (T1w) FL3D-VIBE sequences before and after contrast agent injection were obtained. The tumors were divided into three groups based on histology and immunocytochemistry (Ki67) findings. Volumes and signal intensities (SI) were calculated. Qualitative assessment of each MR measurement was adapted from the VASARI (Visually AcceSAble Rembrandt Images) MRI feature set (2) Introduction: Glucose is the main energy substrate for the brain but lactate is also important since it can be used as an efficient neuronal substrate (1) especially during brain activity (2) . Interestingly, there is a shift in lactate perception, that start to be no longer considered as a waste metabolic product but rather as an important molecule for the brain. Indeed, lactate has also been shown to be important in physiological processes such as memory consolidation and long-term memory formation (3). So, if lactate is an important molecule for the brain, could it be neuroprotective? We tested this hypothesis on a neonatal model of brain hypoxia. Methods: Hypoxic-ischemic (HI) damage was produced in neonate Wistar rat (P7) (carotid artery ligation ? 2 h hypoxia, n = 70). Pups received an lactate injection before (HILb) or after HI (HILa), or glucose (HIG) or pyruvate (HIP), or NaCl (HIC), or lactate ? oxamate (LDH inhibitor, HILO). In the last group, pups received 3 injection of lactate (3 h, 24 h and 48 h after HI, HI3L). In the HIL group, some pups received [3-13C] lactate and MRS was performed on brain biopsies. Brain lesions were assessed by DWI at 4.7 T (Bruker Biospec) 3 ? 48 h after HI. ADC and FA were measured. Behavioral tests were also performed (righting reflex, mNSS and NOR tests). Results: Results showed that lactate was neuroprotective when administered after the insult, rather than before. Brain lesion volumes (BLV) (3 h after the HI event) were 40% of the total brain volume (TBV) in HIC compared to 30% in HILa. BLV 48 h after the HI were 19% of the TBV in HIC compared to 8% in HILa. When oxamate was co-injected (inhibition of the metabolic use of lactate), the neuroprotection was lost. BLV were even bigger in HILO than in HIC, suggesting that endogenous lactate itself could be neuroprotective. Moreover, glucose or pyruvate, two other brain substrates, were not neuroprotective. 13C-MRS of brain biopsies 3 h after [3-13C] lactate injection indicated that lactate was metabolized by the brain (13C incorporation into amino acids). Finally, when lactate was injected 3 times, neuroprotection was the most effective: BLV were around 1% 48 h after the insult. ADC values (measured at P7) were also better after lactate injection, indicating a lower edema severity. Conclusion: A single lactate injection induced a 25% and a 58%reduction in BLV compared to HIC 3 h and 48 h after HI, respectively, indicating a rapid and efficient neuroprotection. When oxamate was co-injected, the neuroprotection was completely abolished, highlighting the role of lactate metabolism in this protection. After 3 lactate injections, pups presented the smallest BLV and a complete recovery of neurological reflexes, sensorimotor capacities and longterm memory, demonstrating that lactate administration is a promising therapy for neonatal HI insult. [1] . Moreover, amplitude of the somatosensory network of TG rats at 18 months of age was strongly correlated with the behavioral performance in a working memory task, the delayed non-matched to sample test (DNMS), revealing the importance of this network for a proper cognitive outcome. In this study we present preliminary results regarding the effect of early behavioral training during 8 months on somatosensory networks in non-aged wild type (WT) and TG rats. Methods: The experimental groups were as follows: Non-trained: WT n = 11; TG n = 11; Trained: WT n = 9; TG n = 12. Training and DNMS test were performed as previously described [1] starting at 3 months of age until animals were 11 months old, when MRI was acquired. rs-fMRI was acquired in a 7 T scanner by using a single-shot gradient-echo EPI sequence. Images had 600 volumes of 64 9 64 9 34 voxels and 0.4 9 0.4 9 0.6 mm 3 /voxel with TR = 2 s and TE = 28 ms. Image preprocessing included: slice-timing, motion correction, skullstripping, spatial normalization, spatial smoothing, detrending and regression by motion parameters, and temporal filtering (0.01-0.1 Hz). 30 independent components were obtained using FSL MELODIC [2] considering the whole cohort. Based on our previous studies we selected the somatosensorial network as it was the focus of our analysis (Fig. 1A) . The standard deviation of the time-series (Amplitude) of the component was computed for each subject and the differences between groups were evaluated using Kruskall-Wallis test. Results: The amplitude, as a measure of the magnitude of the BOLD activity, in the somatosensorial network was significantly increased in the trained TG group compared to the non-trained TG group (p \ 0.05). This was not observed between the respective WT groups (Fig. 1B) . Despite the trend to lower amplitude values in non-trained TG versus non-trained WT, no significant differences between genotypes were observed in any of the groups. Discussion: In our previous work [1] significant differences between cognitive-trained TG and WT rats were observed in the amplitude of somatosensory network only at 18 months of age, while in a similar non-trained cohort, functional alterations in somatosensory cortex connectivity were observed from 6 months on [3] . Although no significant differences between genotypes were observed, there was a significant difference between non-trained and trained TG rats, pointing to an impact of cognitive training in the somatosensorial network of TgF344-AD rats, that could compensate for the initial pathology-related alterations in these animals. were measured in the corpus callosum and external capsule (P = 0.012), the cortex (P = 0.0185) and the thalamus (P = 0.0403) of heterozygous mice. Longitudinal diffusivity (k1) was increased in the cortex (P = 0.0236) and the amygdala (P = 0.0157) and radial diffusivity in the cortex (P = 0.0374) and the thalamus (P = 0.0347). The analysis of proton spectra from 19 control mice and 16 Tshz3 ? / lacZ mice showed an increase in the Glycine/H20 and N-acetylaspartate (NAA)/H 2 O ratios (P = 0.0161 and P = 0.0288 respectively). Discussion: DTI results show a general increase in diffusivity in brain structures with reduced expression of Tshz3 (cortex and amygdala) as well as in the thalamus. The increase in k1 suggests that water diffusivity along or within the axons is increased, whereas the increase in radial diffusivity could indicate a modification in the myelin sheath in mutant animals. The higher levels of glycine, an excitatory neurotransmitter and of the neuronal marker NAA point to neuronal dysfunction. The unexpected accumulation of NAA could be linked to an altered axonal transport of this compound which is hydrolysed in oligodendrocytes. Introduction: Obese patients and male murine models fed on a high fat diet (HFD) are known to experience an inflammatory response of the hypothalamus, the main brain regulator of energy homeostasis [1] . Likewise, this inflammation has been detected in the reward and mesocorticolimbic centres [2] . Preliminary magnetic resonance imaging (MRI) studies have supported evidence of inflammation in male mice brain, produced within only 15 days of HFD [3] . Methods: We conducted a study with 18 female mice randomly divided into HFD and control (CTRL) groups. We followed their respective food intake, body weight changes and blood glucose levels during 10 weeks, and performed T2-weighted and magnetization transfer (MT) images (7 T Bruker BioSpec) of the brain, 1, 2, 4 and 10 weeks after diet diversification. Mean values of T2 and MT ratios (MTR) were calculated in the hypothalamus, nucleus accumbens, infralimbic area (ILA) and hippocampus. Statistical analysis was performed by a 2-way ANOVA (diet, area) and post-hoc T tests. Results: Our results show that weight gain and calorie intake were significantly higher on HFD mice, and no differences were detected in blood glucose levels ( Fig. 1 left) . MTR and T2 showed a significant diet*area interaction effect at day 7, with lower T2 values and higher MTR in ILA (p \ 0.05) (Fig. 1 right) . In summary, our work shows that both physiological and MRI biomarkers are altered by HFD, but to a lesser extent than in male mice [3] . However, one week of high-fat diet was sufficient to trigger an increase of MTR and reduction of T2 values in the mouse ILA, on agreement with inflammation. Introduction: Concussion or mild traumatic brain injury (mTBI) is one of the most widespread types of trauma. To date, cerebral metabolic alterations after mTBI are studied poorly, especially when it comes to the levels of neurotransmitters. Spectral-edited magnetic resonance spectroscopy allows noninvasive measurement of the gamma-aminobutyric acid (GABA) concentration as well as the concentration of glutamate and glutamine (Glx). The literature data on the [GABA] change after acute mTBI is based on the measurement of GABA ? levels, where GABA signal is contaminated with macromolecular (MM) compound. The aim of this study is to measure GABA-in the PCC cerebral region in the acute phase of mTBI in children. Nineteen patients with acute mTBI (12-70 h since the injury, 16.2 ± 1.4 y.o.) and twenty-one healthy control (18.5 ± 2.3 y.o.) participated in the study. Philips Achieva 3.0 T was used, standard MRI protocol for TBI patients (T2-, T1wighted images, FLAIR, SWI, DTI) revealed no pathological lesions in brain tissue of any subject. MRS voxel (50 9 25 9 25 mm) was located in posterior cingulate cortex (see Fig. 1 ). MEGA-PRESS pulse sequence without MM contamination was used [1] : TR = 2000 ms, TE = 80 ms, 180-editing pulses applied on 1.9 ppm and 1.5 ppm, NSA = 288 (acq.time * 10 min). Spectra were processed in Gannet 3.1 [2] , GABA/Cr, Glx/Cr, and GABA/Glx values were obtained. Results: The data of two controls and three mTBI patients did not match the inclusion criteria (GABA fit error \ 12%). GABA-/Cr (p \ 0.01) and GABA-/Glx (p \ 0.05) were statistically significantly higher in mTBI group than in the normal group (see Fig. 2 , 3), while Glx/Cr was unchanged. No correlation between the metabolite levels and age was revealed in both groups. Discussion: It was demonstrated [3] that the measurements of macromolecule-suppressed GABA-may be more effective in search for the metabolic changes than GABA ? measurements. In our previous study, we have demonstrated that GABA-/Cr increases in anterior cingulate cortex of children with mTBI [3] . To our knowledge, the mTBI-caused increase in GABA-in PCC, revealed in current study, is reported for the first time and agrees with the previous results. The lack of correlations between age and metabolite levels agrees with the literature data [4] and eliminates possible bias of the results caused by group age difference. This study provides insight into metabolic alterations caused by mTBI and may facilitate better understanding of the long-term mTBI consequences. Introduction: Concussion is a mild form of traumatic brain injury (mTBI) with short-term loss of consciousness. Long-term consequences of mTBI can include headaches, attention deficit disorder, sleep problems, etc. [1] . However, with standard MR-tomography or computed tomography (CT), anatomical changes and abnormalities are not observed, which indicates the need to search for biochemical changes in the brain using magnetic resonance spectroscopy (MRS). The aim of this work is to study the effect of mTBI in the acute phase on the pH of the brain and the concentration of cerebral metabolites available in the methods: PRESS and TE-averaged PRESS. Materials and methods: The study involved 16 patients with mTBI (scanning was carried out in the acute phase-up to 3 days after injury), the mean age was 15 ± 3 years, and 17 healthy volunteers of the same age constituting the control group. Philips Achieva dStream 3.0 T and 32-channel SENSE quadrature coil were used. The study protocol included a sequence of PRESS pulses with the following parameters: TR = 2 s, TE = 80 ms, N points = 2048, BW = 2000 Hz, NSA = 288; and TE-averaged PRESS: TR = 2000 ms, TE from 35 to 185 ms in 2.5 ms steps. The voxel (50 9 25 9 25 mm) was located in the posterior cingulate cortex (Fig. 1) . The values of the central frequency of the signal in the region of 7 ppm were determined, and the pH values were calculated using the expression from [2] . Results: In mTBI a shift in the resonance of the protons in the region of 7 ppm is 0.01 ppm. This corresponds to a 1.2% decrease in pH. At the same time, the change in the chemical shift of NAA and Cr is less by 2 and 3 orders of magnitude, respectively, which excludes an erroneous shift along the frequency axis (Fig. 3) . Changes in metabolite concentrations according to PRESS and TE-averaged PRESS data not found. A significant correlation was also found between myo-inositol (mI) and glutamate (Glu) (Fig. 2 ). Discussion: The revealed correlation can be associated with the correlation obtained in [3] between NAAG and mI in hCI. It is due to the fact that NAAG can act as an agonist of Glu receptors [4] . NAAG binding to mGluR3 receptors and initiates Ca 2? signaling, in which inositol trisphosphate (IP3) is involved, indirectly associated with mI. The lack of changes in the concentration of metabolites in the PRESS and TE-averaged PRESS spectra is also often found in the literature [5] . A decrease in PN may indicate a violation of blood microcirculation and the accumulation of lactate in the area of interest. However, due to the insignificance of the effect (1%), additional studies are needed, which are difficult for MRS methods. Conclusion: 2HG MR spectroscopy has a competing diagnostic accuracy, considering its non-invasive, non-contrast and radiation free nature. It may be an integral part of advanced MR tumor protocol, aiding prognostic assessment and multidisciplinary treatment planning for glioma patients. Introduction: Mild cognitive impairment (MCI) represents a transitional stage between normal aging and dementia. This stage is characterized by a high rate of progress to various types of dementia [1] . In this study, the search for possible metabolic markers is based on proton MR spectroscopy in the posterior cingulate cortex (PCC), a part of the brain with high metabolic activity, playing central role in the default mode network [2] . Consequently, the study of changes in the concentration of metabolites will help to understand the metabolic changes in PCC during MCI. To date, we found no literature data on the change in absolute concentrations of metabolites. The measurements were carried out on a Philips Ingenia 3.0 T scanner, 27 subjects participated in the study: 1)1) Normal (clinically and according to MRI scan)-14 subjects (12 f ? 2 m), aged 47-79 2)2) Patients with clinically diagnosed MCI-13 subjects (10 f ? 3 m), aged 46-83 Spectroscopic voxel sized 20 9 20 9 30 mm was placed in the PCC, a typical location is shown in Fig. 1 . To acquire spectra, PRESS sequence was used with the following parameters: TE = 35 ms, TR = 2000 ms, NSA = 80, the number of spectral points: 2048, BW = 2000 Hz. The Reference spectrum of the unsuppressed water signal was also acquired. Results: Average SNR of spectra of the normal subjects and of patients with MCI were 19.4 and, 19.2 respectively. The values of absolute and relative concentrations of metabolites are presented in the Fig. 2 and 3 . Statistically significant increase in the Cho/Cr, as well as the increase in the absolute concentration of Cho and a decrease in NAA/mI ratio were revealed during MCI. Discussions: The increase in [Cho] indicates a violation of phospholipid metabolism in this cerebral zone. This finding may witness for a shift in the balance between synthesis and degradation of cell membranes towards their degradation [3] in MCI. The decrease in NAA/mI may signify the change in the neuron-glia ratio towards the gliolysis of the brain [4] . The results of the study are promising as a basis to create the non-invasive markers of dementia progression. about the metabolism of the brain. The aim of this study is to define metabolic changes at different cognitive stages in Parkinson''s disease (PD) using 1H-MRSI at 3 T. Methods: Twenty-one patients with Parkinson''s disease dementia (PDD), 37 mild cognitively impaired PD (PD-MCI), and 30 cognitively normal PD (PD-CN) were included in this prospective study. The Institutional Review Board approved our research protocol, and all subjects provided written informed consent. Multi-slice 1H-MRSI data was acquired using point resolved spectroscopy (PRESS) sequence at 3 T (TR/TE = 1000/52 ms). LCModel software was used for quantification of total N-acetyl-aspartate (tNAA), total creatine (tCr), choline (Cho), glutamate/glutamine complex (Glx), and inositol (Ins). Oryx-MRSI [1] was used to create metabolite maps and overlay them onto the reference anatomical MRI after chemical shift correction. Then, the metabolite intensities were corrected for the cerebrospinal fluid (CSF) fraction at each voxel. Metabolite maps were then registered onto the MNI152 brain atlas using FMRIB Software Library (FSL), and the mean metabolite intensities and ratios were calculated at 400 different functional parcellations of the cerebral cortex [2] . A Kruskal-Wallis test followed by post-hoc Dunn's tests were performed to define metabolic differences among the three patient subgroups. Results: The mean age of PD-CN, PD-MCI and PDD patients were 60 ± 9, 63 ± 8, and 71 ± 6, respectively. PDD patients had statistically significantly lower tNAA/tCr at the right visual network (P = 0.0008 for PDD vs. PD-CN, and P = 0.022 for PDD vs. PD-MCI) than the non-demented PD, and at the right frontoparietal control network than PD-CN (P = 0.0006) after CSF correction (Table 1) . Additionally, PDD patients had lower tNAA/tCr at right somatomotor network (P = 0.0007) and lower tNAA/Ins at left default mode network (P = 0.003) than PD-CN (Fig. 1 ). Introduction: Parkin gene (PRKN) mutation is the most common form of autosomal recessive parkinsonism, which causes the familial form classified as PARK2. Proton magnetic resonance spectroscopic imaging (1H-MRSI) provides information about brain metabolism in Parkinson's disease (PD). This study compares the metabolic profiles of PD patients with PRKN mutation with that of the idiopathic PD patients and healthy control (HC) subjects using 1H-MRSI at 3 T. Material-methods: Sixty patients were enrolled in this prospective study (13 PARK2 patients, 30 idiopathic PD patients with normal cognition (PD-CN), and 17 healthy control subjects (HC)). Our research protocol was approved by the Institutional Review Board, and all subjects provided written informed consent. Multi-slice 1H-MRSI data was obtained using the point resolved spectroscopy (PRESS) sequence (TR/TE = 1000/52 ms). LCModel software was utilized for quantification of total N-acetyl-aspartate (NAA), total creatine (tCr), choline (Cho), glutamate/glutamine complex (Glx), and myoinositol (Ins). Metabolite maps were created using Oryx-MRSI [1] , followed by overlaying them onto the reference anatomical MRI after chemical shift correction. Afterwards, metabolite intensities were corrected for the cerebrospinal fluid (CSF) fraction at each voxel. FMRIB Software Library (FSL) was used to register metabolite maps onto the MNI152 brain atlas and mean metabolite ratios were obtained at 100 distinct functional parcellations of the cerebral cortex [2] . A Kruskal-Wallis test followed by post-hoc Dunn's tests were performed to define metabolic differences among the three patient subgroups. Results: The mean age of PD-CN, PARK2, and HC patients were 60 ± 9, 44 ± 4, and 60 ± 6, respectively. PARK2 patients had significantly higher Glx/tCr in the left salient/ventral attention network (SN/VAN) (P = 0.0012 for PARK2 vs HC) than the healthy control subjects after CSF correction (Table 1) . Discussion: Increased Glx/tCr might be associated with glutamate/ glutamine toxicity within the SN/VAN, which would lead to eventual cell death. Primary function of this network is to direct attention to unexpected stimuli. VAN Introduction: The introduction of new treatments has improved the perspectives of patients with brain tumours. To determine the optimal treatment, there is a need for imaging techniques that not only provide information on anatomy but also on physiology of brain tumours. GlucoCEST, with its potential to indirectly measure D-glucose in tissue and blood, is a promising new MR technique to measure changes in uptake after treatment in brain tumours. The uptake curve reflects delivery, transport and metabolism of D-glucose. Although performing glucoCEST is challenging at clinical scanners (with relatively low B0 field strength), several centres have demonstrated this possibility [1] [2] [3] . Here, we describe the development of glucoCEST from phantom to healthy subjects on a 3 Tesla MR scanner. Methods: We first conducted a phantom study to identify optimal parameter settings in the glucoCEST sequence, at the 3 Tesla MR scanner of our PET/MR modality, with a 24 channel head coil (General Electric, Chicago, USA). We used 3D snapshot CEST [4] . The phantom consisted of 9 plastic 50 ml tubes with a pH of 7.4 and D-glucose concentrations of 2.5 mM, 10 mM and 40 mM. Four tubes were combined with 8% cross-linked bovine serum albumin (BSA) to simulate human physiology. Several sequence parameters were explored, including B1 power (1lT, 1,5lT, 2lT and 3lT), radio frequency (RF) pulse (80 and 160 pulses), saturation time (ST) and interval (40 ms ? 20 ms and 20 ms ? 20 ms) and frequency offset (between -100 ppm and 100 ppm with 53 steps). Eight healthy subjects (age range 19-24 years, 2 males and 6 females) were enrolled. They received an intravenous cannula in each arm; one for glucose bolus injection (injection duration of 3.2 min, 25 g dextrose in a 50 ml solution) and the other for regular blood draws to monitor venous blood glucose levels during and after injection. Baseline and post-injection glucoCEST images, as well as dynamic glucose enhanced (DGE) images, were obtained from each subject. Z-spectra and MTR-asymmetry curves for white matter (WM), grey matter (GM) and superior sagittal sinus (SS) were created for MR signal analysis. Results: In the phantom, the strongest glucoCEST signal was obtained in the tube with 40 mM glucose with a B1 power of 1.5lT, 80 RF pulses, saturation module with 20 ms and 50% duty cycle, and frequency offset at 1.2 ppm (Fig. 1) . Figure 2 shows the difference in Z-spectra from DGE images before and after glucose injection in a healthy subject. Discussion: In healthy subjects, glucoCEST was optimized for 3 Tesla MR using a PET/MR modality. Although the detected signal was still limited in healthy subjects, glucoCEST is a promising new MR technique to measure D-glucose uptake in brain tumours. Therefore, we will further evaluate and optimize glucoCEST in patients with brain tumours. Homocarnosine (Hc) is a dipeptide of Gamma-Aminbutyric Acid (GABA) and histidine. Hc acts as neurotransmitter in human brain. NMR signals of Hc and GABA are quite similar and differ only with the presence in Hc spectrum of imidazole protons resonance line. We can receive Hc signals by using pulse sequence MEGA-PRESS (M-P PS) [1] . It is possible to measure pH-value due to Hc signals [2] . Hc levels and pH-value in response to visual stimulation (VS) are estimated in this study. Methods: The study involved 17 healthy people. Scanning was performed on Philips 3 T scanner. Spectra in rest and during continuous 8 Hz flashing checkerboard VS were accumulated (10 min each), by using M-P PS (TE = 80 ms, TR = 2000 ms, NSA = 288). Data from scanner were processed in Matlab using packages FID-A and Gannet. Phase and frequency correction was performed, as a result, resonant lines of creatine protons (Cr) (d = 3.027) in all spectra were reduced to 3 ppm. Then spectra were averaged among participants. The intensity was determined as one of the approximation parameters for Hc (d = 7.0) and Cr (d = 3.027). Maximum values of Hc peaks and their chemical shifts have been found, pH-values were calculated using the equation: ,86 is the logarithm of the acid dissociation constant, dAH = 7,27 ppm is the chemical shift of Hc acid, dA = 6,92 ppmof Hc base [2] . Hc contribution to the peak d = 3 ppm in M-P spectrum at rest was estimated using a model spectrum, as it was done in the work [3] . Hc peaks in absolute values during VS and at rest are shown in Fig. 1 . Discussion: The decrease in pH during VS can be explained by an increase in lactate concentration. The existence of an inverse correlation between these values is proved by various studies [4] . During prolonged VS an increase in lactate concentration in human visual cortex is observed, which confirms our results [5] . (1) in the late 1950s featuring a frequency response profile that is always symmetric around the on-resonance (e.g. see Freeman and Hill (2)). However, it was reported by Miller (3) that various tissues intrinsically exhibit a strong and rather unexpected asymmetry in their frequency response profile, which was attributed to local asymmetric intra-voxel frequency distributions (3, 4) . It has been shown that this asymmetry is especially strong in brain white matter where myelination is the probable cause (4). In a previous work (5), we showed that the bSSFP profile becomes symmetric in the limit of TR ? 0, where voxels ''forget'' about the underlying spectral composition and become apparently ''pure''. Whereas this limit can hardly be reached at high-field scanners, lowfields might offer symmetric bSSFP frequency response functions. Here, the profile asymmetry at 3 T and 0.55 T is compared. Methods: Imaging was performed on the head of a healthy volunteer at 3 T and 0.55 T (MAGNETOM Prisma and MAGNETOM Free.-Max, Siemens Healthineers, Erlangen, Germany), using a Cartesian bSSFP product sequence, with a flip angle a = 10°, TR = 3 ms, FOV = 256 9 192 9 160 mm 3 , imaging matrix: 128 9 96 9 80, yielding 2.0 mm isotropic resolution. At 3 T a 20-channel and at 0.55 T a 12-channel head coil were used. The frequency response was sampled by N = 36 (3 T) and N = 32 (0.55 T) scans with equally distributed linear RF phase increments in the interval [0, 2p) . Scanning took 20 min at 3 T and 17 min at 0.55 T to complete. Results: Representative bSSFP profiles of a white matter ROI are shown in Fig. 1 . Using a TR of 3 ms, the asymmetry is obvious at 3 T, whereas the profile at 0.55 T exhibits almost no asymmetry. As a result, for brain at 0.55 T the symmetric regime of bSSFP is already reached for common TR settings. We have shown that at low-fields a symmetric frequency response function can be expected for bSSFP brain imaging even for common TR values. The limit of ''pure'' bSSFP imaging, as proffered by low-field scanners, is of fundamental interest especially within the context of quantification where any mismatch between theory and experiment can lead to a systematic bias in the estimated parameters. 1 H multi-volumetric imaging. In children with cystic fibrosis, amapping with breath-holding imaging strongly correlated with lung function markers of ventilation impairment 4 . Transition to freebreathing a-mapping may increase its feasibility. In this work, we propose to derive a-mapping from data acquired on a low-field MR with a respiratory self-gated 3D balanced steady-state free precession half-radial dual-echo pulse sequence (bSTAR 6 ). Methods: The index a reflects a ventilation measure derived from parenchyma signal intensity modulations arising during breathing 5 . The value of a expresses the change in volume of a voxel as a function of the whole lung volume change; e.g. a = 1 indicates the ventilation in the voxel is the same as the whole lung mean ventilation, a = 2 that the voxel ventilation is double. a-mapping was evaluated in healthy volunteers and in accordance with the local ethics. Imaging was performed in free breathing on a 0.55 T clinical MR-scanner (MAGNETOM Free.Max, Siemens Healthineers). The bSTAR 6 parameters were: TE1/TE2/TR = 0.08/ 2.1/2.2 ms, hard RF pulse = 100 us, 150,000 radial half-spokes, FOV = 34 9 34 9 34 cm 3 , 300 interleaves, FA = 25°, BW = 1002 Hz/pixel, TA = 5:30 min. Before image reconstruction, a trajectory correction of the acquired data was performed. Six volumetric datasets were reconstructed with an isotropic resolution of 2 mm via retrospective respiratory signal gating. The lung volume of each dataset was determined from 3D deep learning lung segmentations 7 . Image registration was done with Elastix, and the data median filtered 5 (kernel 5 9 5 9 5). Finally, a was computed 5 . Results: Representative bSTAR images and segmentations are shown in Fig. 1 . The sequence shows potential for diagnostic imaging. Exemplary 3D a-maps are presented in Fig. 2 , revealing ventilation homogeneity on iso gravitational planes (coronal view) and a ventilation gradient from anterior to posterior lung. The derived 3D ventilation maps at 0.55 T show image features and quality comparable to breath-holding imaging 4-5 at 1.5 T and other ventilation techniques (e.g. increased ventilation in dorsal lung as quantified with SPECT).The proposed free-breathing scheme offers good prospects for 3D morphological imaging and ventilation mapping. In conclusion, we demonstrated initial feasibility to derive a-mapping from multi-volumetric data acquired during free breathing using bSTAR on a commercial 0.55 T low-field MR scanner. In future studies, the clinical value of the method will be further investigated. Balanced steady-state free precession (bSSFP) is an example of a clinical sequence limited by B 0 dependent susceptibility and specific absorption rate (SAR), especially at high field strength. Typically, for bSSFP the repetition time (TR) is minimized to mitigate off-resonance-related banding artifacts. Due to the reduced B 0 inhomogeneities at low field, large field-of-view can be imaged with less stringent restrictions to TR. Furthermore, at low field, bSSFP benefits from increased T 2 /T 1 (i.e. higher signal) and the possibility to use optimal flip angle settings, due to reduced SAR limits. Functional lung imaging with matrix pencil (MP) decomposition 2 relies on dynamic free-breathing acquisitions with a dedicated ultrafast bSSFP (ufSSFP) imaging sequence in which the TR is kept ultrashort (i.e. * 1.5 ms at 1.5 T). MP-MRI is based on naturally occurring respiratory and cardiac signal modulations in the pulmonary tissue, and from the acquired time-resolved data, maps of fractional ventilation and perfusion are computed. MP-MRI and similar techniques have recently shown compelling results 3 and are highly desirable for longitudinal chronic pulmonary disease management since non-invasive. As previously mentioned, lung imaging at low field might be beneficial. Therefore, this work aims to demonstrate the feasibility of MP-MRI on a clinical 0.55 T low-field MRI scanner. Results: Exemplary images acquired with ufSSFP, bSSFP, and GRE, as well as functional maps are presented in Fig. 1 . The quality of baseline ufSSFP images acquired at 1.5 T appears superior to bSSFP at 0.55 T, in which artifacts are present (see the liver). In terms of visual image quality, bSSFP at both field strengths outperform GRE. Overall, functional imaging with bSSFP is feasible at both field strengths. We demonstrated perfusion and ventilation mapping with bSSFP on a low-cost commercial 0.55 T MR scanner. Further sequence optimization is required to improve MP-MRI at low field. Na coil and a 32-channel 1 H head coil. Acquisition: Two 23 Na sequences (flip angles 30/70) acquired using density-adapted 3D radial projections (4) to obtain total sodium content (TSC). Protocol details in Fig. 1 . Reconstruction: 23 Na images reconstructed with BART (5) using gridding and a thresholded l 1 wavelet for noise reduction. 23 Na images were reconstructed to 1 mm 3 before calculating TSC and T1 Map images using the methods outlined by (6) . DTI images were eddy corrected. Registration: Performed using FSL and gradient unwarping (7). Details in Fig. 2 . Analysis: TSC calibrated linearly based on mean signal in regions in eyes (140 mM) and the corona radiata (38 mM) (6, 8) . An atlas of white (WM) and grey matter (GM) regions was constructed and warped into subject space with FSL. Whole brain mask ''bins'' were created based on the range of FA values appropriate for either WM or GM regions (i.e. 0.4 \ FA \ 0.5). Mean and standard deviation TSC was measured for each atlas region and separated into the FA bins. Results: Figure 3 : (a) representative TSC, T1 and FA images (b) plots for mean TSC and relative volume in GM and (c) WM versus FA. Discussion: This study examined correlative measurements of 23 Na MRI and DTI in a healthy population. Other groups have investigated 23 Na and diffusion measurements in cancer and traumatic brain injury (9, 10) , but this work combines TSC, a registration pipeline with a fine-grained analysis of FA in specific WM and GM regions. In WM near the lateral ventricles (i.e. corpus collosum) there was an inverse correlation between FA and TSC. One possible explanation of this is point spread function of 23 Na MRI perturbing WM regions by high 23 Na signal from the lateral ventricles. This is a pilot study and further examinations of these effects are required. A TR of 24 ms results in a 23 Na sequence which is T1-weighted. A sequence without T1-weighting would require a TR of 120-160 ms: a significant increase in acquisition time. The multi-flip angle approach used here allows a high resolution TSC map in a clinically feasible time. Introduction: By collating repetitive T2*-weighted images into movies, time-lapse MRI allows non-invasive detection and tracking of intravascular moving single cells to study the immune response under inflammatory stimulus. 1, 2 Since the generated contrast of ironlabeled cells is motion-dependent, we address the detectable velocity range by simulating movement-induced blurring. To validate the findings, a rotating phantom system was constructed, and time-lapse MRI of iron labeled monocytes in agarose was performed. Methods: Time-lapse contrast was simulated using MATLAB. A signal voids position was stepwise increased in a synthetic phantom and artificial k-space was filled by fractions of the Fourier transforms of the individual images. The final image was acquired through inverse Fourier transform (Fig. 1a) and then multiplied with an in vivo time-lapse MR image of mice brain for an overlay. To imitate moving cells, a system was built to rotate agar gel phantoms with embedded ION-labeled monocytes inside an MRI-scanner (9.4 T Biospec (Bruker Biospin) with cryogenic probe) (Fig. 2) . A stepper motor (Nanotec) with a mounted planetary transmission is used as drive to achieve cell velocities around 1 lm/s. Repetitive T2*weighted images were then acquired with a gradient echo sequence (in-plane resolution: 67 9 55 lm 2 , scan time: 8 min 12 s) for a static phantom and a rotation of 4.4 9 10 -3 rpm. Results: The superposition of simulated moving cells by assembled artificial k-space and real MRI data accounts for noise, anatomical structures, and imaging artifacts. A direct comparison of real and simulated cells is possible showing that the artificial cells represent the experimental signal voids well (Fig. 1b) . The increasing blurring of time-lapse contrast becomes evident supposedly resulting in a maximum detectable speed of 1 lm/s. Moreover, the MR images of iron-labeled monocytes inside a rotating agarose phantom (Fig. 3b) are in good agreement with the simulations: In general, slowly moving signal voids close to the axis of rotation remain visible. On the other hand, the contrast of monocytes linked to higher velocities becomes increasingly blurred resembling the simulated time-lapse contrast in shape and size. Discussion: Overall, the performed simulations are in line with in vivo data and can be validated by phantom measurements. However, the detectable velocity range of migrating cells is not only dependent on the speed, but also on the contrast in a static position (Fig. 3a) . Hence, some cells approaching the edge of the phantom generate enough contrast, while others are not detectable anymore. We understand all three approaches as valid methods to further analyze the detection boundaries of time-lapse MRI regarding size and velocity of iron-labeled cells, which we aim to specify in the future for our standard time-lapse sequence and different acquisition accelerating methods. Introduction: The increased SNR available at higher field strengths, like 7 T, allows for higher undersampling factors (R) and thus faster acquisitions. SNR is however not only reduced by HR, but also by the g-factor and that limits the obtainable acceleration of conventional techniques like SENSE. Wave-CAIPI [1] can minimize this penalty by spreading the aliasing and thus fully utilising the 3D coil sensitivity profiles. The amount of spreading is limited by the PSF of the sine waves and thus by their strength and amount of cycles. The usage of a highperformance insert gradient [2] allows for better sine waveforms and thus more spreading, even during short readouts. Previous simulation work has shown that this increases the efficiency of Wave-CAIPI for a single axis insert [3] . This work demonstrates highly accelerated Wave-CAIPI acquisitions using a gradient insert while avoiding PSF calibration pre-scans and compares their performance against SENSE. Methods: To compare the effects of the accelerations techniques, phantom and in vivo 0.8 mm whole-brain MP-RAGE scans (matrix 320/236/320, Te/TR/TI = 3.0/6.0/1000 ms, a = 6 deg, TFE factor 300, Bw = 380 Hz) were made using a Philips Achieva 7 T (Best, Netherlands) scanner. Both SENSE and Wave-CAIPI were 8 9 oversampled in the readout direction and undersampled by 3.3 9 in both phase-encoding directions for a total acceleration of 9.9. 10 sinus cycles were played during, with the Z-wave shifted by 90°. To study the effect of wave amplitudes on the Z-axis, the highest requestable strength of 8mT/m was used for the body gradients and both matched and doubled by the insert. Offline reconstruction in MATLAB recreates the gradient waveforms and convolves them with the impulse response measured using a Skope field camera (Zurich, Switzerland) for each channel. BART was used for CG-SENSE reconstruction and RMSE was iteratively minimised for auto-calibration of the exact trajectory [4] . Results: Figure 1 shows the resulting predicted gradient waveforms and how they are obtained from the different gradient axes, without the need for additional pre-scans. The phantom results depicted in Fig. 2 show sharp artefacts produced by SENSE, that are less visible in the Wave-CAIPI #2 acquisition, but failed correct wave unfolding in the lower and higher amplitudes. The third Figure shows that the optimal in vivo Wave-CAIPI acquisition improves on SENSE in certain areas (green arrow) and adds artefacts on others (red arrows). Discussion: Wave-CAIPI is a promising acceleration technique that performs better with higher-performance gradients. The utilised calibration-less reconstruction produces artefacts when the gradients are pushed to their limits, which PSF pre-scans might mitigate. When then utilising the increased performance of a gradient insert, Wave-CAIPI might be attractive in even the shortest of readouts such as EPI. King's College London, Department of Biomedical Engineering, London, GB Introduction: The short, but non-zero, time taken to switch between transmit and receive leads to the dead-time gap in Zero Echo-Time (ZTE) imaging. This gap causes reconstruction artefacts and so methods have been developed for filling the gap 1 . We previously showed that 1D GRAPPA can fill small dead-time gaps 2 . We show here that gap filling is implicit during iterative reconstruction, and that sensitivity maps can be extracted from either a separate lowresolution scan or from the ZTE data using ESPIRiT 3 . This may be beneficial in dynamic acquisitions where acquiring extra data to fill the gap with the correct contrast may be problematic. We simulated a 3D Shepp-Logan phantom, matrix 64 9 64 9 64, radial oversampling 2, sampled first with a complete trajectory (no dead-time gap), and then with a gap of three samples. We reconstructed the complete dataset using sensitivities directly extracted from the data 4 , followed by 8 iterations of cgSENSE 5 . Dead-time gap data was reconstructed using sensitivities extracted using ESPIRiTwith a kernel size of 7 9 7 9 7, calibration region 23 9 23 9 23. The dead-time gap was excluded from the calibration region. Volunteer data was acquired on a 3 Tesla scanner (GE MR750) equipped with a 32-channel head coil (Nova Medical) with a ZTE sequence, 3 mm voxel size, matrix 72 9 72 9 42, RBW ± 32.5 kHz, 2 9 read oversampling, angular undersampling 6, flip-angle 1°, dead-time gap four samples, and included a low-resolution k-space. Data was compressed from 32 to 8 channels 6 . Sensitivity maps were extracted four ways: merging low-and high-res k-spaces ? direct extraction, direct extraction from low-res data, direct extraction from high-res data, and by ESPIRiT from the high-res data. The latter three were then compared to the first as the ground truth. Results: Figure 1 shows the phantom reconstructed with the different methods. The iterative reconstruction shows no dead-time gap artefacts. Figure 2 shows the sensitivity maps for the volunteer data. The merged and low-res only maps are similar, ESPIRiT shows different channel phase (as expected), while the high-res only maps are clearly corrupted. Fig. 3 shows the ESPIRiT and low-res maps lead to correct reconstructions, while the high-res maps result in artefacts. Conclusion: Small dead-time gaps can be implicitly filled during an iterative recon, similar to missing lines of k-space in traditional parallel imaging. High fidelity sensitivity maps are required, but these can be taken either from a separate low-resolution scan or extracted from the data itself using ESPIRiT. The exact size of the gap that can be filled will depend on the coil geometry. 5 Pruessmann KP, MRM 2001 DOI:10.1002 /mrm.1241 6 Huang F, MRI 2008 DOI:10.1016 /j.mri.2007 S3.P8. Quality-aware cine cardiac MRI acquisition and reconstruction from undersampled k-space data Cine cardiac MRI is an important diagnostic tool in cardiovascular medicine. However, cardiac MRI is associated with an inherently slow acquisition process. Reducing the number of k-space profiles and recovering the image via undersampled reconstruction is a common approach to speed up the scan. Several deep learning (DL) undersampled reconstruction methods have been proposed to accelerate image reconstruction. However, the undersampling factor is commonly fixed to conservative values before acquisition to ensure diagnostic image quality, potentially leading to unnecessarily long scan times. Here we aim to address this limitation by assessing image quality from undersampled k-space during acquisition, creating an ''active'' acquisition process in which only enough k-space data are acquired to enable the reconstruction of an image that can pass automated quality control (QC) checks. The image analysis pipeline consists of a DL reconstruction algorithm and a DL image QC step to detect poor quality reconstructions. We simulated an active acquisition process by first creating k-space data from cine short-axis cardiac MR images from 270 UK Biobank subjects. We utilised a similar strategy to [1] to generate synthetic phase and use a radial golden angle sampling pattern to simulate undersampled k-t-space data containing increasing numbers of profiles. These were then reconstructed using a deep cascade of convolutional neural networks (DCCNN) [2] (F.1) and subsequently automatically checked for quality. The QC was framed as a binary classification problem and addressed using a ResNet classification network. Binary image quality labels from 225 images of different levels of undersampling were generated by visual inspection and validated by an expert cardiologist. The ResNet network was trained for 200 epochs with a binary cross entropy loss function. The simulated acquisition terminated when the reconstructed images passed the QC check. Our results show that by using DCCNN for cine cardiac MRI reconstruction, we can pass image QC checks after approximately 4 s of simulated acquisition (i.e. undersampling factor of 4.5) . This would result in a reduced scan time for 2D cardiac cine MRI, which takes * 12 s in our clinical protocol (spatial resol. = 1.8 9 1.8 9 8.0 mm 3 , temporal resol. = 31.56 ms, undersampling factor = 2). For the QC step, the average balanced accuracy, sensitivity and specificity were 0.93, 0.86 and 0.99 respectively. Image quality results are shown in T.1. This work demonstrates the feasibility of a DL-based framework for automated quality-controlled ''active'' acquisition of undersampled cine cardiac MRI without previously defined undersampling factor. Future investigations will consist of linking image acquisition and reconstruction to further downstream tasks such as segmentation, volume curve analysis and estimation of cardiac functional parameters. [1] Haldar, J. P. (2013). [2] Hammernik, K. et al., (2018) . [3] Oksuz, I. et. al (2020) . This work was funded by EPSRC (EP/P001009/1) and HDR UK Health Data Research UK. Introduction: LORAKS (low rank modeling of local k-space) a calibration-less constrained low-rank modeling framework estimates the linear dependence structure from the under-sampled acquisition and allows for more flexible sampling schemes [1] , it imposes support and phase constraints and is expected to provide more accurate reconstruction than classical parallel imaging approaches. This flexibility of LORAKS makes it applicable to Wave-CAIPI [2] , and it has been shown in [3] that LORAKS outperforms SENSE [4] in the reconstruction of this type of acquisitions. For a more general scope, we simulate different Wave-CAIPI parameters to further investigate their impact on image quality, here we present the preliminary results. Methods: A fully sampled GRE of a healthy volunteer was acquired on a 7 T SIEMENS scanner, using the following parameters: matrix size = 240 9 240 9 60, voxel size = 1 9 1 9 2 mm 3 , BW/px = 80 Hz, TE/TR = 11/29 ms. Coil compression from 32 to 16 channels and Coil Sensitivity maps were obtaining using ESPIRiT [5, 6] . Wave acquisitions were simulated from a fully sampled Cartesian data using the wave equation [2] (Fig. 1a) and different wave parameters (Fig. 3) . The k-space data was then under-sampled with a factor of 16 (R y = 4, R z = 4) using a CAIPIRINHA mask [7] (Fig. 1b) and reconstructed using the direct SENSE reconstruction. The same data was also under-sampled using three different masking strategies (Fig. 1b) and reconstructed using a modified version of the opensource LORAKS implementation [1, 8] to allow for 3D reconstructions and incorporating the Point Spread Function formulation. The rank 1,000 and regularization k = 1 were selected as suggested in [3] , neighborhood radius was r = 3. To assess the quality of the reconstructions the NRMSE and High Frequency Error Norm were computed. Results:The image quality improves with higher wave gradient amplitudes and number of cycles (Fig. 2) , as expected from the larger spreading effect and as previously reported in [9] . Nevertheless, we did not observe a marked improvement in image quality when reconstructing with LORAKS compared to SENSE approaches while the reconstruction time of LORAKS is [ 10 times larger (Fig. 3) . Discussion: In this work the Wave-CAIPI and LORAKS parameters were selected arbitrarily. Further work needs to validate the results shown here with more combinations of LORAKS parameters selected in a more informed manner. Since we used simulated data, a comparison with prospectively under-sampled data for the most promising parameter sets will be performed. Introduction: Reducing acquisition time in MRI is a common goal within the MR community. To achieve this, a usual approach is to undersample the k-space and reconstruct the images with the support of involved methods. Both numerical optimization methods and, more recently, Deep Learning (DL) based methods have been reported. Specifically, MoDL [1] is a framework that satisfactorily combines both approaches. However, fully-sampled data are needed for training due to its supervised learning character and the best solution reported (a 10-stage network unroll with weight sharing), it is high computationally expensive. Materials and methods: The aim of this abstract is to design and test a self-supervised DL architecture for the implementation of the solution reported in [2] . The function there optimised included a data fidelity term and a regularization term with a motion compensation operator T(Á). These terms are weighted by a parameter k. T(Á) is a groupwise elastic registration, implemented by means of a pre-trained network [3] . The architecture is fed with 30-frame 320 9 320 pixel multislice 2D cine images. which are resampled to 1 mm 2 and cropped to 160 9 160 pixels. The database is composed by 9-14 slices from 88 patients, divided into 4 subsets, namely, 60% for training, 10% for early stopping, 10% for validation (i.e. selection of k) and 10% for testing. Provision was made so that slices from the same patient belong only to one subset. An Acceleration Factor (AF) of 10 has been used and we have let k range within the interval (0.1, 0.8) . Operator T(Á) consists in three executions of the network in network [3] . Training has a maximum of 300 epochs and early stopping has a patience parameter of 40 epochs Validation selected k = 0.3. Results and discussion: Table 1 presents the results on the test set for SSMoComp vs. MoDL with 5 layers and 10 iterations, trained for 50 epochs. Figure 1 shows a reconstruction example. Results show that SSMoComp is more efficient (9 86) than MoDL, with comparable quality results. MoDL has less trainable parameters and no free parameters, but very large GPU memory requirements. S3.P11. Improving spatial resolution of myocardial T1-mapping using a model based super-resolution reconstruction *F. Cachado 1 , A. Gaspar 1 , R. G. Nunes 1 1 University of Lisbon, Institute for Systems and Robotics / Department of Bioengineering, Lisbon, PT Introduction: Over the last years, T1 mapping has become an important tool for myocardial tissue characterization, including detection of fibrosis1. For a detailed tissue evaluation, a high in-plain resolution and sufficient signal-to-noise ratio (SNR) are required, thus, thick slices are often used, sacrificing the through-plane resolution. To increase it, we propose to incorporate a T1 signal recovery exponential model into a super-resolution (SR) reconstruction. Methods: SR was implemented by considering Low Resolution (LR) acquisitions with a relative offset (half of the slice thickness) for each inversion time TI. Ground truth (GT) T1 weighted (T1w) images were simulated by applying the MOLLI2 exponential model to the MRXCAT3 phantom (matrix size = 180 9 140, and slice thickness = 1 mm). To build LR images (y) with a slice thickness of 2 mm, a convolution matrix A (accounting for the slice profile) was applied to the phantom. Random Gaussian noise was added (Signalto-noise ratio of 40). To reconstruct High Resolution (HR) images (x), we considered both data and model consistency terms (Eq. 1). The exponential signal recovery model in Eq. (2) featuring the parameters: M0, equilibrium magnetization, k, a constant to account for inversion efficiency and T1, relaxation time was used: minfx; ugnormðAx À y; 2Þ 2 þ lambda à normðx À u; 2Þ ð 1Þ where Using the ADMM4 approach to minimize Eq. (1), we first fixed the model parameters and predicted the HR images x, using the closed form solution derived from the generalized Tikhonov regularization applied to the data term, with a regularization weight of 1.0. The next step consists of applying soft-thresholding, using lambda = norm(\ AT,y [ ,?), to the model term to minimize the difference between the x and u images, where u is the T1w image predicted from Eq. (2) after fitting the relaxation model to the previous x estimate. To evaluate the quality of the reconstruction, the mean absolute error (MAE) was calculated for the input LR and reconstructed HR images compared to the GT. Results: The example T1w reconstructed images (longest inversion time), as seen in Fig. 1 , show a MAE of 0.86% whereas for the LR images the MAE is 3.95%. The estimated HR T1 maps were correctly estimated compared to the GT map, as can be seen in Fig. 2 University of Lisbon, Lisbon, PT Introduction: Quantitative MRI provides valuable information regarding cardiovascular pathologies as myocardial T 1 changes may reflect infarction, diffuse fibrosis or inflammation [1] . The clinical approach consists of acquiring multiple T 1 -weighted images with a MOLLI (MOdified Look-Locker) sequence and fitting a 3-parameter model: S(TI) = A-B.exp(-TI/T 1 *); where T 1 * T 1 *(B/A-1) [1] . However, many other factors affect the magnetization state (e.g. effective flip angle and heart rate variability) which lead to T 1 bias. Another issue of conventional cartesian MOLLI is the lengthy breathhold required. To address this, we propose coupling an accelerated golden-angle (GA) radial MOLLI with a model-based regularized reconstruction SALSA. Methods: MOLLI acquisition was simulated using Extended Phase Graphs (EPG) [2] : 5(3)3 scheme with a GA radial trajectory and bSSFP readout (FA = 35°, 2 inversions with TI = 231 and 331 ms, TR = 517 ms, heart rate = 70 bpm). MRXCAT [3] heart phantom was considered with 3 tissue regions: healthy myocardium, fibrotic lesion and blood pool (T 1 = 1000/1300/1400 ms and T 2 = 50/70/250 ms at 1.5 T- Fig. 1 ). Gaussian noise was added (SNR = 40). EPG dictionary computed with T 1 = [150:1:2300]ms and T 2 = [25:25:350]ms. Accelerated GA radial data (R = 4) was reconstructed with NUFFT [4] and SALSA [5] . Considering that the noisy k-space (y) of an image (x) can be described as y = Bx ? n (B is the direct operator and n the added noise); SALSA solves the constrained optimization problem min 1/2||Ax-y|| 2 ? /(v), A = BW and W contains the wavelet elements, using an alternating direction method of multipliers (ADMM). The first term describes the L 2 data-fidelity term and /(v) (if x = v due to variable splitting) is the wavelet-based regularizer term. T 1 maps were obtained by 3-param fit to the accelerated MOLLI images, and by matching the NUFFT and SALSA reconstructed MOLLI images to the dictionary. Different SALSA weights (k) were investigated (Fig. 2) . Results: A reduction of aliasing artifacts present in the MOLLI SALSA reconstructed images is noticeable until an optimal minimum T 1 error point at k = 0.001. Spatial sharpness of the myocardial wall and fibrotic lesion boundaries was also improved at k = 0.001 compared to 3-param fit and NUFFT T 1 maps (Fig. 2 g) . However, increasing k past this will lead to excessive smoothing of the T 1 map and loss of sharpness. Proposed SALSA-EPG resulted in improved T 1 accuracy and precision (Fig. 3 ). Discussion: SALSA-EPG appears to be a promising method to obtain robust and fast T 1 mapping of the heart. Future work includes adding coil sensitivity information to increase acceleration and validating simulation data with phantom and in vivo acquisitions. Introduction: Undersampling speed up DWI that helps to reduce the off-resonance, chemical shift, susceptibility and geometric artifacts 1 ; however, undersampling leads to aliasing artifacts 2 . This paper proposes a new method 2D Dense U-Net for the reconstruction of artifact free DW image from the acquired 1D variable density undersampled k-space data. The proposed 2D Dense U-Net is tested on ten patients'' entire brain volume data. The results show better image quality, both visually and in terms of evaluation criteria than contemporary Compressed Sensing (CS) reconstruction 2 . Method: Figure-1 illustrates a schematic diagram of the proposed method. 2D Dense U-Net has 10 convolution layers with a learning rate of 3 9 10 -5 ; weight decay factor of 0.1; ReLU as an activation function and RMSProp optimizer. These parameters are chosen after an extensive experimentation. Firstly, the variable density undersampled (input) and fullysampled (labels) images are used to train the 2D Dense U-Net. The 2D Dense U-Net (after training) is expected to remove the artifacts from the undersampled image, however, it also updates the originally acquired k-space samples. To retain the initially acquired k-space data, an additional step of k-space updation 3 is applied. Finally, IFFT of the updated k-space provides the solution image. The proposed method is trained and tested on the Oasis-3 DWI datasets 4 acquired from a 3 T Siemens scanner with the following parameters: Scanning sequence = ss-EPI, TR = 14.5 ms, TE = 0.11 ms, and Flip angle = 90°. 2D Dense U-Net training is performed with a training set of 5433 images and a validation set of 605 images while the trained 2D Dense U-Net is tested on 20,160 image 4 . The proposed method was trained on Intel(R) Xeon(R) CPU,128 GB RAM, and GPU NVIDIA GeForce GTX 1080Ti with Python 3.8. The network requires approximately 18 h for training in our experiments. The reconstruction is performed on the whole brain volume with different AF 4 and 6 and (0 B b values B 800) s/mm 2 . Results: Figure-2 shows reconstruction results of our experiments at AF = 6 for a single patient. Comparison of different evaluation parameters for the CS and 2D Dense U-Net at AF = 4 and 6 is given in Table- 1. The results demonstrate that the 2D Dense U-Net provides 57.0% lower mean AP, 52.0% lower mean RMSE, and 24.5% higher mean SSIM values than the contemporary CS for human brain DWI data in our experiments. Introduction: BLADE sequence (also known as the PROPELLER MRI [1] or MultiVane) acquires multiple overlapping rectangular k-space patches, which cover the circular region in k-space while sharing k-space center. The motion insensitivity of BLADE MRI is achieved by retrospective correction for translational and rotational motions. It has been accepted that BLADE MRI reduces the motion artifacts and benefits the scanning process for uncollaborative patients. For collaborative patients, however, the current BLADE protocols for collaborative patients yield suboptimal image quality, mostly due to the trade-offs between scanning time, matrix size, and the sensitivity encoding (SENSE) factor. To improve the image quality of BLADE MRI using the index tensor solvers and to evaluate the MRI image quality in a clinical setting. In this study, we investigated the index tensor notation for common Cartesian and non-Cartesian MRI encoding methods. We generalized the index notation to recent inverse reconstruction methods in MRI, using simulation and in vivo MRI data. Methods: BLADE MRI reconstructions using two tensor solvers (the least-square solver and the L1TV-LAD solver) were implemented on the graphics processing unit (GPU). The BLADE rawdata were prospectively acquired and anonymized before the assessments by two independent radiologists with random orders of showing the images. Evaluation scores were examined for consistency and then by repeated-measures analysis of variance (ANOVA) to identify the potential superior algorithm(s). Results: As shown in Fig. 1 , the complex imaginary is influenced by different coil estimation methods. The simulation showed the structural similarity index (SSIM) of various tensor solvers range between 0.995 and 0.999. Evaluation of images (Fig. 2) showed substantial inter-reader agreement (Cohen's Kappa = 0.63, 95% confidence interval: 0.57, 0.68). The image score of L1TV-LAD was significantly higher than vendor-provided image and the least-square method. The image score of the least-square method was significantly lower than the vendor-provided image. No significance was identified in L1TV-LAD with the regularization strength of k = 0.4-1.0. Introduction: Undersampling can be used to accelerate time-consuming MRI acquisitions, but leads to artefacts in the image domain. Non deep learning based reconstruction processes mitigating this effect can provide diagnostic image quality, but can also require significant reconstruction time and computational resources. Deep learning techniques, on the other hand, have shown significant potential in the MRI reconstruction problems with great results and fast evaluations [1] . While deep learning application have already been proposed in the k-space, we intend to offer in this work a novel approach to simplify the training task and ensure data consistency within the network both at training and evaluation. Methods: All following methods were trained and tested on 4860 knee images from the NYU fastMRI database [3] . The chosen deep learning architecture is a Unet [4] , as it provides an easily trainable convolutional approach while having proven its capacities. To accomodate the coil-separated and complex information, the same neural network weights are applied parallelly to each of these dimensions. The consistency enforcement consists in setting the kspace values that were measured to their original values after the correction, by use of a inverse mask and the addition of the original measurement, as described in the joined figure. In order to evaluate the gain from the consistency enforcement, two Unet are trained with kspace data, one with the data enforcement approach, and one without. The network loss is a combination of an L1-norm in the kspace domain and a modified SSIM [5] score in the image domain. The network trained with the data enforcement method outperforms the other method with a optimum loss of 0.39 compared to 0.45 and the data enforced version converged at least 5 times faster to its final value. Discussion: The convolutional approach, although fairly close to the GRAPPA kernel approach, might be improved upon by designing a network architecture with intrinsic kspace properties and chosen trajectory in mind. Nevertheless the trajectory masking technique can be applied to any such approach, and will help shortening training phases. Furthermore the use of complex numbers in both kspace and image space, will be ultimately fit for phase contrast applications. Introduction: Methods such as MOLLI 1 have had a major impact in the clinic for myocardial tissue characterization, but warrant further improvements (e.g. increased slice coverage, sampling flexibility). These can be hindered by the need for access/knowledge of the vendor specific programming environments. To overcome this limitation, we introduced the open-source Prototyping Myocardial T1 mapping sequence (ProMyoT1) 2 using Pulseq 3 . Here, we report recent improvements to ProMyoT1, reducing artifacts on the T1 weighted (T1w) images, and implementing an undersampled (US)-ProMyoT1 version for faster acquisition. Methods: Two modifications were performed to the linear filling scheme of ProMyoT1 (github.com/ANG13/ProMyoT1)-(TR/TE = 3.04/1.52 ms, linear ramp-up of 11 pulses, flip angle = 35°, slice thickness = 6 mm, FOV = 200 9 200 mm 2 , and matrix size = 128 9 128). 2 First, a gradient crusher was added before each T1w image acquisition to remove echo-pathways responsible for undesirable image artefacts. Second, a US-ProMyoT1 version was developed to: 1) reduce the acquisition window (390 ms for fully-sampled FS-ProMyoT1); 2) shorten the first inversion time (TI) (TI 1 = 218 ms for FS). The current US-ProMyoT1 enables an acceleration factor of 2 with 24 auto-calibration lines, acquisition window of 231 ms, and TI 1 = 150 ms. FS-ProMyoT1 and US-ProMyoT1 were tested with and without the additional gradient crusher in a Siemens Aera 1.5 T scanner in the ISMRM/NIST phantom 4 . A clinical MOLLI scan was acquired and reference tabulated phantom T1 values used for comparison. Image reconstruction (inverse FFT for FS and GRAPPA for US) and T1 estimation were performed offline using Matlab; for T1 mapping the function lsqcurve was employed (3-parameter model) . Results: T1w images show a reduction of artefacts when adding crusher gradients for both FS and US ProMyoT1-see TI 2 images in Fig. 1 , estimated T1 maps in Fig. 2 . Crusher application impacted T1 values of FS-ProMyoT1 for vial #3 (T1 no Crusher = 984 ± 3 ms vs T1 crusher = 1023 ± 3 ms) ( Fig. 2 and 3) , while T1 values of FS-ProMyoT1 differ mainly for vial #6 (T1 no Crusher = 413 ± 6 ms vs T1 crusher = 340 ± 6 ms). FS-ProMyoT1 did not provide correct T1 estimates for reference values \ 100 ms. Discussion: This work shows that ProMyoT1 can easily be adapted and improved. T1w artifacts which locally hindered T1 estimation were removed with the addition of gradient crushers. This allowed improving T1 estimates for the affected vials especially for US-Pro-MyoT1. The US-ProMyoT1 enabled T1 estimates similar to FS-ProMyoT1 with the advantages of increased flexibility in vivo for different heart rates. Future work will include testing US-ProMyoT1 in different scanners and in vivo for reproducibility. Introduction: Reconstruction of the under-sampled DW-MR data is challenging because the higher b value images usually have low SNR; however high-quality reconstruction is often needed for clinical use. We propose a DWI Cascaded-Net to recover the artifact free DW-MR image using accelerated 1D Cartesian variable density k-space data. Our experiments show improved performance of the proposed DWI Cascaded-Net both visually and in terms of assessment parameters than contemporary U-Net and ResNet reconstruction results. Subjects and method: Figure-1 depicts block diagram of the proposed scheme. In DWI Cascaded-Net (U-Net ? ResNet), initially IFFT is applied to the 1D variable density under-sampled k-space data which provides aliased images. The aliased images (input) and the artifact free reference images (label) are fed to train the cascaded-Net i.e. U-Net ? ResNet 1,2 . Once the network has been trained, the under-sampled unseen data is fed to the network to get the output, which recovers the zero-filled spaces of the input under-sampled kspace data. To retain the originally acquired k-space points, an additional step of data consistency 3 is applied. Finally, IFFT of the updated k-space provides the solution image. The proposed DWI Cascaded-Net is trained by using a training set of 5433 images and a validation set of 605 images. In the proposed method, RMSProp optimizer with a learning rate of 5 9 10 -4 is used for optimizing the weights of the network. The network training was implemented on Python 3.8 using TensorFlow, batch size = 5, epochs = 300 with early stopping criteria and 12 h training. Results: The proposed DWI Cascaded-Net is trained and tested on publicly available OASIS-3 (https://central.xnat.org/)3 DWI neuroimaging data having matrix size = 256 9 256 with TE = 0.11 ms, TR = 14.5 ms and Flip angle = 90°. In our experiments, the b values vary between 0 and 800 s/mm 2 . Reconstruction results of DWI Cascaded-Net, conventional U-Net and ResNet at AF = 4 are presented in Figure- 2. Table-1 shows the results of DWI Cascaded-Net, conventional U-Net and ResNet in terms of evaluation parameters at AF = 2,4 and 6. Discussion/conclusion: Proposed DWI Cascaded-Net shows improved results as compared to conventional U-Net and ResNet. The proposed method utilizes all the advantages of both the networks together without losing the spatial information during training and reconstructs the high-quality DW-MR images. Introduction: MRI of hard biological tissues remains technically challenging due to the fleeting lifetime of their MRN signals [1] . Special-purpose protocols, such as the Zero Echo Time (ZTE) sequences, are arguably most suitable for ultra-short T 2 samples [2] . However, current approaches to ZTE make it incompatible with slice selection. Here we present a new protocol for slice selection with ZTE (SS-ZTE, [3] ). Methods: SS-ZTE sequence is composed by four main building blocks (Fig. 1): a) sample excitation, where a hard RF p/2 pulse is applied after the onset of the slice selection gradient (G ss ); b) slice selection, where a spin-locking (SL) RF pulse (B 1SL ) detuned as c(B 0 ? G ss z 0 ) selectively locks the sample magnetization at slice z 0 , spoiling the rest due to T 2 decay during gradient aplication (slice thickness D = 2B 1SL /G ss ); c) preservation, to protect the magnetization and coherence of the selected slice from being affected by gradient reconFigurations; and d) acquisition, after the onset of the readout gradient and termination of G ss gradient. Results: As a first demonstration of SS-ZTE performance, we have run tests on our 0.26 T DentMRI scanner [1] . The sample is a cylindrical tube filled with 3% CuSO 4 doped water (T 2 = 3 ms). For this T 2 , we can use a simplified version of the preservation block in Fig. 1 (top) which does not require the storage/spoiling pulses. Figure 2 (left) shows 1D projections of slices for G ss = 30 mT/m and several SL pulse durations. SL times above 1 ms result in rather clean Lorentzian slice profiles (square profiles would result from a sinc SL pulse). Figure 2 (middle) presents the dependence of D as a function of B 1SL strength for G ss = 30 mT/m and t SL = 1 ms, with D ranging from 2 to 10 mm. Figure 2 (right) shows three (superimposed) 4 mm slices selected with B 1SL = 160 lT during, t SL = 900 ls and G ss = 80 mT/m. For comparison, we have also included 4 mm slices obtained with 450 ls and 100 lT sinc RF pulses for standard slice selection based on gradient echo (GRE). In Fig. 3 we compare 2D images with slice selection (6 mm) based on standard GRE (left) and SS-ZTE (right). The GRE parameters are the same as in Fig. 2 (right) , with the shortest possible echo time (3.1 ms) . For SS-ZTE, we chose B 1SL = 120 lT, t SL = 1 ms and G ss = 40 mT/m. Acquisition starts 400 ls after SL pulse. SS-ZTE enables a SNR improvement by a factor 2.8. Discussion: SS-ZTE removes one of the main constraints associated to previously existing ZTE sequences: their inability to produce 2D images. A second limitation is the need for a wait time between the excitation and acquisition windows, due to dead times in RF electronics [1, 2] . In SS-ZTE, the preservation block can also be implemented with dynamical decoupling schemes (e.g. CHASE [4] ), instead of SL, making the wait time unnecessary. S3.P19. Implementation of stack-of-spiral arterial spin labeling at 7 T using Pulseq and MR field monitoring Maastricht University, Faculty of Psychology and Neuroscience, Maastricht, NL Introduction: Pulseq [1] is an open-source MR pulse sequence development package for implementing sequences with novel acquisition and sampling strategies, without going through the timeconsuming vendor-specific sequence programming for initial investigation of sequence behavior. Field monitoring with NMR probes [2] enables high-fidelity measurement of the gradient performance and therefore field correction during image reconstruction. In this study, a 3D stack-of-spiral (SOSP) Arterial Spin Labeling (ASL) sequence was developed and tested at 7 T combining these two techniques. Methods: FAIR-QUPISS II labeling module using a tr-FOCI inversion pulse [3] was implemented for the SOSP ASL sequence using Pulseq in Matlab. Acquisition parameters include: TR/TI2/TI1 = 3000 ms/1800 ms/700 ms, FOV = 200 9 200 mm 2 , in-plane resolution = (2.1 mm) 2 , slice thickness = 2 mm, 36 slices per slab, undersamping factor = 1.6, FA = 20°, BWTP = 25. FAIR-QUIPSS II labeling blocks and imaging blocks of the sequence are shown in Fig. 1 . MR images were acquired on a water phantom and a subject. To examine reconstruction performance, we acquired additionally Cartesian-sampled GRE reference images with identical resolution to their spiral counterpart on the phantom. The in vivo scan contained 50 control/label pairs. Field monitoring was performed after image acquisition, using 16 F 19 NMR probes (Skope, CH) placed at the iso-center. GPU-accelerated algebraic reconstruction with up-to-2 nd -order field correction based on 3D SENSE model [4] was programmed in Matlab. Off-resonance maps were calculated using the topup function in FSL by acquiring a pair of echo-planar images with the opposite phase encoding and readout direction. Results: As shown in Fig. 2 on reconstructed phantom images, reconstruction with off-resonance correction effectively reduced geometric distortion and blurring from spiral acquisition, yielding comparable image quality with Cartesian-sampled GRE images. The subtraction images from in vivo acquisition, as in Fig. 3 , showed perfusion signal primarily in the cortical area. The central brain showed darker signal most likely due to over-flipping as can be also seen from the control images. Discussions: The combination of Pulseq and field monitoring makes it feasible to quickly implement and examine the performance of novel MR acquisitions, as shown by our pilot study on SOSP ASL acquisition. The 3D-SENSE based reconstruction with dynamic field and off-resonance correction was successfully implemented, which will be even more beneficial when concurrent field monitoring is performed to correct for physiological fluctuations. The in vivo ASL acquisition indicates non-ideal RF implementation. Next steps include optimizing RF design and improving k-space sampling efficiency. References: 1, Layton et al. 2017 MRM 77:1544 -1552 . 2, Barmet et al. 2008 MRM 60:187-197. 3, Ivanov et al., 2017 MRM 78:121-129. 4, Wilm et al., 2011 MRM 65:1690 -1701 S3.P20. Skull lipid signal suppression by means of an outer volume crusher coil for diffusion-weighted half fourier acquisition single shot turbo spin echo with smooth transition between pseudo steady states at 3 T *A. Arbabi 1 , D. G. Norris 1 1 Radboud University Medical Centre, Nijmegen, NL Introduction: The most widely used method to suppress the bright fat signal in turbo spin echo (TSE) images is the Fat-Sat technique, where a low-bandwidth RF pulse tuned to the fat resonance frequency and an extra gradient pulse to crush the fat signal are used immediately before the imaging scan. All these lead to higher specific absorption rate (SAR), increased scan time, and sequence design complexity. We demonstrate the application of a crusher coil for the skull lipid signal suppression, which works based on the principle of surface spoiling gradients 1 . We also introduce a variation of diffusionweighted half Fourier acquisition single shot TSE with smooth transition between pseudo-steady states 2 (TRAPS-DW-HASTE). This method provides a significantly lower SAR level, the same level of sensitivity, and similar image quality as compared to the standard manufacturer''s sequence at the same field strength. Theory: DW-HASTE is characterized by long echo times (TE) due to a long time block reserved for the inserted diffusion preparation part. To achieve the shortest attainable TE, a centre-out phase encoding scheme is followed, which preserves the full sensitivity for the central k-space region. In TRAPS-DW-HASTE, 180°refocusing RF pulses are used for the early echoes. Afterwards, the nutation angle is gradually ramped down to a fixed 90°flip angle (Fig. 1) . If the magnetization is continuously in a pseudo-steady state, the greatest signal can be obtained. For crusher coil application, a trigger signal with a variable duration was implemented at the start of the diffusion preparation event block. Methods: MRI scans were performed on a whole body 3 T Siemens Prisma scanner (80 mT/m strength and 200 mT/m/ms slew rate) with a 32-channel receive-only head coil in 2 healthy volunteers. Data were collected with DW-HASTE and TRAPS-DW-HASTE for b = 0 (4 reps and 112 s scan time) and 1000 s/mm 2 (10 reps and 280 s scan time): TE/TR = 109/2000 ms, FOV = 200 * 200 mm 2 , * 1 mm isotropic in-plane resolution, 14 axial slices (3 mm), and GRAPPA = 2. A 3D diagonal diffusion weighting scheme was used. Fat saturation was performed by either standard Fat-Sat technique or the crusher coil. SAR value measured on the scanner was 78% and 42% for DW-HASTE and TRAPS-DW-HASTE, respectively. Results: Figure 2 shows DW-HASTE and TRAPS-DW-HASTE example images. Figure 3 compares the fat saturation methods. Discussion: TRAPS-DW-HASTE deposits significantly lower RF power at 3 T while maintaining image quality. The crusher coil can be used in conjunction with any pulse sequence due to its easy implementation. Due to its rigid design in combination with varying head sizes and shapes, it is possible that skull fat is partially spoiled, or parts of the brain tissue are unwantedly crushed. This problem could be resolved by introducing the capability to locally adjust the strength of the crusher field. References: 1. Boer et. al., MRM 73:2062 -2068 ). 2. Hennig et. al., MRM 49:527-535 (2003 . A driven equilibrium approach to in-vivo diffusionweighted imaging of the liver using spatial encoding Introduction: Spatiotemporal encoding (SPEN) is an alternative to Fourier imaging that is robust against distortions and motion [1] . Recent studies have shown the potential of SPEN for diffusionweighted imaging (DWI-SPEN) of challenging body regions like the pregnant mouse abdomen [2] . DWI-SPEN attains TEs [ 35 ms, which hinders its application to the liver at B0 3 7T, where T2£ 23 ms [3] . We propose a driven equilibrium approach [4] to DWI-SPEN by placing a diffusion preparation module in front of the SPEN sequence, here named as DP-SPEN, which obtains in-vivo images of the liver with TEs as short as 26 ms. Methods: The DP module (Fig. 1) consists of 3 non-selective block pulses-p/2x-py-(p/2-x)-and one pair of monopolar diffusion gradients. Spoiler gradients are placed around and after the py and (p/2x) respectively. The tip-up pulse is phase-cycled between (p/2-x) and (p/2-y) to correct for Eddy current artifacts [5] . DP is followed by the SPEN sequence [2] . Experiments on a 9.4 T horizontal bore scanner (Bruker-Biospin, Karlsruhe, Germany) using a 4-element array cryocoil (Bruker BioSpin, Fallanden, Switzerland) for signal reception. Animal experiments were preapproved by the institutional and national authorities. DWI images acquired from water (WP) and brain (BP) phantoms. The WP consisted of a 30/70% mixture of H20/D2O. The BP consisted of a fixed rat brain immersed in Fluorinert (Sigma Aldrich, Lisbon, Portugal). Images acquired with DWI-EPI and DP-SPEN (Fig. 2) . In-vivo DP-SPEN images of the liver (Fig. 3 ) acquired in N = 2 mice (wild type, females, 8 weeks). SPEN images reconstructed as described in [2] . For DP-SPEN, images of the two phase-cycles were combined, using the sum of squares, to obtain the final Eddy current artifact free images. Results: Fig. 2 shows that DP-SPEN provides mean diffusivity (MD) and fraction of anisotropy (FA) estimates that agree well with those of DWI-EPI. For WP, the MD/FA estimate by DP-SPEN was slightly higher in the upper region of the phantom due to an incomplete correction of Eddy current artifacts. In-vivo results in Fig. 3 show that DP-SPEN accurately images the targeted FOV and successfully maps diffusion in the liver and surrounding organs, in both axial and coronal views. Discussion: DP-SPEN can map liver diffusion at TEs that are * 10 ms shorter than the minimum TEs of DWI-SPEN, endowing DP-SPEN images with higher SNR, which is relevant when mapping diffusion in the liver. B-matrix calculation is greatly simplified when compared to DWI-SPEN because shorter TEs minimize diffusion effects from SPEN imaging gradients. Finally, the decoupled diffusion preparation module offers DP-SPEN larger flexibility in the type of diffusion preparation to be implemented. Paris-Saclay University, Gif-sur-Yvette, FR Introduction: By Chemical Exchange Saturation Transfer (CEST) Line Scanning (LS) localized Z-spectra can be acquired in less than a minute [1] . This allows to detect kinetic reactions of enzymes and stereoisomeric changes e.g. mutarotation [2] . Conventionally mutarotation is measured by optical methods, although large sample volumes are required. Further, measurements by NMR often require deuterated environments. In contrast, CEST LS can be performed in small sample volumes under physiological conditions and is therefore a promising alternative. The aim of this study is to measure mutarotation of glucose by CEST LS and to determine the kinetics of its stereoisomers over time. Methods: For the measurements of glucose mutarotation kinetics, a 75 mM solution of a-glucose was prepared in aqua injectabilia (n = 3) and was measured by CEST LS for a total time duration of 4 h. Reference datasets were acquired by a CEST-EPI and -RARE sequence. All datasets were corrected by WASSR [3] . Measurements were performed at 20°C on a 9.4 T small animal MR Biospec system (Bruker) equipped with a 72 mm quadrature coil. To quantify the ratio of the two stereoisomers, the asymmetry for a at 2.1 ppm and b at 2.8 ppm was calculated. Data were fitted by using a modified model of 2-side chemical exchange (Fig. 2a) including a correction factor for the sensitivity of detection c and for the uncertainty of the starting point t d . Based on CEST measurements of a concentration series (15-75 mM) at the equilibrium, the ratio of isomers in the glucose solution of 75 mM was determined at the endpoint of mutarotation. Results: Z-spectra acquired by LS agree very well with data measured conventionally by CEST-EPI or -RARE (Fig. 1) , and lead to a significant reduction of the measurement time (EPI: 1280 s., RARE: 32,640 s., LS: 40 s.). In the LS spectra, a decrease in the peak of the a isomer and an increase in the b isomer is detected over time and the corresponding asymmetry values approach an equilibrium state (Fig. 2) . From the ratio of the two fits, including the correlation between concentration series and asymmetry, the isomer ratio R in the steady state can be calculated (Fig. 3) : R calc = 1.61 ± 0.02 (expected value R theo = 1.78). Discussion: Currently, the quantification of mutarotation kinetics by CEST LS shows an accuracy of approximately 90%. To further improve the analysis, it is necessary to enhance the model by optimizing the sensitivity factor in dependency of Larmor frequency, concentration, pH and ionic environment. Nevertheless, even in a simplified model assuming the sensitivity factor to be linear and equal for both exchanging resonance frequencies, glucose mutarotation can be measured by CEST LS and the isomer ratio can be verified in good agreement to conventionally measured values. Introduction: The impact of eddy-currents (EC) [1] can be corrected by determining the gradient impulse response function (GIRF). Here we test recently introduced open source (OS) calibration pulse sequences [2] on 1.5 T and 3.0 T scanners, to which we added the ability to measure B0 EC with the method proposed in [3] . Methods: We implemented two pulse sequences to measure the GIRF: H1 (M1) [1, 4] and (M2) [3] using the OS tool PyPulseq [5]- Fig. 1 . To speed up the acquisition, we decreased the repetition time while spoiling undesirable echo pathways.The GIRF was measured using each method in two different systems in doped water phantoms, through autonomous MRI on a 3.0 T Siemens Prisma Fit [6] and in a Siemens Aera 1.5 T. The data were processed in Matlab Ò with an inhouse script making use of the GIRFOS_tool [6] . Finally, we explored the ability to measure the B0 EC (H0 response function) with [5] on both scanners. As this estimation was found to be more sensitive to noise, a Savitzky-Golay filter was applied to the measurements and a regularization factor included during the deconvolution process. To test whether a plausible H0 function had been estimated, its convolution with the input gradient to produce a predicted B0 EC phase was compared with the measured phase. Code available at github.com/ ZemaTimoteo/GIRFOS_tool. The magnitude GIRF (H1) functions for both the x and y axes are depicted in Fig. 2 for both scanners and methods; in all cases the low-pass characteristics were observed as expected. As predicted, the H1 is a scanner specific function, and in most cases it was possible to obtain comparable curves using different methods and number of repeats; an exception is the shape of H1 for the x axis at 1.5 T as the estimated shapes differed between methods. Figure 2 also demonstrates that the profile of H1 was similar for long and short TR acquisitions, confirming that undesirable signal contributions could be avoided. The H0 functions depicted in Fig. 3 also display low-pass filter properties for both scanners and for the two measured axes. Comparison of the measured and predicted phases suggest that the implemented scheme is not yet suitable to capture fast phase polarity changes. The best post-processing approach for limiting the impact of noise was found to be scanner specific. We demonstrated the ability to perform fast and reproducible GIRF and H0 calibration measurements. At present, the measured H0 does not yet capture fast changes in phase polarity contrarily to [3] ; a possible explanation is the limitation to use a longer dwell time of 10 ls (4 times larger) compared to the original implementation; the differences between methods also requires further investigations. Acknowledgments: FCT grants UIDB/50009/2020,SFRH/BD/ 120006/2016 and LISBOA-01-0145-FEDER-029686 1 Vannesjo MRM 2013,69. 2 Fernandes ISMRM 2021 . 3 Robinson MRM 2019 ,81. 4 Duyn JMR 1998 ,153. 5 Ravi Op Sourc Softw 2019 ,4. 6 Ravi MRM 2020 S3.P24. Blurring artefact correction on MR images with spiral k-space by modifying k-space trajectory *O. B. YURDAKOS 1 , Y. GOKPEK 1 , O. Doganay 1 1 Health Sciences Institute, Department of Basic Oncology, Izmir, TR Introduction: The magnetic field inhomogeneities results in blurred MR image artefact due to the frequency off set with spiral k-space sampling [1] . The blurring artefact may be introduced by deviations in gradient imperfections that are normally corrected by the use of modified k-space trajectories rather than the original k-space trajectories for improving image resolution in the reconstruction process [2] . However, there is no general methodology to guide the correction of k-space necessary to achieve a robust offline blurring correction. The purpose of this study is to calculate and correct the blurring artefact due to the frequency off-set and/or gradient imperfections by modifying the original k-space trajectory. Methods: The original spiral k-space [3] , was modified multiplying original k-space by the correction functions including Stairs and Dirac Delta functions in k-space without changing the number of point as shown in Fig. 1 . Then, Spearman''s correlation was used to measure the improvement of blurring artefact in comparison to a Reference image that does not include any of resonance blurring. Two sets of phantom images were used with Df = 100 Hz representing the blurred image and Df = 0 Hz representing the reference image. The blurred image was corrected with kernel step size of 0.01 from 1 to 10. A line profile was selected from corrected images and compared to the reference image. Results: Figure 2 shows the blurred image in (a) the improvement the image resolution as a function of Stairs in (b) and Dirac Delta in (c) with respect to the Reference image in (d). with the line profiles from point A to B (e). The best correlation in line profiles between the reference image and the corrected were summarized in Table 1 , was 0.983 showing that the best kernel was 1.421 achieved by Stairs correction with having 4 steps providing an improvement of image resolution of approximately 80% with a small ringing artefact as shown in Fig. 2c . Discussion: Both Stairs and Dirac Delta functions were successful at correction of blurring artefact with an improvement of 80% and 60%. Stairs function corrected image created a ringing artefact along the external and internal side of the phantom edges, seen in Fig. 3c . Dirac Delta function has created the best image in terms of increasing foreground intensity while having slightly less correlation coefficient compared to the Stairs function. Nevertheless, both corrected images have shown better spatial resolution than the blurred image. Introduction: A 3D Magnetic Resonance Fingerprinting (MRF) acquisition scheme has recently been proposed to enable retrospective motion correction for improved quantitative T1, T2 and proton density (PD) mapping [1] . In this study we evaluate the further improvement provided by the use of a convolutional neural network (CNN) to learn and correct for residual motion artifacts [2] . Methods: 3D MRF acquisitions [3, 1] of 26 patients with negative radiological report were included in this study. 13 were paediatric patients (age: 4-10, acquired on an 1.5 T HDxT scanner) while 13 were adults (age: 19-74, acquired on a 3 T MR750 scanner). Motioncorrected k-space data [1] were transformed to image space and matched with a pre-computed dictionary of MR signal evolutions to obtain quantitative maps of T1, T2 and PD [4] . Maps went through a multi-scale 3D patch-based CNN to correct for residual motion artifacts [2] : using artificially motion-corrupted data of healthy adult volunteers, the CNN was trained to output residual maps, i.e. difference between motion-corrupted and motion-free maps, which were then used to produce corrected maps. Synthetic images mimicking conventional MRI contrasts were generated from the T1, T2 and PD maps, both before and after CNN-based motion correction, by using the equations describing signal intensities of clinical T1w FSPGR, MP2RAGE, T2-FLAIR, T2-FSE sequences. One neuroradiologist evaluated and compared the synthetic contrasts before and after the CNN-based motion correction. A 6-point scale (0: not assessable; 5: very good) was used to assess image quality and anatomical detail of each sequence in each lobe, limbic system, basal ganglia and cerebellum. The Wilcoxon signed rank test was used to assess differences in assigned scores before and after correction. WM-GM contrast and SNR were computed [5] on maps of T1, T2 and PD, as well as on synthetic FSPGR, MP2RAGE, T2-FLAIR, T2-FSE images before and after the use of the CNN, by using automated tissue segmentations of conventional T1w images coregistered to the MRF data. Funding: Grant GR-2016-02361693 (Italian Ministry of Health). Results: A significant improvement in image quality was observed by radiological evaluation in images that underwent CNN-based motion correction (p \ 0.005 uncorrected in all comparisons with the exception of the basal ganglia in synthetic MP2RAGE images; Fig. 1 and 2). The use of the CNN determined a significant increase in WM-GM contrast and SNR in most cases (Fig. 3) . Discussion: Residual motion learning and correction with a CNN agnostic to the anatomy of evaluated patients enables a significant further improvement of quantitative maps and synthetic radiological contrasts obtained with MRF. Introduction: Quantitative T1 mapping of the abdomen is a noninvasive technique that can help to characterize pathologies and monitor therapy effects in liver and kidney 1 . However, respiratory motion due to breathing of the patient can lead to inaccurate T1 quantification 2 . For an accurate diagnosis, it is important to minimize such motion artifacts. Commonly, T1 maps are acquired during a breath hold, which is often strenuous for patients. Here, we present a respiratory motion correction approach that uses an image-based navigator for improved correction of respiratory motion. The resulting motion-corrected T1 maps were evaluated in two healthy volunteers. Methods: Data was acquired in two healthy volunteers (m-29y, f-26y) at a 3 T MRI Scanner (Siemens Healthineers). We used a multi-step approach (Fig. 1) for motion-corrected T1 mapping similar as proposed for cardiac T1 mapping 3 . We used a continuous 2D golden-radial trajectory 4 and applied 7 equally spaced inversion pulses during a scan time of 16 s (TR/TE = 4.9/2.2 ms, flip angle = 5°, resolution = 1.3 9 1.3 9 8 mm 3 ). In a first step, we reconstructed real-time images (100 ms per image) using iterative SENSE 5 . The position of the diaphragm was tracked and spline-interpolated to obtain an image-based respiratory navigator. Next, we used this image-based navigator to bin the original k-space data into 10 respiratory motion states for the reconstruction of respiratory-resolved images at higher spatial resolution. Respiratory motion fields were estimated using non-rigid spline-based image registration 6 . In the last step, we carried out a motion-corrected image reconstruction of images at different inversion times (TI images). Motion-corrected T1 maps were calculated as T1 values for each pixel using the inversion recovery Look-Locker concept 7 . Results: The calculation of the image-based navigator was successful in both volunteers. The image-based navigator (Fig. 2) shows 4 breathing cycles during 16 s. Fig. 2 further illustrates the reconstruction of respiratory-resolved images using the image-based navigator. The registered respiratory-resolved images show that motion correction was successful. T1 maps for two volunteers are displayed in Fig. 3 . Compared to the uncorrected T1 maps, the corrected T1 maps show less artifacts at the diaphragm and sharper blood vessels and kidney structures. T1 maps with breath hold are given as reference. Discussion/conclusion: In this study, feasibility for respiratory motion-corrected T1 mapping in 16 s with a voxel resolution of 1.3 9 1.3 9 8 mm 3 at 3 T was shown, using continuous data acquisition and an image-based navigator for motion correction. S3.P27. Image registration shortcut for prospective motion correction using unseparated multiband slices Introduction: When multiband (MB) excitation is used, image-registration based prospective motion correction (PMC) is typically performed after the MB-acquired slices are separated into individual slices. This study explores the possibility of programming a pulse sequence to register the unseparated/collapsed MB slices in a Prospective Acquisition CorrEction (PACE) scheme and explore its effects. In PACE, translational and rotational motion estimates are calculated by registering the current imaging volume to a Reference volume. These estimates are then used to adjust the slice position and orientation [1] . This also results in a time delay, in addition to the mutlislice separation-related delay [2] . Registering the volumes without a prior separation of the MB images will allow for a quicker processing and an increase in temporal resolution. Methods: The study was performed using retrospective reconstruction of scan data acquired from a phantom on a Siemens MAGNETOM Terra 7 T scanner. A 2D EPI with BOLD (blood oxygen level dependent) contrast time series was acquired (30 volumes, 36 slices, MB factor 3, TR/TE 2210/32 ms, matrix size 64 9 64, slice thickness 3 mm). Using a processing algorithm developed in the Siemens' Image Calculation Environment, volumeto-volume registration (as in PACE [1] ) was performed for two different cases: 1) MB slices are unseparated; 2) MB slices are separated using the slice-GRAPPA method. Translational and rotational parameters were then calculated. Results: Figure 1 shows a sample of the unseparated and separated slices. Figure 2 shows part of the results with the x-axis translation and y-axis rotation estimates for the unseparated (blue) and separated (red) image registration cases. Discussion: It is possible to integrate the technique into the processing pipeline. Figure 2 shows that the unseparated slice registration's shift pattern is in close agreement with the separated slice registration, however the parameter values are small due to the phantom study's inherent limitations. Volunteer studies involving motion protocols were planned to validate the method further in brain imaging, but these have been postponed due to ongoing scanner and recruitment restrictions related to the COVID-19 pandemic. The use of a subset of slices instead of a full volume for registration akin to in MS-PACE [3] will also be evaluated. Overall, pending further validation, it shows a potential in applying direct registration of unseparated images for PMC. Figure 1 . An unseparated MB slice and its corresponding slices after slice-GRAPPA separation. Figure 2 . Image registration results for translation along the x-axis and rotation along the y-axis. The red curve shows the separated slice registration, while the blue shows the unseparated slice registration. S4.P1. Separating tissue transit effects to improve robustness of multi-TE ASL-based blood brain barrier integrity measurements Introduction: Arterial spin labeling (ASL) utilizes water as an endogenous tracer and thus can act as a sensitive method for detecting minute and subtle damage to the blood-brain barrier (BBB). The general kinetic model assumes instantaneous exchange of water molecules with surrounding tissue. But even restricted exchange does not fully describe reality, as labeled blood has to first reach the capillaries within a voxel, before exchange with tissue takes place. In this work, we have extended the T2-based two-compartment model [1] by incorporating an additional delay called ''Intra-voxel Transit Time'' (ITT) to account for transit within the voxel. The main goal of incorporating ITT is to separate the two distinct mechanisms of tissue transit and exchange time (Texch). We compare extended model with 2-CM using simulations and in vivo data. Modeling details can be found in a separate abstract (Nr. 085). Methods: Data was simulated for two multi-TE Hadamard [2] (HAD) ASL protocols: HAD-8; SBD: 400 ms, PLD = 200 ms & HAD-4; SBD = 1000 ms, PLD = 600 ms and both with 8 9 TEs = 13.8-207.6 ms, increment of 27.7 ms. Both datasets were generated with arterial transit time: 500 B ATT B 2500 ms, Texch = 10 B Texch B 1000 ms; ITT = 200 ms and CBF = 60 ml/100 g/min. Texch error in terms of percentage deviation was calculated relative to the simulated Texch value. In-vivo data from ten healthy subjects (4 females, aged 28-40 years) was acquired with the same protocol at 3 T (MAGNETOM Skyra, SIEMENS Healthineers AG). Data was analyzed with FSL fabber [3] and Texch was estimated with both models. Extended model also estimated ITT. Mean gray matter (GM) values were calculated. Results: Figure 1 shows the Texch error estimates from the two models. As can be seen, Texch is overestimated with the 2-CM. Figure 2 (A) shows the Texch and ITT maps from a representative subject. Figure 2 (B) shows the mean GM Texch estimated with the two models (P \ 0.002). Table 1 shows mean GM Texch and ITT values. Discussion and conclusion: Data was simulated with ITT = 200 ms and 2-CM not considering ITT, overestimates Texch especially for lower Texch values (\ 300 ms). This shows that fast exchange dynamics-which would be expected in a disrupted BBB-show the highest error and are partially disguised by the ignorance of ITT. This highlights an important issue that if the tissue transit effects are not modelled carefully, this may lead to a low sensitivity of the technique to differentiate a disrupted BBB from a healthy one. The extended model shows stable results for in-vivo data. The mean GM Texch estimated with 2-CM (342 ± 85.3) was 50.6% higher relative to the extended model (227.9 ± 37.9, P = 0.002). Mean GM ITT was 310.3 ± 52.9 ms. Robustness of Texch estimation as a proxy measure of BBB permeability can be improved by separating transit effects from the exchange dynamics. Introduction: Most data scientists spend about 45% of their time on tasks like data loading and cleansing [1] . This is especially problematic in ASL-MRI, which is available in a variety of acquisition flavors, export forms, and differs between vendors and laboratories. ASL-BIDS [2] has recently addressed this issue by defining a standardized data structure. Current DICOM to BIDS conversion tools are still working on a complete support of ASL-BIDS [3] . Here, we introduce a tool within ExploreASL [4] , that converts DICOMs to ASL-BIDS using DCMTK [5] and dcm2niiX [6] . Methods: This tool comprises four modules. The DCM2NII module flexibly converts DICOMs to NIfTI images and descriptive JSONs. In this conversion step, relevant fields ( Fig. 2) are extracted. Full ASL-BIDS rawdata are generated by the second module. Output of the first module is improved by interpretation of vendor specific fields, such as Philips scale slopes and Siemens Phoenix protocols. The tool validates the metadata files for completeness according to BIDS 1.6.0. Module three allows defacing of structural scans. Module four converts the BIDS rawdata to the ExploreASL format. To reliably support a variety of ASL sequences, anonymized single-subject data from 48 previously-processed studies were converted (Table 1) . Integration and unit testing scripts were written to consistently inspect data transfer results. Results: Improved processing parallelization, improved reproducibility, and platform-independence were achieved through dockerization of ExploreASL including the ASL-BIDS conversion tool. ExploreASL v1.7.0 fully supported the conversion of the sequences in the ASL consensus paper and ASL-BIDS 1.6.0. A variety of relevant fields were read from DICOM metadata (Fig. 2) . Testing was done internally using datasets from 10 different scanners of 3 vendors with 3 different sequences and 25 software versions. Discussion: This toolbox improves ASL data curation, creating an easy-to-use tool to convert DICOM source data. The described workflow creates a standardized structure ready to be analyzed and shared between researchers. The strength of this tool is the diversity of supported ASL sequences, vendors, and scanners. A limitation of ASL-BIDS in general is that more advanced ASL sequences are not yet supported. We anticipate that this work will increase research time efficiency and allow pooling multi-center ASL datasets for advanced statistical analyses. Acknowledgments: This work is part of the Eurostars Project ASPIRE 01QE2026A, which is funded by the German Federal Ministry of Education and Research, Innovate UK, and the Netherlands Enterprise Agency. HM and MD are supported by the Dutch Heart Foundation (2020T049). FB and XG are supported by NIHR funding through the UCLH Biomedical Research Centre. Introduction: Arterial spin labeling (ASL) is a non-invasive perfusion MRI scan technique used in pathologies such as Alzheimer''s disease, cerebrovascular disease, and brain tumors [3] . To obtain absolute quantification of cerebral blood flow (CBF) in mL/100 g/ min, the blood equilibrium magnetization (M0b) is required [1] . For that, M0-scan acquisition is recommended [1] ; however, in practice, clinical studies often lack an M0-scan. A control image can be used as an alternative only when background suppression (BSup) is not used. Here, we investigate the possibility to reconstruct an M0 scan from a control image with BS and compare it with acquired M0 scans from the same patients. Methods: Five healthy volunteers (mean age 66 ± 13 years, 1 female) were scanned in two sessions using a Philips 3 T MRI. Each session consisted of a 3D T1w scan, and two pCASL scans with 2D EPI and 3D GraSE readouts: 4 BSup pulses at 716/1949/2875/ 3391 ms after labeling start, inter-slice timing 33.8 ms (2D), and an M0 scan without BSup, details in [2] . ASL and T1w data were processed using ExploreASL [5] . For each slice, the level of BSup in GM/WM was calculated assuming 90°pre-saturation, 95% inversion efficiency of BSup and respective tissue T1-time (Fig. 1) . The slicewise BSup levels were used to correct the control images, generating a pseudo M0 image. Two models were tested: 1) single-tissue, assuming either 100% GM or 100% WM voxels, and 2) mixed-tissue, allowing voxels with GM and WM partial volume-there, the signal was separated using partial volume correction [4] , BSup-corrected using respective tissue models, and then aggregated. Mean relative error between CBF quantified with the pseudo-and true-M0 was obtained for the whole brain in the central slices (3-13 in 2D, 3-12 in 3D) . A voxel-wise within-subject coefficient of variation (wsCoV) in CBF was calculated between sessions. Results: Examples of pseudo-M0 are in Fig. 2 . For 2D EPI, the average error in CBF was 6.1% for mixed-tissue and 14.8% for single-tissue model (Fig. 3) . The mean wsCoV was 5.74% for mixedtissue and 6.75% for single-tissue model. For 3D GraSE, the error in CBF was 7.4% for mixed-tissue and 8.6% for single-tissue. The mean wsCoV was 5.38% for mixed-tissue and 4.22% for single-tissue model. The mixed-tissue approach provided the best accuracy. The between-subject variation was comparable to previous results using a real M0 scan [2] . Lower level of BSup in 2D EPI allows a more robust M0 estimation than with 3D GraSE. Using M0-estimation from the control scans with BSup is a feasible option to CBF quantification if M0-scans are missing. Introduction: Background suppression (BGS) in ASL leads to perfusion images with higher SNR compared to ASL without BGS 1 . BGS is obtained by applying multiple inversion pulses before and during the post-label delay (PLD). The optimal inversion times (TI), and therefore the quality of the BGS, depends on the relaxation times of the underlying tissue and on inhomogeneities of the magnetic fields. Although this results in inter-subject differences, current ASL protocols make use of one set of predefined TI for all subjects, primarily because these inter-scan variations are not known at the moment of scanning. This means that the quality of the perfusion images is not optimal for all subjects. In this work, we develop a feedback loop (FBL) that optimizes the TI of ASL BGS pulses realtime on the scanner, generating individually optimized perfusion images for each subject. Methods: Data acquisition: Experiments were performed in 2 subjects (informed consent obtained), using a 3 T MR system (Philips, NL) with a 32-channel head coil. PCASL data were acquired with a single-shot EPI readout: label duration/PLD = 2050/1750 ms, TE/ TR = 17/4000 ms. Initial TI (683/1948/2980/3597 ms) were optimized via simulations for suppression of CSF, gray matter, white matter and corpus callosum. Feedback mechanism: After each dynamic, label/control images were sent to an external computer via the remote connection software XTC (Philips, NL). On this computer, we developed a Python tool that receives and processes the images in real time. Updated TI were sent back to the scanner and imported during scanning (Fig. 1A) . TI optimization: 4 TI (2 during labelling, 2 after) were optimized in real-time using Nelder-Mead 2 (80 dynamics), such that the label signal was minimized while maximizing the perfusion signal to avoid magnitude subtraction errors for near-optimal BGS, i.e. with m L/C the label/control image obtained with TI = [TI 1 ,TI 2 ,TI 3 ,TI 4 ] T . Results: Figure 1 shows the TI, the cost function and the label images during the FBL scan in 1 subject. Figure 2 shows the averaged ASL images over the last 10 dynamics of the FBL scan for k = 5 (regularized) and k = 0, compared to a scan with and without standard BGS. Figure 3 shows the results for a stimulus scan at the end of a FBL scan, showing a signal increase in the visual cortex both in perfusion and control images. Discussion: The developed FBL scan increased the SNR of the perfusion images by 23% compared to 15% for standard BGS. Regularization prevents the FBL scan to converge to label and control images with opposite signs, otherwise leading to magnitude subtraction artifacts. The improved BGS leads to a control image in which the perfusion signal can directly be appreciated. This patient-optimized approach could therefore be a first step towards subtractionless ASL to allow twice the temporal resolution when monitoring neuronal activation. Introduction: Background suppression (BS) of static tissue signal plays a significant role in successful application of arterial spin labeling to avoid subtraction errors due to subject motion or signal fluctuation. However, common BS approaches can be difficult to set up properly and premise prior knowledge about the tissues T1 distribution [1] . Therefore, a novel approach for adaptive BS in pseudocontinuous arterial spin labeling is presented. Instead of assuming the T1 distribution in an organ of interest, the actual spectrum of encountered T1 values is calculated from the preceding M0 scan, which is then used to suppress the static tissue to a desired level of residual magnetization. Methods: In typical ASL experiments, a M0 scan with multiple saturation times is performed for quantification purposes. In this work M0, T1 and the saturation efficiency M sat are fitted for each voxel from the M0 scan using a monoexponential. A two-dimensional histogram with a bin size of 100 ms for T1 values and 2.5% for the saturation level is calculated from resulting T1 and M sat maps. The resulting 2D histogram is then used to find optimal timings of n FOCI inversion pulses by minimizing the functional as given in Fig. 1a . In short, the algorithm tries to suppress all T1 values present in the histogram to a user defined level of residual magnetization. The algorithm is evaluated in a watermelon and in-vivo by adjusting the adaptive BS to levels of 0%, 4%, 8% and 12% with application of one to four inversion pulses. Additional imaging parameters are given in Fig. 1b . Results: Figure 1c shows a comparison between measured and predicted image contrasts. Figure 2 shows the mean level of suppression for different numbers of inversion pulses. Quality of perfusionweighted images is shown in Fig. 3 . The proposed BS technique produces accurate levels of residual magnetization in the watermelon phantom, even for a small number of inversion pulses (cf. Fig. 1c, top) . For in-vivo data of the brain, the T1 distribution is broader such that only one inversion pulse is not able to produce the desired suppression (cf. Fig. 1c, bottom) . Using four inversion pulses, predicted and measured BS get closer to the desired levels of magnetization. However, larger deviations are observed in contrast to the phantom data, which might result from restricted inversion efficiency in-vivo. This might also explain why BS levels using two or three inversions are closer to desired levels (cf. Fig. 2 ) which needs further investigation. Figure 3 shows that high quality perfusion weighted images are obtained using the proposed BS with levels of 4%-12% and two-four inversions. We conclude, that the proposed technique is able to produce high quality perfusionweighted images while simplifying the BS adjustment in pCASL experiments. Introduction: Arterial Spin Labeling (ASL) is highly promising for non-invasive perfusion imaging in cerebrovascular diseases. 1 Pseudo-continuous ASL (pCASL) measurements in patients with unilateral internal carotid artery stenosis (ICAS) previously showed ipsilaterally decreased cerebral blood flow (CBF). 2, 3 However, lower CBF may also originate from partial volume effects (PVE) 4,5 and arterial transit time (ATT) delays. 1, 6 The aim of our study was to evaluate PVE and ATT bias on pCASL-based CBF in grey matter (GM) of ICAS patients and age-matched healthy controls (HC). First, PVE were assessed by a linear regression algorithm. 7 Second, ATT effects were evaluated by means of spatial coefficients of variation (sCOV). 6 Methods: 15 asymptomatic ICAS patients (70.2 ± 4.4 years) and 24 HC (69.9 ± 7.3 years) were scanned on a 3 T Philips Ingenia (Best, NL). The imaging protocol and derived parameters are summarized in Fig. 1 , including single post label delay pCASL. PVE correction (PVEc) 7 on GM-CBF was compared globally as well as differentiated between anterior and posterior circulation using a perfusion territory atlas. 8 Moreover, CBF asymmetry indizes 2 (AI, Fig. 1 ) between hemispheres were evaluated and sCOV compared between patients and HCs. Results: Exemplary data show PVEc induced GM-CBF increases, with strongest effects in frontal regions of ICAS patients (Fig. 2 ). Group analysis in ICAS showed globally increased GM-CBF (? 10.3 ± 2.5%, Fig. 3A ) by PVEc, most pronounced in the anterior circulation (Fig. 3B) , and preserved CBF symmetry (AI&0,47, Fig. 3C ). HCs'' GM-CBF increased by PVEc (? 9.1 ± 3.1%), without regional differences. Spatial COV was 0.37 for HC and 0.33 for patients (Fig. 3D ). Our results indicate systematic underestimation of uncorrected GM-CBF in patients and HCs, in line with literature. 4, 9 Therefore, PVEc is recommended to better differentiate disease driven CBF changes. Interestingly, PVE of HCs were spatially homogenous, while patients showed strongest effects in the anterior circulation. This may be due to accelerated parenchymal volume loss in areas supplied by branches of a (stenosed) ICA, as ICAS is known to be related to atrophy. 10 Unchanged AI indicates similar bilateral effects and absence of severe unilateral atrophy, while analysis of sCOV indicates no severe delay effects. Prospectively, vessel selective ASL 11 could further differentiate PVE within perfusion territories and time-encoded ASL 12 could exclude delay effects. To conclude, patients' lowered ipsilateral CBF seems to be pathophysiologically driven, as we found neither effects of PVEc on AI nor severe ATT artefacts. Therefore, widely available standard pCASL is applicable for clinical perfusion imaging. Towards free-breathing liver perfusion imaging (using prospective motion correction) However, being a subtractive technique, ASL in abdominal organs is challenging due to breathing motion. Prospective motion correction using additional navigator images, however, is hampered by background suppression used to avoid subtraction errors in ASL. Therefore, a technique for prospective adjustment of saturation and imaging slices of a pCASL sequence is presented, based on motion estimates obtained from a preceding reference cycle and a crosscorrelation matching approach during the ASL experiment. Methods: Prospective tracking of the liver position during an axial pCASL experiment is accomplished by additional 2D EPI navigator readouts as shown in Fig. 1a . First, multiple navigators are acquired over a period of around 10 s, covering the full cycle of breathing motion of the liver (reference cycle) and a look-up table is filled with calculated positional shifts. During the actual ASL experiment, 2D navigator images (ASL navigators) are acquired before the slice-selective presaturation pulse as well as the 3D GRASE imaging module. These ASL navigators are matched into the reference cycle by calculation of the normalized cross-correlation between the navigator and all reference images. After that, translational shifts, corresponding to the position of the image with maximum correlation value, are sent back to the sequence, adjusting the saturation/imaging position. The algorithm is validated for protocol parameters as given in Fig. 1b . A phantom is placed on a movable base (cf. Fig. 2a) , which is shifted continuously during the acquisition of the reference cycle. During the experiment, the phantom is moved between two distinct positions. Motion experiments are performed, with/without application of prospective correction. Results/discussion: Figure 2b shows the motion trajectory estimated from the reference cycle, corresponding to the applied shifts. Figure 2c shows an exemplarily plot of the correlation metric between the first navigator of the ASL experiment and all navigators from the reference cycle. Maxima are clearly visible such that correctly identified corresponding shifts from the lookup table were sent back to the sequence during the experiment (cf. Fig. 2d) . Figure 3 shows successful prospective adaption of the sequence to the changes in positioning of the object when compared to the scan without adaption of the sequence. Additional retrospective in-plane motion correction using a PROPELLER readout will be further investigated. Conclusion: Successful tracking of the objects position during the ASL experiment was achieved using the proposed reference cycle correlation matching technique. Therefore, future work will focus on the application of the proposed technique to free-breathing ASL imaging of the liver. [1, 2] . Methods: Synthetic ASL scans were generated by the ASL-DRO [1] using average CBF and segmentations from the EPAD study [3] as DRO input (Fig. 1) . To study the influence of motion, pseudo-random motion was introduced to the rawdata using the default ASL-DRO. 3D translations and rotations were defined for the ASL time-series starting with zero for the first volume. For the following volumes, a uniformly distributed random number within a specific range was added to the previous volume. The synthetic data were processed with ExploreASL using default settings. Estimated motion and mean CBF were compared with ground truth motion and CBF. Results: CBF results based on the EPAD ground truth (Fig. 1) were 36-58% underestimated. ExploreASL estimated motion with minor errors for most motion settings . Both the euclidean norm as well as absolute displacement ( Fig. 2j -k) display a high effect of rotation on overall voxel-wise motion. The voxel-wise error in CBF increased by 3.8 between -100% and 0% and 19.4 ml/100 g/min between -100% and ? 200% (Fig. 3a) . For peripheral brain regions, the error increases with motion exacerbated by the sharp contrast between brain and background in the DRO (Fig. 3b) . CBF underestimation increased with motion based on increased blurring [4] (Fig. 3c) . The increased effect of motion on individual smaller regions (Fig. 3d ) led to higher variation in CBF. The ASL-DRO proved to be useful for software QC and evaluation. It was found that a motion measure based on displacement of a single-voxel is not representative and mean-voxel motion needs to be used. CBF underestimation was partly caused by input data undersampling, partly by partial volume artifacts. Both the high impact of rotational movement and the increased effect of motion on thinner or smaller regions is visible in the results (Fig. 3a, b) , and the DRO can be utilized in the future to help correct for these artifacts. Integration of DRO based QC workflows is a promising way to improve methods stability and pipeline results. Introduction: Arterial spin labeling (ASL) is a noninvasive perfusion imaging technique with great clinical potential, and a consensus paper has been published with recommendations for its clinical implementation 1 . In particular, a segmented 3D GRASE readout is indicated for optimal SNR. Because this is known to be affected by susceptibility artefacts, appropriate distortion correction techniques are often used 2 . One study showed that distortion correction increased diagnostic precision of ASL perfusion images 3 , but the impact on ASL data quality and ensuing perfusion measurements has not been systematically assessed. Here, we investigate the effects of susceptibility distortion correction on perfusion imaging by pCASL with 3D GRASE in terms of temporal SNR (tSNR) as well as perfusion. Methods: A group of 15 women (20-48 years, 7 controls and 8 migraine patients) was studied on a 3 T Siemens Vida MRI System using a 64-channel head RF coil. Perfusion imaging was performed using pCASL with 3D GRASE (TR = 5.6 s, TE = 18.4 ms, labeling duration = 1.8 s, post-labeling delay = 1.8 s, background suppression, 4 repetitions) 1 . A fieldmap (TE1/TE2 = 4.92/7.38 ms) and a T1weighted structural image (1 mm isotropic resolution) were also acquired. Image analysis was performed using FSL (fsl.fmrib.ox.ac.uk). After motion correction, two preprocessing options were considered: without or with distortion correction (FSL's PRELUDE&FUGUE). Relative perfusion maps were obtained by averaging pairwise controllabel subtraction images with spatial regularisation (FSL's BASIL). The tSNR was calculated voxelwise on the pairwise control-label subtraction images as the temporal mean value divided by the temporal standard deviation (tStd). The structural image was segmented and 8 anatomical ROIs were defined. Registration was performed between the relatively distortion-free structural images and each of the two preprocessed ASL datasets (without and with distortion correction) (FSL's FLIRT). Results: An example of the effects of distortion correction on the tSNR and perfusion maps is shown in Fig. 1 . The tStd significantly decreased and tSNR significantly increased in all ROIs (Fig. 2) . For perfusion, we found significant changes in the occipital lobe, cerebellum and brainstem (Fig. 3 ). Hospital da Luz, Imaging Department, Lisbon, PT Introduction: Arterial spin labeling (ASL) has great clinical potential for the noninvasive evaluation of cerebral blood flow (CBF), as well as other haemodynamics parameters such as the arterial transit time (ATT) if sampling multiple post-labeling delays (PLD) [1] . The intrinsically low signal to noise ratio (SNR) of ASL has motivated the development of various acquisition and post-processing strategies. In particular, Independent Component Analysis (ICA) has been shown to be valuable in differentiating the ASL signal of interest from structured noise sources [2] . Here, we further investigate the impact of ICA-based denoising on ASL perfusion imaging comparing two groups of participants: Small Vessel Disease (SVD) patients and agematched controls. Methods: Multi-PLD pulsed ASL data was acquired from 17 SVD patients and 12 age-matched controls on a Siemens Verio scanner using a PICORE-Q2TIPS sequence with a 2D multi-slice GE-EPI readout (TR/TE = 2500/11 ms; 11 TIs: 400-2400 ms). Data analysis was performed using FSL (fsl.fmrib.ox.ac.uk). ICA was performed on the motion-corrected control-label subtraction time series. Noise-related Independent Components (ICs) were manually identified and removed following 2 approaches: conservative ICA (ICAc)-removal of ICs clearly related to motion, susceptibility and multiband artifacts; and aggressive ICA (ICAa)-removal of all ICs not consistent with the expected ASL signal location/dynamics (Fig. 1) . These two approaches were compared to the non-denoised data (noICA) in terms of the % change in the temporal SD of the subtraction time series as well as the gray matter mean values of CBF and ATT and respective variances estimated by fitting an extended kinetic model to the data using BASIL [1] . Results: The voxelwise temporal SD changes produced by the 2 ICA denoising approaches (Fig. 2) were up to ± 10% greater in patients relative to controls. The CBF and ATT estimation variances were significantly reduced with ICA denoising, indicating an improvement in estimation precision, with a trend for greater effects in patients (Fig. 3) . Although global gray matter CBF and ATT values did not change, ICA denoising did impact the voxelwise statistical analysis between patients and controls: after denoising, significant CBF differences found in apparently artifactual locations were eliminated (results not shown). Discussion: Our results support the use of ICA denoising of multi-PLD pulsed ASL data to improve the quality of CBF and ATT estimates, and further suggest that it may differentially affect data collected from patients and controls, significantly influencing the statistical analysis of group differences in perfusion parameters. ICA denoising appears to have more impact in the patients'' group, most likely because these are more prone to artifacts such as head motion. Physiological noise correction usually demands additional experimental setup to record cardiac and respiratory signals. In EEG-fMRI acquisitions, the ECG is typically recorded anyway; since it is modulated by respiration, an ECG-derived respiration (EDR) signal may also be derived without extra equipment 1 . We evaluate the performance of EDR signals for physiological noise correction in restingstate EEG-fMRI, by comparison with measured respiratory signals. Methods: Resting-state EEG-fMRI data was acquired from 6 female migraine patients on a 3 T Siemens Vida system with a 64-channel RF coil using 2D-EPI (TR/TE = 1260/30 ms, in-plane GRAPPA-2, SMS-3, 60 slices, 2.2 mm iso resolution). Respiratory signal (Resp) was acquired with the integrated BioMatrix Sensors at 400 Hz and ECG as part of the MR-compatible EEG system (Brain Products) at 1000 Hz. Gradient artifact correction 2 , downsampling (250 Hz) and bandpass filtering were applied to ECG. EDRs were extracted with 7 different methods: ECG envelope (ENV), heart-rate variability (HRV), amplitude modulation (AM), QRS area modulation (QRS-AM), principal component analysis (PCA), kernel PCA (kPCA), and empirical mode decomposition (EMD) 1 . The similarity between each EDR and Resp was assessed by correlation and coherence 3 . Three general linear models (GLM) of physiological noise were fitted to the average BOLD signal in grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF): Respiratory only (RETROICOR respiratory terms; respiratory volume per time, RVT), cardiac only (RETROICOR cardiac terms; cardiac rate, CR) and full model with all regressors 4, 5, 6 . The variance explained (VE) was computed based on the adjusted R 2 (R 2 adj ). Results: Among the EDR methods yielding the highest correlation coefficients for each category, EMD, PCA and kPCA achieved the highest correlation values, but they were surpassed by HRV in terms of coherence (Fig. 1) . Regarding the physiological noise models, all EDRs yielded significant VEs and no significant differences were found relative to Resp (Figs. 2, 3) . Discussion: The EMD, PCA and kPCA methods showed the highest similarity with Resp, in line with previous work outside the MR environment. The VE obtained from the EDRs is equivalent to the one obtained from Resp, indicating the feasibility of using EDRs for physiological noise correction in resting-state EEG-fMRI. Although derived from ECG, EDR-based regressors explained more BOLD signal variance than using only cardiac regressors. Functional connectivity (FC) measured by resting-state fMRI (rs-fMRI) can be analysed using graph theory to characterize brain networks' topology over time. Since these are spatially embedded, the question arises whether the structure captured can be explained exclusively by proximity constraints determined by the brain's underlying structure, or whether there is some degree of functional specialization responsible for the patterns found. This question has been recently addressed in terms of the static functional connectome (sFC) [1] . Here, we further investigate the community structure captured over time by analysing dynamic FC (dFC) against a spatially informed null model. A rs-fMRI dataset collected from 9 healthy subjects at 7 T was used [2] . The data was parcellated into 68 regions using the Desikan atlas, and dFC was computed (for each TR = 1 s) using phase coherence [3] ; this was averaged over all time points to obtain sFC. The networks were thresholded by keeping the giant component structure for most time points. Community analysis was performed using the Louvain algorithm, before and after thresholding both sFC and dFC networks, for all time points deviating from a rewiring null model. To investigate the impact of spatial embedding, we used a degree-constrained spatial null model and applied a modified Louvain algorithm regressing out this influence [4] . Finally, the community structure captured with and without the proximity constraints was compared using Normalized Mutual Information (NMI). We found an overall increase in the modulary values when thresholding both sFC and dFC (Fig. 1 ), as expected [5] . However, this increase was substantially higher for dFC, pointing to a more modular structure captured over time than for the temporal average. Comparing against the rewiring null model, dFC modularity was statistically significant for all time points, while about 88% were selected when using the spatial null model, resulting in higher z-score values than for the sFC (Fig. 2) . This suggests that dFC community structure is less influenced by spatial constraints, whereas sFC favors short-range connections expected by the spatial embedding. This observation was further confirmed by comparing the community structure with and without regressing out this influence, since we found higher similarity values for dFC (Fig. 3) . We found that, although the topology and community structure of the rs-fMRI functional networks is mainly explained by the spatial embedding, a significant degree of functional specialization can still be detected which is space-independent, particularly for the dFC. Introduction: For performing robust and generalizable machine learning functional neuroimaging studies, combining multi-site data has been essential. Considering the impact of scan parameters on fMRI images, researchers usually trim multi-site data to the same number of time points. However, the effects of trimming BOLD signal data in terms of functional connectivity (FC) are still poorly understood. Methods: Resting-state functional MRI data from thirty healthy subjects were pre-processed for five different numbers of time points. Individual FC matrices were generated by performing dual regression with the brain template from Shirer et al. [1] . The correlation matrices were binarized for several thresholds, which excluded weak correlations and included both positive and negative correlations for the analysis. To study FC differences between different numbers of time points, network degrees were computed for each brain region, and the degree distributions were compared. In addition, the median degree numbers per subject and per brain region were subtracted between different trimming conditions. Wilcoxon tests were then performed to evaluate which brain region degrees were significantly different between different trimming conditions. Results: Increasing degree numbers for the global network were observed, together with a non-significant right-shifting trend of the degree distributions, associated with fewer time points (Fig. 1 left) . Additionally, when comparing the degree distribution per brain region, it was observed that the majority of resting-state networks show increased degree numbers, whilst only the auditory network showed a decreased degree for fewer time points (Fig. 1 right) . Discussion: The overall increase in the degree numbers for fewer time points, suggests that there are fewer stronger correlations between brain regions with longer scanning times. In fact, some brain networks seem to present a more dynamic pattern than others, namely the higher visual, language, and left executive control networks. In contrast, the auditory, the visuospatial, the basal ganglia, and the dorsal default mode networks present a more stable and less dynamic pattern. This may be related to environmental demands (e.g. at rest during MRI scanning, visual stimuli) that engage some networks to change communication patterns more than others. Finally, although the overall characteristics of the global networks seem to be maintained, care should be taken to account for the trimming effects on FC at the regional level. Introduction: Functional magnetic resonance imaging (fMRI) employs the blood oxygen dependent level (BOLD) effect to determine stimulus induced brain physiological state changes [1] . The BOLD effect leverages hemoglobin''s oxygenation dependant magnetic properties to describe changes in blood-flow and metabolic need [1] . The hemodynamic response function (HRF) mathematically characterizes the BOLD response, to describe neural activity via amplitude, latency, and duration measures [2] . Although BOLD signal temporal correlation with the HRF is the accepted norm to determine neural activation, it is non-ideal because of its assumed canonical form [3] . An assumed HRF model also masks information to be gained from its variability. As such, a model independent method is proposed. Methods: Six participants (5 male, 3 right-handed) performed finger tapping during two fMRI runs. Each run was divided into six repeated sets of two 30 s blocks: (i) right finger tapping (RT) then rest (ii) left finger tapping (LT) then rest. Data was collected using a 32-channel head coil and GE 3 T MR750 scanner. 3D fSPGR (0.5 9 0.5 9 1 mm, TE/TR/flip = 2.36/7.86 ms/120) images provided anatomic reference. Functional images (3.44 9 3.44 9 4 mm, TE/TR/flip = 35/ 2000 ms/900) were preprocessed (eddy current and motion correction), prior to co-registration using FSL [4] . Co-registered images were temporally (0-0.05 Hz), and spatially (5 mm FWHM Gaussian) filtered. Voxel-wise analysis was performed based off the BOLD time series phase space. Phase space separability into tap versus rest states was assessed with a linear discriminant classifier. Voxel-wise HRF correlation analysis served as an activation reference. Phase space metric and the correlation maps were Fischer transform z score normalized and spatially warped to the MNI-152 atlas for group analysis. Statistical maps were generated with AFNI [5] . Results: Bonferonni corrected T-tests showed significantly active regions for correlation and separation metrics for RT and LT trials. A larger activation using the separation metric, as noted for RT (Fig. 1) , also held for LT. ANOVAs comparing HRF correlation to boundary separation activity, with participants separated by handedness, were performed for RT and LT. The dominant tappers showed less similarity between the two metrics (Fig. 2) . Regardless of handedness, the metrics show a greater similarity during left hand tapping. The results indicate that the separability metric agrees with HRF correlation activation patterns yet is model free. It is sensitive to hand-dominance and thus could show activation as a gradient due to brain state organization and not, subject specific, amplitude. The results suggest that handedness affects motor activation, with left-handed patterns tending to show additional and bilateral recruitment. A feasible extension would compare paretic and non-paretic tapping activation for stroke patients to characterize activation and assess rehab induced changes. Introduction: Task functional MRI (fMRI) has been well-established combining the measurement of hemodynamic response function (HRF) with the general linear model (GLM) to detect blood-oxygenlevel-dependent (BOLD) changes in rat brain structures 1. However, it is unclear whether species (rat vs. mouse) or sex differences in mice exist regarding HRF. Further, a strong bias exists towards male study subjects in the literature 2. Therefore, we investigated HRF in male and female mice using 2 natural stimulation modalities (pinprick (pp) and von Frey (vF) ) and compared these findings to our previously published rat HRF1. Methods: 8 to 10-week-old C57BL/6 J mice were stimulated on their right hind paw and scanned using a 9.4 T MRI scanner with a Cryoprobe. After shimming, task fMRI measurements were performed using a single-shot gradient EPI sequence (TR/TE 1000/18 ms, Matrix 76 9 66, Resolution 200 9 200 lm 2 ) under medetomidine sedation combined with 0.2% isoflurane. Experiments started 40 min after bolus injection. For mechanical pp and vF stimulation, we used an inhouse developed rotating stimulator and a block paradigm with 10 s stimulation/20 s rest and a frequency of 1 Hz and a pulse duration of 0.5 s (Fig. 1) . SPM12 was used for MR data preprocessing and to perform GLM-based analysis (using the FIR basis set). Time courses of BOLD responses of S1HL were extracted with MarsBaR. Time courses were fitted ( Fig. 2A) , HRFs were calculated (Fig. 2B ) and compared with a functional t-test using custom written Matlab scripts as previously published 1. Results: We found significant differences in the HRFs between both sexes during pp stimulation (p = 0.029), which was accompanied by 0.5 s delayed onset in male mice (Fig. 3A) . Further, we detected a different HRF during pp and vF stimulation in female mice (p = 0.034) (Fig. 3B) . Again, the female pp-HRF showed a significantly faster onset. Next, we tested GLM performance with different model orders. Preliminary data show comparable performance of mouse-derived and rat-specific HRFs at the first order, while human HRF performed poorly. Using a 3rd order model all HRFs showed similar performance. In conclusion, our study shows sex-and stimulation modality-specific differences in HRFs. These differences might account for the increased sensory sensitivity of female compared to male mice. Further, the difference in modality might arise from increased recruitment of peripheral nociceptors when stimulating with pin prick. Since most preclinical studies are performed with male animals, we therefore recommend to use sex-specific HRFs or to use models of higher order to form regressors for the GLM. Introduction: To achieve sufficient spatial and temporal resolution to distinguish features of laminar BOLD responses, we recently implemented line-scanning fMRI 1 . Here, a single line is acquired with * 250 lm resolution and * 100 ms TR, requiring very precise positioning. To negate the convoluted folding of the cortex and maximize the resolution in laminar direction, we developed a framework for planning of the line based on functional (population receptive field [pRF] parameters) and structural (minimal curvature) information. Our planning strategy allows us to direct the line through a patch of cortex with distinct functional and structural properties, thereby greatly enhancing the utility of line-scanning data. Methods: The pipeline includes two different scan sessions at different days (Fig. 1) . In the first session, we acquired a high-resolution MP2RAGE and performed pRF-mapping at 7 T. The line-scanning acquisition in session 2 used a modified 2D gradient-echo sequence where the phase-encode gradients are removed and outer volume suppression (OVS) is used to suppress signal outside the line. Rigidbody transformation mapping of session 1 to session 2 was applied to the coordinate and normal vector of the selected best vertex to plan the line (Fig. 1) . Four 5-min line-scanning runs were acquired with flashing checkerboard stimuli presented on the left and right side of fixation in a randomized event-related design to elicit lateralized BOLD responses (Fig. 2BC) 2, 3 . The line was perfectly perpendicular to the cortical band covering our patch of interest ( Fig. 2A) , indicating successful navigation of the best vertex from surface to scanner. To determine which voxels of the line belonged to grey matter, we used the tissue segmentation obtained in session 1 (Fig. 3A) . Finite impulse response (FIR) time-courses showed a strong BOLD response in the left hemisphere following stimulation to the right, contralateral, visual hemifield (Fig. 3B ). This response was absent or inverted for ipsilateral stimulation, which caused negative BOLD-response in a few voxels per subject (Fig. 3C ). We have presented a method that combines functional and structural data to increase the accuracy of planning and enrich the information obtained with line-scanning fMRI. A hemifield experiment evoked a strong positive BOLD-effect in response to stimulation of the right visual field, while stimulation of the left visual field evoked more negative responses. We aim to build upon these promising results by optimizing the pipeline further in a variety of ways; eye tracking (effect of eye blinks on laminar processing), motion correction and denoising strategies, and an attention task to elicit stronger negative BOLD responses 4 . Purpose: The study is aimed at explanation of the mechanism of neurovascular coupling a pulse increase in the supply of oxygen and energy metabolites to the excitation area of central nervous system in response to neurostimulation. Methods and materials/background: Study group consisted of 15 healthy participants in age from 18 to 28 years (mean age 24 years). All MRI studies were performed on a Philips Achieva dStream 3.0 T scanner equipped with a 32-channel Philips dStream head coil. A 5 min fMRI echo planar imaging (EPI) sequence was acquired (TR = 3000 ms, echo time [TE] = 30 ms, 100 dynamics with dynamic scan time = 3 s). A flashing chessboard with a frequency of 8 Hz will be used as a visual stimuli for fMRI study (Fig. 1) . All fMRI studies for each participant were repeated twice: after the first fMRI scan all participants took a pill of acetylsalicylic acid (aspirin) to inhibit the prostaglandin synthesis-one of the main ways of vasodilation. An intergroup analysis (before/after pill) was estimated using paired t-test. Results: In all subjects, the visual cortex was activated statistically significantly (p \ 0.05, FWE). Rest of brain structures do not show a reliable response to neurostimulation ( Fig. 2A ).The study made it possible to discover and study effect of habituation-desentization of induced vasodilation upon repeated video stimulation. The resistance of BOLD signal level in visual cortex to repeated video stimulation and effect of acetylsalicylic acid (aspirin) was revealed. BOLD was found to be sensitive to aspirin in thalamus (p \ 0.09, FWE), a structure involved in the transmission of visual information (Fig. 2B) . Conclusion: The observed decrease in hemodynamic response to visual stimulation in thalamus under the action of aspirin, absence of the effect of aspirin in the visual cortex and a strong correlation between changes in BOLD in both structures indicate a different ratio of the activities of drug-initiated vasodilation and vasoconstriction processes. Analysis of the distribution of prostaglandin E synthetase in the human brain showed that the expression of the corresponding gene in occipital cortex is lower than in the thalamus. The lower content of the enzyme, the precursor of PGH synthetase, may be the reason for the lesser effect of its inhibition by aspirin in the cortex as compared to the thalamus. This creates a statistically significant vasoconstrictor effect of the drug in thalamus during video stimulation. Introduction: Recent methodological advances in fMRI contrast and readout strategies allow researchers to approach the mesoscopic regime of cortical layers. Specifically, high-resolution blood-volume sensitive vascular space occupancy (VASO) (Lu 2003) enables functional mapping of cortical information processing within and across brain systems (Bollmann 2020). While layer-fMRI VASO has commonly been performed at 7 T, the goal of this work is to investigate how well it can be extended to other field strengths, 3 T and 9.4 T. Methods: We used an SS-SI VASO implementation at 3 T, 7 T, and 9.4 T based on the 3D-EPI sequences by (Poser 2010) and (Stirnberg 2020) , respectively. All experiments were performed on Magne-tomSIEMENS scanners with 32-channel (3 T, 7 T) and 31ch (9.4 T) receive coils, using a 12-15 min finger tapping task. Layer analysis and within-layer smoothing (FWHM = 0.8 mm) was done in LayNii (Huber 2021). Here, data from two traveling heads were augmented with separate cohorts across field strengths (N = 12). 9.4 T experiments used a modified adiabatic TR-FOCI inversion pulse and toggling between two separate pTx shims for inversion and excitation (Huber 2018). Results: We find that a single run for BOLD and VASO can provide reliable task-induced signal changes at sub-millimeter resolutions across all tested field strengths (Fig. 1) . The GE-BOLD activation map shows more significantly activated voxels than VASO. VASO's higher spatial specificity is also visible in profile plots as two separate peaks (blue arrows). The blurry nature of the GE-BOLD signal is expected due to the locally unspecific signal of pial and diving veins. 3 T vs. 7 T: For VASO at 3 T, we find that the relatively low SNR can be partially accounted for by means of the divergence of blood and tissue T1 at lower field strengths. Furthermore, layer-specific smoothing, longer runs, and NORDIC-based (Vizioli 2021) component selection can result in additional sensitivity gains. 9.4 T vs. 7 T: At 9.4 T, we find that the practical constraints of B1 ? -challenges can be largely accounted for by advanced pulse specific pTx-shimming with 9.4 T-optimised excitation and inversion pulses. Discussion and conclusion: In this work we find that high quality layer-specific CBV-fMRI data can be acquired across a wide range of field strengths, as suggested previously (Markuerkiaga 2021). Each field strength has specific challenges that can be largely accounted for. We believe that the usability of sub-millimeter VASO sequences across field strengths will contribute to an increased accessibility of laminar-fMRI across a wider community. Bollmann & Barth 2020, PNB, Huber et al., 2021 , NI, Huber et al. 2018 . NI, Lu et al. 2003 , MRM, Markuerkiaga et al., 2021 , NM, Strinberg & Stöcker, 2020 , MRM, Vizioli et al., 2021 S4.P19. Total glutamate and glutamine concentration dynamic after short visual stimulus The peculiarity of the kinetics of changes in [Glu] , measured using MRS, in a short period is that the resulting change in [Glu] can be caused not only by slow metabolic mechanisms 3 but also by the rapid movement of the neurotransmitter from the vesicles 4 . Hence, it becomes necessary to determine the kinetic characteristics of Glu dynamics during neuroactivation. For this purpose, in this work, the total concentration of Glu and glutamine (Gln) was measured in response to a short visual stimulus in 24 s. Methods: MRI and single voxel 1H-MRS were performed using the Philips Achieva dStream 3 T MRI System (Eindhoven, Netherlands) using the 8 coil SENSE Head coil. Eighteen healthy subjects (11 males; mean age = 23.6 ± 2.2 years) took part in the study. The 4 Hz flashing checkerboard was used for stimulation in similar blocks (3 sstimulus, 21 s-black screen). FMRI images were acquired using gradient-echo echo-planar imaging (GE EPI) sequence with TR/ TE = 3000/40 ms, flip angle = 90°, in-plane resolution = 2.4 9 2.4 mm, slice thickness = 4 mm, number of slices = 30 and 120 measurements. An activation map was obtained using SPM12 in response to visual stimulation. The spectra were obtained using the PRESS sequence (TR = 2000 ms, TE = 35 ms, NSA = 444 (12*37), voxel size-20 9 30 9 20 mm) and were localized in the activated region of the visual cortex (Fig. 1) . Spectra were preprocessed. We identified the impact of BOLD effects in metabolite spectra by estimating the differences in total creatine (tCr) singlet (3.03 ppm) linewidth between each time point and the value at -1 s (1 s before the start presentation of the stimulus. Metabolites were quantified using LCModel with a basis set of 17 simulated metabolites. The difference between the obtained values (Glx/Cr, NAA/Cr, and linewidth of tCr) corresponding to individual time points was statistically determined by repeated measures of ANOVA statistical test. Post hoc Dunnett's test was then run. Results: Dunnett's post hoc test showed a statistically significant increase in Glx/Cr between -1 s and 1 s, 19 s (Fig. 2 , p \ 0.05, * ? 7% at each point; rmANOVA: p = 0.06). There is no correlation between the change in Glx/Cr concentration and the amplitude of the fMRI response (r = 0.24, P [ 0.05). Discussion: Analysis of the main rates of reactions for the synthesis of Glu and Gln showed that there is no reaction with the rate obtained in the present study (* 0.8 mM/s). It follows from this that the result obtained is most likely a consequence of the transition of Glu from the ''invisible'' (vesicular) pool to the ''visible'' one 5 . The peak at 19 s is considered as the end of the Glu reuptake process. S4.P20. The iron-complex of Deferasirox: the molecular and functional characteristics that can make it a viable alternative to Gd-based agents *E. Gianolio Cage Chemicals, Novara, IT Introduction: Currently, the search for alternatives to gadoliniumcontaining MRI contrast agents addresses the field of Fe(III) bearing species with the view that the use of an essential paramagnetic metal ion may avoid the issues raised by the exogenous Gd. In this context attention is specifically devoted to highly stable, hexacoordinated Fe(III) complexes [1] [2] [3] . Particularly interesting are Fe(III) complexes with ligands that are represented by clinically approved iron sequestering agents, as their established biocompatibility properties may provide a good support for facilitating their clinical translation as MRI CAs. Herein, the in vitro and in vivo studies on Fe(deferasirox)2 (Fe(DFX)2- Fig. 1 ), a Fe(III) complex with a sequestering ligand largely used for thalassemic patients [4] , are reported. Methods: T1 and T2 relaxivities were investigated in water and in human serum at 25°C and 37°C in the range of magnetic field strengths 0.01-80 MHz on a Stelar SpinMaster FFC-NMR relaxometer. The interaction of Fe(DFX)2 with human serum albumin was studied using the proton relaxation enhancement method and the involved binding sites on the protein have been identified by relaxometric competitive assays. The absence of any inner sphere water molecule coordinated to Fe(III) ion was assessed through a variabletemperature 17O-R2-NMR experiment at 14.1 T. Fe(DFX)2 efficiency as MRI contrast enhancer was investigated in vivo on a Bruker BioSpec 3 T scanner in a TS/a tumor bearing mouse model and compared with the contrast enhancement obtained by Gd(DTPA). Results: Fe(DFX)2 owns an outstanding thermodynamic stability and a good relaxivity, comparable to those of clinically used Gd-based CAs, even if lacking any inner sphere water molecule. Its high affinity for HSA enables, in serum, the formation of a supramolecular adduct with three units of complex bound to the protein. The binding sites were identified to belong to IB, IIA and IIIA subdomains. Preliminary in vivo imaging studies on a tumor mouse model indicate that, on a 3 T MRI scanner, the contrast ability of Fe(DFX)2 is well comparable to the one shown by the commercial Gd(DTPA) agent. Discussion: The relaxation enhancement capability, the very high stability, the overall biodistribution and excretion properties and the expected good biocompatibility of Fe(DFX)2, make this system a promising candidate as an alternative to the currently used Gd-based MRI CAs. Introduction: Dynamic contrast-enhanced (DCE) MRI asks for exact determination of the dynamic arterial blood concentration (arterial input function, AIF). Image derived AIF measurement is heavily biased by partial volume or motion effects which leads to insufficent quantification of perfusion MRI especially in small animals. In a previous study, we could introduce a novel method to measure the AIF DCE-MRI in mice in an extracorporeal circulation [1] . In the current study, we simultaneously recorded rapid MRI extracorporeal AIFs and radioactive contrast agent analogs in a hybrid cross-validation approach. Methods: 12 intracranial tumor bearing nude mice were measured in a 9.4 T Bruker Biospec small animal MRI. 35 mM Gd-DO3A-butrol (Gadovist, 7 mice) or Gd-DTPA (Magnevist, 5 mice) were co-injected i.v. in 100 ll mixed with their radioactive analog; 68 Ga-DO3Abutrol (mean 7.2 Mbq ± std 2.4) or 99m Tc-DTPA (23.1 MBq ± 6.2). An extracorporeal circulation was applied shunting the femoral artery to the tail vein. The circulation featured 2 reservoirs in the MR field of view and a MR compatible unit (Swisstrace, Twilite) for measurements of blood radioactivity (Fig. 1A) . A Golden-angle Radial Sparse Parallel (GRASP) sequence was used for DCE covering the whole brain and reservoirs in isotropic 0.2 mm voxels [2] . Compressed sensing MP2RAGE with identical spatial resolution was employed for T1 mapping [3] . Results/discussion: Integrated simultaneous recordings of PET and DCE-MRI AIFs using the Twilite measuring unit were technically feasible. Shunt flows varied across different animals but were constant in individual measurements as established by laser doppler flow recordings (Fig. 3A/B ). Dynamic acquisition employing rapid MRI techniques demonstrated very little noise at 5 s temporal resolution while allowing 3D isotropic whole brain coverage. Compressed sensing MP2RAGE established quantitatively plausible 3D T1 maps with identical coverage and spatial resolution in less than 6 min. Well quantitative correspondence on the whole range of dynamic contrast agent and radiotracer concentrations was established when applying a fixed correction factor of 1.8 (Figs. 1B and 2) . Although our preliminary analysis does not yet feature dispersion correction, the findings point towards a high precision across individual animals with both contrast agent and radioactive analog pairs. The robust quantitative relationship of simultaneous recordings of different contrast agent/radiotracer analog pairs establishes a basis for quantitative comparison of AIF recordings in hybrid small animal PET/MRI. Moreover, it enables to perform integrated PET/MRI modeling to explore quantitative relationships of contrast agent and radiotracer kinetics. Evaluation of the cerebrovascular reactivity (CVR) by use of breath-hold-triggered fMRI (bh-fMRI) as an index of the CPR has been proposed as a readily available and reliable alternative [1] . Recent findings suggest the use of resting-state fMRI (rs-fMRI) to estimate the CVR [2] . The aim of this study was to compare the rs-fMRI approach to bh-fMRI and [ 15 O]water PET in patients with MMA. Methods: A comparison of 25 rs-fMRI data sets of patients with MMA to the corresponding bh-fMRI data sets and, in a subgroup of 7 patients, to [ 15 O]water PET data sets was performed. The rs-fMRI images were realigned, normalized, spatially smoothed (12 mm FWHM) and frequency filtered (0.02-0.04 Hz). The cerebellar signal time-course was determined and a cross-correlation analysis was performed in which the correlation coefficients between the cerebellar Reference time-course and the individual voxels' signal time-courses were calculated. The data sets of all three modalities were segmented into 12 standardized VOIs [3] . A comparative analysis between rs-fMRI and bh-fMRI as well as [ 15 O]water PET was performed by calculating the correlation between the rs-fMRI CVR values (correlation coefficients) of the 12 VOIs and the corresponding bh-fMRI CVR values (relative signal change after breath-hold stimulation) and [ 15 O]water PET CPR values (relative CBF change after ACZ stimulation) [4] . Results: Both, the comparison of the 25 rs-fMRI data sets to the bh-fMRI data sets (Pearson's r = 0.71 ± 0.13, minimum = 0.35, maximum = 0.95) and the 7 rs-fMRI data sets to the corresponding [ 15 O]water PET data sets (Pearson's r = 0.80 ± 0.19, minimum = 0.41, maximum = 0.95) revealed high level of agreement [4] . Exemplary maps of one patient with high agreement and the corresponding scatterplots can be seen in Figs. 1 and 2. Discussion: The present analysis indicates that rs-fMRI might be a promising readily available method for hemodynamic evaluation with almost no patient cooperation required. Further studies are needed. Feasibility of glioblastoma tissue response mapping with physiological BOLD using precise oxygen and carbon dioxide modulation identified by applying an iterative analysis to the data to determine the correct voxel-wise temporal CO2/O2 shift to match the start of the hypercapnia/hypoxia/hyperoxia induced BOLD signal change. %BOLD signal change/mmHg during hypercapnic, hypoxic and hyperoxic stimulus was calculated for the whole brain, tumor lesion and segmented regions of interest (contrast-enhancing-CE-tumor, necrosis, edema). Results: Mean %BOLD signal change/mmHg during hypercapnic stimulus was 0.017 (SD 0.072) in whole brain, 0.028 (SD 0.069) in the tumor mask, 0.029 (SD 0.07) in CE tumor, 0.029 (SD 0.069) in necrosis and 0.047 (SD 0.052) in edema. Mean %BOLD signal change/mmHg during hypoxic stimulus was -0.014 (SD 0.006) in the whole brain, -0.020 (SD 0.017) in the tumor mask, -0.019 (0.015) in CE tumor, -0.029 (SD 0.025) in necrosis and -0.021 (0.021) in edema. Mean %BOLD signal change/mmHg during hyperoxic stimulus was 0.004 (SD 0.001) in the whole brain, 0.004 (SD 0.002) in tumor mask, 0.009 (SD 0.013) in CE tumor, 0.004 (SD 0.003) in necrosis, 0.003 (SD 0.002) in edema (Table 1) . Discussion: Our preliminary study, shows a good feasibility of BOLD with standardized and precise CO2 and O2 application as an emerging physiological imaging technique to detail specific GBM characteristics. In particular, specific tissue responses for hypercapnia, hypoxia and hyperoxia were found (Fig. 1) , whereas unique whole brain patterns could also be appreciated (Fig. 2) . This emerging BOLD technique poses two advantages, since standardized and precise gas application results in a high inter-and intrasubject agreement between follow-up studies, whereas the BOLD MRI contrast provides a high imaging contrast covering the entire brain. The unique tissue response patterns generated from this technique, can therefore be tested as novel imaging markers to better detail GBM lesions and gauge treatment response. Comparing common blood-brain barrier (BBB) permeability models for potential use in neuropsychiatric diseases *A. Gilbank 1, 2 , B. N. Frey 3 Introduction: Dynamic contrast enhanced MRI (DCE-MRI) as a tool for quantifying microvasculature permeability has been used to investigate many brain pathologies. In this technique, a series of T 1weighted images are acquired as a bolus of gadolinium-based contrast agent is injected to record its time-course through the tissue. A mathematical model is then used to fit the data and find physiological parameters, such as mass-transfer rate (K trans ) and plasma space. The brain's tissue environment has very low permeability due to the tightly-regulated nature of the blood-brain barrier (BBB). Certain pathologies, such as multiple sclerosis, have been found to strongly disrupt this barrier [1] . However, other disorders like bipolar disorder have been hypothesized [2] and found to cause more subtle disruption [3] , which is harder to detect and quantify. In this study, different mathematical models were compared in healthy brains to find if any model had higher precision with the goal of using the optimal model for our neuropsychiatric studies. Methods: Four models were compared: Tofts, modified Tofts, Two-Compartment Exchange, and the Uptake model. Eight healthy subjects (22 ± 4 years old, 2 male, 6 female) were scanned, 6 of whom returned for a repeat scan. Subjects were scanned with a 3Tesla MRI and 32 channel head coil (GE Healthcare, Milwaukee WI). A precontrast whole-brain(?) T 1 map was calculated from an acquired B 1 ? map and T 1 -weighted and proton density weighted images. Subsequently the DCE data was acquired. For DCE Gadovist (Bayer; 0.1 mmol/kg body weight) was injected with a 3 cc/s injection velocity into the left antecubital vein. Region-ofinterest (ROI) analysis was performed bilaterally on putamen, thalamus, hippocampus, caudate head and amygdala, and on midline structures: pons and cerebellum. The posterior pituitary was used as a ''maximal permeability region'', as its vasculature is void of a BBB. Data was fitted with each model in Matlab (MathWorks Inc., Natick MA) and the R 2 values were calculated. ANOVA and ANCOVA tests were applied to find significant differences between models (p \ 0.05). Results: The Tofts model was found to have a significantly lower R 2 in ROIs with a BBB, compared to the modified Tofts and Uptake models. Additionally, ROI size (voxel count) was found to be a significant covariate in the ANCOVA comparing the models' R 2 values. Discussion: As there is no ''gold standard'' for DCE-MRI modelling, measurement of model accuracy is not possible. Model precision, however, can be estimated through the error in model fit. The Tofts model may have performed worse due to its assumption of negligible tracer concentration in plasma space, or because it had the least parameters. Based on model error there was no difference between modified Tofts, Two-Compartment Exchange, and Uptake models in brain regions having an intact BBB. More data is needed to determine any differences between these three models. Introduction: Relative cerebral blood volume (rCBV) is a clinically well-known parameter and reflects information of both microvasculature density and diameter in brain tumors. Studies prove that rCBV can be used for predicting tumor aggressiveness in glioma (1) . However, in lower grade oligodendroglioma, due to ''chicken-wire'' nature of its vasculature, rCBV could also be elevated (2) . Vessel size imaging (VSI) is an emerging MRI technique enabling us to calculate the mean diameter of vessels within an image voxel (3) . In this study we aimed to investigate the added value of VSI in two nonenhancing glioma subtypes of oligodendrogliomas and astrocytoma. Subjects and Methods: A retrospective dataset consisting of 33 patients with confirmed non-enhancing glioma was used in this study (4) . All patients underwent 3 T MRI scanning (GE, Milwaukee, WI, USA) including hybrid EPI (HEPI); a 2D, simultaneous acquisition of GRE-and SE-EPI DSC MRI with acquisition parameters of 122 TRs, TR: 1500 m, 15 slices, voxel size: 1.88 9 1.88 9 4.00 mm 3 , TE GE/ SE: 18.6/69 ms. DSC MRI was performed with administration of 7.5 ml of gadolinium-based contrast agent with a pre-load bolus of equal size. rCBV maps were calculated using DSC data from the gradient-echo. Estimates of mean vessel diameter were made according to previously described methods (5) . The average of vessel size measurements and normalized rCBV within a tumor ROI (hyperintense on T2/FLAIR) were calculated for each group of patients. Results: The histological classification includes oligodendroglioma (grade II/III) and astrocytoma (grade II/III) ( Table 1 ). As seen in Fig. 1 , no significant differences between VSI and rCBV values between grade II and grade III oligodendroglioma and astrocytoma were found within this study. In the oligodendroglioma subgroup both average rCBV and VSI are higher in grade II compared to grade III (T Test, p = 0.21 and p = 0.08 respectively). Whereas, in astrocytoma subgroup both rCBV and VSI are higher in grade III compare to grade II (p = 0.51 and p = 0.41 respectively). See Fig. 2 . for visual inspection. Discussion: The results of this study suggests that rCBV and VSI are not able to predict the grade of oligodendrogliomas. Moreover, rCBV and VSI measurements in oligodendrogliomas grade II inclines to be even higher than astrocytoma grade III, while the clinical prognoses shows that the latter is more aggressive than the former. Note that the IDH mutation status of both groups (Table 1) is a likely reason for not finding any significant differences. Future investigations will include increasing the sample size as well as investigating correlations of molecular status with rCBV and VSI. Introduction: Diffusion MRI (dMRI) is a prominent MRI modality to provide contrast to the tissue structure at the microscopic scale, and many techniques have been developed so far [1] [2] [3] [4] . Among them, the SANDI approach 5 has been recently proposed to map the apparent soma size and density and neurite density based on dMRI data. SANDI employs a 3-compartment model and powder averaged diffusion measurements up to very high diffusion weightings, thus the estimates might show bias due to the Rician noise floor 6 . Here we investigate the effects of using either magnitude or real data on the SANDI parameters in the mouse brain, in-vivo. Methods: Acquisition: dMRI data from N = 6 C57BL/6 J mice were acquired on a 9.4 T Bruker scanner with a 4 channel cryprobe using a PGSE-EPI sequence (D/d = 20/5.5 ms) with 8 shells with b = {1, 2.5, 4, 5.5, 7, 8.5, 10, 12 .5} ms/lm 2 and 40 directions each, TE = 36.8 ms, TR = 4 s, 4 averages, slice thickness = 0.4 mm, 35 slices, in plane resolution = 0.12 9 0.12 mm, matrix = 118 9 100, Partial Fourier = 1.35, per-slice triggering and fat suppression, total time * 3 h. All experiments were performed after approval from the ethics committee following European Directive 2010/63. Data analysis: To compare the effects of using either magnitude or real data, pre-processing was performed as illustrated in Fig. 1 , with calculation of real data 7 after denoising 8 , ghost correction 9 and coil combination 10 . After pre-processing, the powder averaged data 11 was fit using the 3-compartment SANDI model (spheres, sticks and Gaussian diffusion) and a Random Forest regression algorithm in Matlab. Then, parameters estimated from magnitude and real data were compared. Results: Figure 2 shows representative SANDI parameter maps obtained from magnitude and real data. Overall, parameters have similar contrasts for magnitude and real data, although real data yields more homogeneous maps across the brain. The scatterplots presented in Fig. 3 for both white matter (WM) and gray matter (GM) ROIs show that magnitude data, which includes the Rician noise floor, results in slightly higher apparent neurite fraction (f neurite ). This is also observed in the maps in Fig. 2 , especially at the bottom of the brain where the SNR is lower, as the data was acquired with a surface coil. 1 University of Toulouse, Toulouse Neuro-Imaging Center (ToNIC), Toulouse, FR; 2 University of Toulouse, Toulouse, FR Introduction: Diffusion tensor imaging (DTI) is a classic approach to extract parameters related to tissue microstructure. Free water elimination DTI model aims to evaluate fraction of free water inside a voxel. For FW estimation, two shell acquisitions and fitting procedure with weighted linear and non-linear least squared (NLS) are recommended [1] . Another procedure has been used in clinical studies with regularized gradient descent (RGD), which was initially proposed to manage single-shell acquisitions. Our work aims to investigate differences between the two fitting procedures as well as the influence of the partial volume and the angular resolution to estimate accurately FW maps. Methods: We assessed a 3 T MRI exam (Achieva, Philips) on 30 healthy subjects (see parameters in Fig. 1 ). The gradient directions were calculated to be angularly equidistributed in single shell acquisitions and complementary between several sets. All DWIs were processed following: denoising with LPCA [2] filter using DIPY [3] , correcting for susceptibility and eddy current artefacts using SPM [4] and FSL [5] respectively. FW maps were calculated using RGD [6] and NLS [7] from DIPY. At high resolution, to evaluate the number of gradients needed for free water estimation, parametric maps were calculated with combinations of sets. Results: The Fig. 2 shows correlations between FW calculated with NLS and FW calculated with RGD in the corpus callosum (CC) and cerebrospinal fluid (CSF). The correlation is higher in the CC (R 2 = 0.9415) than in the CSF (R 2 = 0.719) and it is better for high resolution than the standard one (CC: R 2 = 0.5966; CSF: R 2 = 0.409). Other white and gray matter structures were investigated and results are the same as for the CC. Considering high resolution, we found that the estimation of FW did not differ using only one set or using all sets. Discussion/conclusion: Our results showed that at high resolution, only one set is enough to estimate FW with accuracy. It allows us to reduce acquisition time or to increase the resolution with the same duration as with more directions. We also demonstrate the fact that RGD results should be interpreted carefully especially for standard resolution. With high resolution the two fitting algorithms corelates more because of the reduction of partial volume effect. This result supports the interest of acquiring DWIs with the highest possible resolution. Fig. 1.b) . A remarkable concordance can be observed, validating the experimental results obtained with the birdcage coil prototype. The profile patterns also show an excellent agreement with those reported for a 7 T birdcage coil [4] . An analytical expression of the magnetic field for the popular birdcage coil was derived, and experimentally validated for high field applications. Introduction: The development of an RF coil with improved performance is still an important quest in MRI. We proposed to modify the standard birdcage coil by adding circular petals to the rungs to improve its magnetic field, B 1 and uniformity. Method: We numerically simulated the magnetic field of the coil prototype as shown in Fig. 1 , using 4 rungs with 4 petals on each rung. The separation between the circles is 3 times the petal radius, to diminish the unwanted effect of the mutual inductance. The commercial software CST Microwave Studio (CST MICROWAVE STUDIO, CST GmbH, Darmstadt, Germany) was used to calculate the electromagnetic fields. A saline-solution spherical phantom was used (r = 0.55 S/m, e = 78.4, q = 998 kg/m 3 , and l = 0.999991). B 1 simulations of a birdcage coil with similar dimensions and 8 rungs were also conducted for comparison purposes. Results and Discussion: Two-dimensional maps were numerically computed and shown in Fig. 2.a) and b) . To investigate the uniformity and sensibility of this resonators design, a comparison plots was computed with the simulation data, as shown in Fig. 2. c) . Profiles were calculated with data taken along the black line in Fig. 2.d) . Finally, a comparison histogram was also done and shown in Fig. 2 .e). A clear improvement of sensibility can be observed from Fig. 2 , considering that the comparison was done against a birdcage coil with 8 rungs. Numerical results show very concordance with data reported previously [1] . These results showed that it is feasible to build a volume resonator with a better sensibility and similar uniformity that a birdcage coil a larger number of rungs. Many low-field MRI systems employ open, biplanar magnets to combine the advantages of low field with enhanced access to the patient. The biplanar volume coil 1 follows the same philosophy, enabling simple positioning and comfortable patient scanning from its open access while maintaining good B1 homogeneity and filling factor. Such a geometry appears particularly handy for MSK imaging in various positions, as opposed to conventional closed volume coil geometries 2 (e.g. solenoids or Helmholtz). Here, we demonstrate an optimized coil composed of two biplanar coils operating in quadrature mode with good intrinsic decoupling at 4.33 MHz (0.1 T). The two coils A and B of the quadrature coil are based on 4 and 5-strip 150 9 100 mm Cu planes connected in series and tuned by variable capacitors at 4.33 MHz (Fig. 1A) . Inductive coupling was used to interface both coils 3 with an optimized coupler design. Good geometric decoupling between the two coils is achieved (S 21 -36 dB, S 11,A -5.4 dB, S 22,B -4.8 dB, Q A 222, Q B 267). MRI was performed on a small footprint resistive biplanar 0.1 T magnet (Bouhnik SAS, France). A noise equalization scaling was applied to the 2nd channel by considering the noise standard deviations in a corner region of the magnitude images. 3D sequences were run on a Cameleon3 spectrometer (RS2D, France) with a variable Gaussian density k-space sampling pattern of 50% for phase-encoding. For in vivo data, the k-space was filtered with a Tukey window (r = 0.2) using MATLAB (Mathworks, USA), and zero-filled to double the matrix size in 3D. GRE images on a silicone phantom (Fig. 2) show a homogeneous sensitivity of both coils at their center (ROI size: 30 9 30 9 12 voxels). The SNR of the combined image in the central ROI was on average * 36% greater than that of the separate channels. The distorted peripheral regions of the phantom, due to B0 inhomogeneity and gradient nonlinearities, are located far from the volume of interest. Figure 3 shows examples of 3D views obtained with the quadrature biplanar coil in vivo (ankle and elbow of two healthy volunteers). bSSFP images exhibit good contrast between bone, skin, muscles and ligaments. Typical bSSFP banding artifacts are visible at the edges of the FOV. A similar contrast is obtained from the first echo of the DESS sequence in the flexed elbow (Fig. 3) . Overall, the quadrature coil enables good sensitivity and enhanced SNR over its entire volume. This optimized design makes in vivo imaging at low frequency with high 3D resolution within acceptable acquisition times possible. In addition, different extremities can be easily scanned in various positions thanks to its three open sides. Complementing the advantages of low-field with open access detectors shows great potential for a variety of MRI applications, such as weight-bearing and intra-operative MRI without artifacts from surgical instruments and prosthetics. [1] . An object of interest is for example the myelin bilayer in the brain, which has signal components with T2s down to 8 us [2] . However, to perform such measurements high performance gradient systems and RF chains are required. One important part of such RF chain is the transmit-receive switch. Not only does such switch has to change state very fast, ideally in less than one microsecond, but they need to withstand high RF power as well. In all of this the induced transient signals in the RF chain have to be low enough not the saturate the LNA, which is challenging for such switch speeds and power levels. There were only a few TR switches fulfilling the base requirements [3] witch were based on PIN diodes. Recently a novel switch was introduced based on GaN MOSFET capable of switching in less than 100 ns and allowing more than 1 kW of peak RF power [4] . In this work we demonstrate that such switch introduces no image artifacts when challenging state of the art sequence is performed. Methods: Figure 1 shows TR switch topology, the GaN mosfet are turned on during transmission insulating the receive port. During reception the mosfets are turned off insulating the transmit port and connecting the TR port to the receive port. Figure 2 shows the implementation of such switch topology, additionally a LDO, mosfet driver and LNA are on the PCB. For testing the TR switch a 3D ZTE sequence with SPI gap filling was performed. Relevant scan parameter were 1 kW peak RF power, 1 MHz bandwidth with a gradient of 200 mT/m. The TR switch was changed state as soon as the coil was sufficiently rung down. Which was in this experiment was approximately 400 ns. For imaging, a proton free loop coil was used. As phantom, a cylindrical water phantom doped with MnCl2 to shorted T2 and NaCl to increase the coil loading and thereby shortening the ring down was used. Results: Figure 3 shows image of the water phantom. For better visibility, the image was besides linear also plotted with logarithmic magnitude. Discussion: As can be seen in Fig. 3 neither the foreground or the background have any artifacts that could be attributed to the transmit receive switch. This shows that the switch is able to support such short T2 measurements without issues. With a switch time of less than 100 ns this switch topology is not limiting short T2 imaging in the foreseeable future and could be one part to enable imaging of ever shorter T2 compounds. Introduction: Using parallel-transmit (pTx) 7 Tesla (T) promises to improve image quality and safety margins. Up to date, pTx is not used extensively due to its very restrictive safety margins (* 5 9 more restrictive compared to single-Tx for commercial coils). These margins are the main ones responsible for the limited use of acquisition parameters. In this work, robust validation of electromagnetic simulation with the experiment will be explored for 7 T pTx infant MRI. Methods: A realistic infant size in-house designed phantom was 3Dprinted and filled with a saline solution ([NaCl] = 5.8 g/L, volume = 4.8L, er = 79, r = 0.95S/m) (Fig. 1A) . The 8ch dipole array 1 (23 cm long, 1.5 cm width on FR-4) was tuned/matched at 297.2 MHz and its S-matrix was measured with VNA (Keysight E5080A-ENA, USA) on the bench using the phantom. In simulation, the 8ch dipole array was modeled (Sim4Life 6.2, ZMT, Switzerland), including the FR-4 substrate (er = 4, r = 0 S/m), and conductive parts defined as lossy metal (r = 5.8e7 S/m). The tuning/matching circuit layout was modeled with an exact design of copper trace. A similar range of values to realistic array components was used with losses for inductors (50-70 nH with series resistors (1-8Ohms)) and capacitors (18-25 pF for series/2-15 pF for parallel,with Q-factor 1500 2 ) to adjust the lumped elements for tuning/matching with the measured S-matrix as a target using a co-simulation method (Optenni Ltd,Finland). MR acquisitions were performed using the phantom on a 7 T MR scanner (MAGNETOM Terra, Siemens Healthcare, Erlangen, Germany). Individual B 1 ? -field maps (magnitude/phase) were acquired 3 , normalized to 1 kW total output power at RF amplifier, and compared to simulated B 1 ? -field maps. The B 1 ? -field distributions in CP mode and for one RF shimmed case were measured with the actual-flip-angle method 4 and compared to simulations. Results: The mean difference for the simulated and experimental reflection coefficients (Sii) is lower than 1%, while it is about 12% for the nearest neighbor couplings (Sij) (Fig. 2) . A good correlation is observed between experimental and simulated B 1 ? -field maps for magnitude and phase (Fig. 3B) . Dipoles 1 to 7 present an individual averaged B 1 ? -field value within ± 10% of the measured value (Fig. 3C ) while for dipole 8 it is 23% higher in simulations. The simulated CP and shimmed B 1 ? maps were correlated better than 90% with measured maps (Fig. 3D ). Discussion and conclusion: Robust correlations were found between experiments and simulations for S-matrices, individual B 1 ? -field amplitude/phase maps and RF shimmed B 1 ? -maps. The absence of coaxial cables in simulations or slight differences in coupling values, notably for dipole 8 with 1 may explain the remaining mismatches. We conclude that a robust validation of electromagnetic simulation with the experimental setup was shown for safe infant MRI at 7 Tesla. Introduction: Glioblastoma multiforme (GBM) is considered the most lethal of the malignant primary brain tumors 1 . In fact, even after aggressive treatment, prognosis and survival remain poor 2 . Recently, nanosystems have become promising candidates for GBM diagnosis and treatment, due to their exceptional magnetic properties, biocompatibility and blood brain barrier (BBB) penetrability 3 . In our previous investigations, metallated doped conjugated polymer nanoparticles (CPNs) (conjugated with fluorescent polymer F8BT) were visualized in tumors by T2 weighted (T2W) MRI 4 . In this sense, this project aims to evaluate the biodistribution of two CPNs with different types of cores, Fe 3 O 4 or NiFe 2 O 4 , in mice bearing GBM flank-tumors and control mice. Methods: In vitro validation of both CPNs was performed in a phantom study with different CPNs dilutions (Fig. 1A) . Afterwards, NOD-SCID mice were injected intravenously with CPNs with a Fe 3 O 4 or NiFe 2 O 4 core. CPN's biodistribution was studied by T2W magnetic resonance imaging (MRI) pharmacodynamics ( Fig. 1C & D) and T2 maps (Fig. 1B) were obtained before and after the CPNs injection in a 7 T system. An hour after injection, mice were sacrificed, organs were resected and studied by fluorescent imaging. Additionally, mice bearing C6-GBM flank tumors were studied by T2W MRI before and 15 min after intertumoral injection of Fe 3 O 4 or NiFe 2 O 4 CPNs. Then, mice were sacrificed and flanks removed for fluorescence studies using a IVIS Lumina II system. Results: We observed a higher CPNs uptake in the liver and a moderate accumulation in the renal cortex and the renal medulla as seen in T2W pharmacodynamics (Fig. 1C & D) . The CNPs liver accumulation seem to be higher with Fe 3 O 4 than NiFe 2 O 4 core nanoparticles (Fig. 1C) . Intratumor injection studies revealed that both CPNs can be visualized in flank tumors by T2W images, showing signal decreasing in the location where the CPNs were injected ( Fig. 2A) . Both CPNs were also detected in the xenograft tumors by fluorescence imaging and not observed in not injected (control) tumors (Fig. 2B) . Introduction: We previously 1 addressed the cause of apparently absent perfusion in the right inferior temporal lobe by optimising the spatial positioning of dielectric pads to achieve a more symmetric B 1 ? field. Our earlier work 2,3 has shown that optimisations to improve labelling efficiency such as tr-FOCI inversion pulses 4 and dielectric pads 5, 6 can yield robust perfusion imaging using ASL at 7 T. Dielectric pads provide a means to locally improve the B 1 ? field, and Pads of smaller size (13 9 13 cm 2 ) have proven to be a preferable compromise for participant comfort owing to differences in head shapes and sizes. We observed that optimised pad positioning allows a symmetric B 1 ? in near whole-brain acquisitions. Thereby the adiabatic condition for the inversion pulse is fulfilled, resulting in robust measurement of perfusion in the right inferior temporal lobes. Dielectric pads have been previously shown to yield improvements for spin-echo based diffusion imaging that is also B 1 ? sensitive 7 . Here, we extend our investigation to also include diffusion-weighted imaging using a PGSE sequence to assess the impact of our setup. Methods: Data were acquired on 6 participants (30 ± 3 years, 2 female) on a Siemens Magnetom 7 T using the 1Tx/32Rx NOVA head coil. Up to three 13 9 13cm 2 CaTiO 3 dielectric pads were employed using (a) 1-1 and (b) 2-1 conFiguration (Fig. 1,2 schematic). B 1 ? maps were acquired using an Sa2RAGE sequence 8 with 2 mm isotropic voxels. Perfusion-weighted (GRAPPA = 3, TIs = 700/1700 ms, TR/TE = 2760/12 ms, BW = 2164 Hz/px) and M0 (TR = 20 s) data were acquired with a pulsed ASL sequence with a FAIR QUIPSS II labelling and 2D EPI readout with 1.8 mm isotropic voxels. Diffusion-weighted data (GRAPPA = 3, b val (#dir) = 1000(35)/2000(53) s/mm 2 including 5 interleaved b0 images, TR/ TE = 6500/62 ms, BW = 1776 Hz/px) were acquired using a 2D-EPI spin-echo sequence with 1.5 mm isotropic voxels. Data processing was carried out using FSL''s oxford_asl, eddy_cuda9.1 and dtifit 9 . Results: Please see figure captions. Discussion: In this 7 T study, we show that whole brain B 1 ? distribution can be improved by optimising the placement of dielectric pads given a priori knowledge of the coil''s inherent B 1 ? asymmetry. We demonstrate that B1 ? has a direct impact on perfusion and diffusion measurements in the right temporal lobes and therefore, on their potential clinical utility at 7 T. We urge caution for translating 3 T protocols to 7 T because the skewed B 1 ? can be limiting on the measured perfusion and diffusion data. Next steps include to study the impact of the B1 ? asymmetry on regional perfusion and diffusion signals. We will use MESMERISED 10 to study the impact of B 1 ? optimisation on high b value acquisitions as well as more advanced diffusion models. Introduction: Deuterium magnetic resonance measurements following injection or ingestion of 2 H-labelled compounds are of increasing interest for metabolic studies [1, 2] . The quadrupole moment of 2 H offers sensitivity to molecular orientation [3] and different relaxation time behaviour compared to 1 H. Analysis of body fluid samples during heavy water loading is used for studies of protein synthesis [4] . Here we report 2 H MR measurements made at 7 T on human subjects loading with D 2 O to * 1.5% concentration over a 6-week period for a protein synthesis study. Methods: Measurements were made on a 7 T Philips Achieva scanner using a dual-tuned 2 H/ 1 H birdcage coil (Rapid). Two subjects were scanned weekly during the loading period. In each session, we acquired 3D multi-echo, gradient echo (MEGE) 2 H images with a range of TR-values to allow calculation of T 1 and T 2 * maps, using simple fitting to the saturation recovery and TE variation. Low resolution MEGE 1 H images were also acquired for comparison. In an additional scanning session, we used the 32-channel Nova receiver coil to acquire high-resolution (0.7 mm isotropic) 1 H MPRAGE and T 2 * -weighted images for tissue segmentation, along with a low resolution MEGE 1 H image for co-registration. Relaxation times were measured in CSF, grey matter and white matter regions defined on the segmented images. Results: Fig. 1 shows exemplar 3D 2 H GE image data (5 mm isotropic resolution, TR = 50 ms, FA = 30°, 4-echoes TE 1 = 6 ms, DTE = 8.5 ms, acquisition time 6.5 min) acquired after 8 days of D 2 O loading. The T 2 * contrast in the image, which is formed from the average of the 4 echoes, mainly differentiates CSF from the other tissues. Fig. 2 shows example 2 H R 1 = 1/T 1 and R 2 * = 1/T 2 * maps (6 9 6 9 10 mm 3 voxels) from two axial slices. These were formed by fitting to images with FA = 60°and TR = 80,120,240 and 480 ms and to 5 echoes with TE 1 = DTE = 9 ms, respectively. Corresponding 1 H R 2 * maps formed by fitting to MEGE data with 5 echoes (TE1 = 8.9 ms and DTE = 5 ms) are also shown. Fig. 3 Introduction: Electrical muscle stimulation (EMS) has been successfully used during MRI imaging 1-3 , but reports on the safety aspects are scarce 4, 5 . Quantitative data on heating in realistic usage situations help to ensure the safety of the subjects and optimize the acquisition parameters and hardware setup for high-SAR sequences. Methods: Temperature was measured with a fiber optical thermometer. The two electrodes of the EMS device were in contact with a muscle tissue-simulating gel for MRI and temperature sensors were placed below them. A 3 rd sensor was inside the medium far from the electrodes and a 4th sensor was inside the bore in the air. An overview of the experiments is presented on Table 1 . The experiment was repeated with the electrodes attached to a large plastic bottle 4 , and in vivo. The objects were placed off-center to achieve an E-field coupling similar to the one experienced on a limb. For the in vivo measurements, two large electrodes were connected to the EMS device and placed on the thigh 2 and two on the other leg, in skin contact and under a non-connected electrode. At 3 T, a balanced SSFP sequence (scan time: 15.6 min, TR = 3.74 ms, a = 27-31°) was used to maximize SAR. At 1.5 T we measured once at maximum SAR and once at half the maximum SAR. In all cases, the calculated B1 ? rms field was recorded. Results: The rise of temperature never reached a safety-threatening level. A rise of temperature was observed on the sensors that were in touch with the electrodes and was different for the two electrodes ( Table 2 ). The highest temperature rise at 3 T was 9.6°C, when smaller electrodes were used. In vivo, the temperature rise did not exceed 4°C. For larger electrodes, maximum temperature rise was 6.5°C. At 1.5 T a higher B1 ? rms was allowed, therefore the resulting temperature rise was higher and for the maximum allowed SAR this was 12°C. Temperature curves (Fig. 1) show that placing the electrodes on top of a bottle lead to an overestimation of the heating rate (1A&B). The rise of temperature in vivo at 3 T was roughly linear and did not exceed 4°C (1C), and was very similar to the free probe. Finally, at 1.5 T we observed a fast and large change of temperature (1D). The largest temperature rise was observed at 1.5 T, with higher B1 ? rms, and at 3 T with smaller electrodes connected to the cable. The highest temperature rise at 3 T was less than 10°C. Therefore, it is suggested that during EMS-synchronized MRI, the electrodes can stay attached to the subject during the scan even in case of high SAR sequences. For an additional safety margin, the cable can be detached from the electrodes when possible. Using electrical muscle stimulation in vivo during MRI scans causes no major risk of excessive RF heating, neither at 1.5 T nor at 3 T. S5.P11. Estimation of RF response at multi-electrode lead hotspots using lead electromagnetic models Introduction: This work focus on identifying the optimal clinical relevant pathway from a given set of trajectories and estimating the RF-induced dissipated power at the multi-electrode array of cortical implant (CorTec GmbH, Germany, see Fig. 1 ) by means of lead electromagnetic models (LEM). To develop the LEM, the transfer functions for each of the electrode are computed along with the tangential electric fields. The unscaled LEM is then used to compute the dissipated power at the electrodes and identify the optimal trajectory. Methods: The lead electromagnetic models as described in ISO/TS 109741 [1] comprise of a transfer function of the implant, the tangential electric field along the lead pathway and a constant. The LEM is given by Here, S(l) is the transfer function, E tan (l) is the tangential electric field along the lead pathway, A is the calibration factor, and l is the length of the lead. The transfer functions of all four leads of CorTec's implant are computed by reciprocity method [2] using ANSYS HFSS (ANSYS Electromagnetics 2020 R1) at resonance frequency of 64 MHz with leads being inserted in gel medium with physical properties defined as per ASTM standard. Further, an MRI birdcage coil at 64 MHz is simulated with a human head model along with 11 clinical relevant trajectories (Fig. 2) . The tangential electric fields along these trajectories are computed and used in Eq. 1 to develop the unscaled LEM. Results and discussion: The unscaled lead electromagnetic model (Eq. 1 without constant A) is then used to compute the power dissipation at the hotspots for all the clinical trajectories in the human head model. Fig. 3 shows the normalized power dissipation for all 11 clinical trajectories computed using unscaled LEM. It can be observed from Fig. 3 10.3.7 and specifies that the measurement must take place at constant around the body temperature (37°C ± 5°C), if the material properties are significantly affected by ambient temperature [1] . However, it has been identified that especially implantable pulse generators (IPG) require an advanced test setup as these are more likely to fail. It is important to note that the measurement setup is permanently exposed to the static magnetic field. In order to fulfil these requirements we have developed a temperature controlled measurement setup to be able to test these critical implants within the MR environment. Setup: The setup shown in Fig. 1 is built to utilize the dB/dt at the top of the MR bore. The setup contains six controllable chambers, and each chamber was equipped a heating resistor (HR) to imitate the behaviour of an IPG. Cooling of the HR is achieved by a continuous flow of tempered air. Fibre optical probes are attached to each HR to monitor the temperature. Two more probes are pinned outside of the chamber to monitor Reference temperature. The Reference temperature was monitored all the time and used to control the air flow for an automatically readjusted of the temperature inside the system. Focus of the whole test setup is the MR compatibility which is achieved by the usage of only MR Safe labelled materials. The green, red, and purple graphs in Fig. 2 , shows the heating of 5.5 W, 3.3 W and 1.1 W at the heating resistor. The first dip in the graph shows the time point where the heating was switched off and the second dip shows where the active cooling was switched on. The yellow graph shows the temperature curve with a maximum cooling for three different heating powers. This was used to determine the maximum power that can be cooled to reach the required temperature range. At a power of 3.3 W the temperature can be cooled down to a value of 38°C which is within the required range. The brown graph shows the constant temperature curve at approximately 37°C using the cooling system. Discussion: This setup shows that it is possible to meet the standard prescribed in the ISO/TS 10974. This approach provides a more realistic way of performing gradient induced vibration measurement, which leads to more reliable measurement results. As a conclusion it is no longer necessary to distinguish between devices that have to be measured at 20°C and those that have to be measured at 37°C, as all AIMD can be measured at body temperature in a simple manner. With the presented setup an advanced temperature control is realised for the vibration measurement of critical implants. Fig. 1 , the neural network in the decoder part has been replaced by an analytic equation solver. This uses an analytic expression derived for the magnetic field generated by a cubic permanent magnet [2] . We call the combination an encoder-analytic hybrid model (EA model). The encoder part is a 3D convolution neural network that takes a 3D target field as an input and generates the magnet design parameters as an output in the form of a 1D vector. This EA model thus does not require a dedicated training data set. It can simply learn from the analytic part of the model, which acts also as a form of regularization. Results/discussion: We put our model in a task to generate the best surface magnet design. This should have a small perfectly linear gradient (G z ) perpendicular to the magnet surface which can be used for spatial encoding perpendicular to the surface of the magnet and at the same time the field should be perfectly homogeneous in the x-y plane in a region of interest (ROI) defined above the magnet. The Figure of merit that is minimized here is thus inhomogeneity defined over a reduced magnetic field (B r = B-zG z ). We start with a 6 9 6 grid of permanent cubic magnets and the optimization task looks to find the best z-positions for each magnet within a ± 5 mm translation (see Fig. 2 ). Our AI model would look for the solution by iteratively learning from the forward map of the problem (the analytic part) and fine-tuning the parameters for subsequently building the inverse map (the encoder part). For the current case, these inhomogeneity values are taken as the error which is then propagated backwards to train the encoder. The ROI of dimensions 16 9 16 9 6 mm 3 was defined at around 30 mm above the zero position of the magnets. Figure 3 (a) shows the result of the inhomogeneity convergence obtained using the EA model. Figure 3 (b,c,d) provides for the best magnet design, the line plots for the B x , B y and B z components in the ROI along the x-, y-and z-directions. We compared this result with a genetic algorithm model and found that the best magnet output of our EA model is about 3-4 times better. S5.P14. A fast 0.5 T prepolarizer module for preclinical magnetic resonance imaging Introduction: We present a description and the experimental test of a home-made PMRI (Prepolarized Magnetic Resonance Imaging), preclinical system designed for simultaneous visualization of hard and soft biological tissues [1] for the HISTO-MRI project [2] . Subjects/methods: The PMRI system is shown in Fig. 1 . This setup is composed of a main, C-shaped permanent magnet, a prepolarization magnet (PM), an RF coil, a cooling system for the PM, and highpower electronics. The PM is a water-cooled solenoid capable of generating a dc pulses with an intensity, Bp, of 0.48 T for a drive current, IP, 255 A. To generate the prepolarization pulses, the system uses high-power electronics based on a high-power switching module (IPM-16P from Eagle Harbor Technologies Inc.). This is driven from a battery bank consisting of 16 commercial batteries connected in pairs, each pair with both batteries connected in series. A more detailed description of this system can be found in [3] . Discussion/results: Two different studies were carried out. Firstly, we demonstrated the SNR enhancement directly in the reconstructed images using the PMRI system. For this we used a mouse brain sample. The images in Fig. 2(a) and (c) are obtained using a RARE pulse sequence which a prepolarization pulse of 700 ms at ?p & 0.48 T, applied at the beginning of a sequence repetition, and where ?dead = 40 ms. The color scale is common to both datasets to highlight the gain in SNR when we prepolarize the sample. The SNR enhancement is evident in both sets of images. To quantify the influence of prepolarization, we plot in Fig. 2 (d) the signal intensity profile along a horizontal line around the middle portion of the images in Fig. 2(a) . The results shown are low pass filtered (in image space) to aid visual estimation of the SNR increase & 1.72. Secondly, we demonstrated the capacity of our system to obtain images with tissue contrast. In this case, we used a sample with different organic samples, (with different T1, see Fig. 3(b) ). The images reconstructed in Fig. 3 (a) are taken for a RARE PMRI sequence with different prepolarization times. The evolution of CNRs as a function of the prepolarization pulse length is also shown in Fig. 3 (c). Unconventional MRI platforms, including low-field scanners, have seen rapid advancement toward clinical and research applications in the past few years [1] . The creation of open-source platforms has lowered the entry barrier for education, clinical and advanced research [2] . In this work we show the status of MaRCoS (MAgnetic Resonance COntrol System), an open-source MR console and software package with powerful features, that offers higher versatility than many closed-source commercial spectrometers, and show its use in several MR setups. Figure 1 shows the MaRCoS stack. It is based on the previouslydeveloped OCRA [3] , upgraded to improve its features. The core hardware is a Red Pitaya SDRLab ( [4] , $600) controlling a custom gradient driver board (either [5] or [6] , * $300). The SDRLab FPGA (Field Programmable Gate Array) runs the flocra firmware. Compared to OCRA, flocra uses a 'FLow-based' structure with two RF transmit and two receive channels with independent frequencies, phases, amplitudes, downmixing parameters and sampling data rates. It also provides four gradient outputs and six TTL I/Os, with a timing resolution of 8.13 ns. The shortest time between RF, TTL or receiver events is 33 ns, and * 1 us for gradients. Sequence length, timing and complexity have fewer restrictions than before, e.g. there are no raster clocks, and sequences with hundreds of coherent TRs, soft pulses, and arbitrary gradient waveforms have been tested. Changes to a parameter (RF properties, receiver rate, etc.) can be precisely timed. As with OCRA, this facility allows, for example, real-time alteration of sequence parameters based on inputs from external sensors. The SDRLab is controlled using Python libraries. Sequences can be written using i) Pulseq [7] , a hardware-agnostic pulse sequence prototyping framework, ii) a simple domain-specific language, or iii) sets of time/value Numpy arrays. A hardware simulator and a test suite greatly simplify new feature development. Figure 2 shows a GUI that runs standard sequences, and an example TSE waveform. Figure 3 shows images acquired at three sites using MaRCoS, demonstrating the capabilities of the system for diverse and high-quality imaging. However, MR investigations of the heart are time-consuming and, thus, not feasible at all institutions. Additionally, the presence of active cardiac devices (such as pacemakers and implanted cardiac defibrillators), poses an additional safety risk in MRI, because of the potential of heating and device malfunction. For these reasons, cardiac MRI at lower field strengths has recently gained interest [1, 2] . Lower field strength scanners are potentially cheaper, and pose reduced safety concerns. The aim of this work was to assess feasibility of cardiac imaging on a novel commercial reduced-cost lowfield scanner, with ultra-wide bore size and reduced gradient performance in presence of active cardiac devices. In contrast to existing dedicated research systems, this is a commercial device that can have immediate application in the clinical routine; however, cardiac scanning on this scanner is currently not supported by the manufacturer. Methods: A prospectively triggered balanced steady-state free precession protocol was implemented on two commercial whole-body MRI systems (MAGNETOM Free.Max, 0.55 T, and MAGNETOM AvantoFit, 1.5 T, Siemens, Germany) with a resolution of 1.7 9 1.7 9 6.0 mm 3 , FOV 320 9 290 mm 2 , flip angle 80°, and parallel imaging factor 2. TR/TE was 4.26/1.78 at 0.55 T and 2.7/ 1.2 ms at 1.5 T. A healthy volunteer was scanned at both field strengths in three conditions: with no cardiac device, with an MR-Conditional Pacemaker (PM, Identity 5386, St Jude Medical, CA), and with an MR-Unsafe Implanted Cardiac Defribillator (ICD, Lumax 540, Biotronik, Germany). The devices were placed on top of the right pectoral muscle to simulate a clinically relevant position. Two-and four-chamber views (CV) of the heart were obtained. Results: In the no-device condition, the images at 0.55 T exhibited a lower signal-to-noise with respect to the 1.5 T images. The contrast between blood and myocardium was visible in both the 2-and 4-CV planes ( Fig. 1 ), but not the trabeculations. In the presence of implants, the PM did not impair the evaluation of the left ventricle, whereas the ICD caused major artifacts, especially visible in the 4CV at both field strengths (Fig. 2 ). Discussion: Despite the theoretical advantage of a lower field strength with respect to susceptibility, the larger bore size of the 0.55 T scanner (80 cm vs 60 cm) and the increase in TR due to reduced gradient performance made the size of the artifacts comparable at both field strengths. The 0.55 T system suffered from highly reduced signal-to-noise, however the delineation of the main structure was possible in normal conditions. Alternative k-space sampling strategies able to exploit the gradients more efficiently while keeping TR short (e.g. radial) are necessary for improved diagnostic performance. Introduction: Processing magnetic resonance spectroscopy (MRS) signals remains challenging due to hardware and physiologic processes, which may lead to frequency and phase shifts (FPS). Thus, frequency-and-phase correction (FPC) is a useful step in MRS signal processing. Deep learning (DL) has proved to be successful in a wide range of tasks, including the MR field. DL applications in MRS have recently emerged 1 . It has been shown that DL can also be used for FPC 2 in the frequency domain with two separated networks. In this study, we proposed a novel deep autoencoder (DAE) network for FPC. We showed that a single DAE network could learn a nonlinear low-dimensional model to predict frequency and phase shifts. Methods: Single voxel spectroscopy in vivo 1H MRS signal was acquired from a rat's right hippocampus in a 9.4 T small animal MR system (Bruker BioSpec) using point resolved spectroscopy (PRESS) sequence (spectral width = 4400 Hz, 4096 points, TE = 16.5 ms, TR = 2500 ms, 256 averages). The proposed DAE ( Fig. 1 ) was implemented in the Pytorch framework. Training and testing set (X) with 10,000 samples were created by applying uniformly distributed artificial frequency and phase offsets in the range of -20 to 20 Hz and -90°to 90°, respectively, to the acquired scan. The training was performed using 9000 samples, the mean-squared error loss, and an Adam optimizer 3 with a batch size of 32, learning rate 1e-5, and 150 epochs. In addition, the test dataset is processed (or corrected) using the spectral registration (SR) 4 method. Results: Figure 2 shows the error (True-DL-Predicted) against the true offset. In the test dataset, the mean error of the frequency and phase shift correction achieved with our proposed method was 0.002 ± 0.042 Hz and-0.009°± 0.254°, respectively, while SR correction errors were 0.048 ± 5.5 Hz and-0.471°± 5.973°. However, it is disputed whether denoising can decrease estimation uncertainties or if it is just a cosmetic tool reducing noise in signalfree areas. We propose a deep learning (DL) approach for noise removal in the time-frequency domain where noise and signal is distributed in a 2D fashion before reconversion to pure 1D spectral or time domain. Here, we judge denoising performance in the relevant spectral regions with a novel score. Methods: Human brain spectra were synthetized in VESPA (TE = 35 ms sLASER). Datasets consist of 16 metabolites with concentrations varying from 0 to twice the typical concentrations in healthy brain. For quantification purposes a constant water Reference is added at 0.5 ppm. To mimic in vivo conditions, uniformly varying parameters are used: shim 2-5 Hz, water SNR 5-40, macro-molecular background by ± 33%. Time-domain signals (4096 points, 1 s) were transformed into spectrograms using a short-time Fourier transform with window size 128 and overlap interval of 97 samples yielding a 128 9 128 matrix. Figure 1 illustrates the DL model, a U-net with symmetric skip connections (adapted from 2D audio signal processing1), that used real and imaginary channels and mapped spectrograms into apparently noise-free representations. The effect of denoising (DQ, Denoising Quality) is evaluated as a weighted and non-weighted mean absolute deviation from ground truth (GT) relative to the mean absolute deviation of pure noise. Without weighting, DQ is defined like fit quality2. Weighting with the GT spectrum makes this measure (DQw) sensitive to the spectral areas crucial for quantification. (DQ and DQw were calculated according to the equation presented in Fig. 3 ). Results and discussion: A novel MRS denoising technique in timefrequency domain was implemented and performed very well as judged visually in Fig. 2 . Results of denoising as judged by the traditional and a novel denoising score are presented in Fig. 3 for a test set of 1 k datasets. While DQ is close to 0 for a wide range of SNR, DQw is close to one on average, suggesting that denoising is mostly effective in non-signal containing areas. For some spectra, it is \ 1, for others [ 1. Conclusions: We show that time-frequency domain denoising by DL is well suited for short TE MR spectra to yield visually pleasing spectra. However, the mean residuals in the spectral areas of the metabolite signals (reflected in DQw) seem to be almost as large as without denoising. Further investigations including model fitting of denoised spectra and investigations of the variance of such results are now needed for final judgment of the practical benefit of denoising in MRS. Introduction: Quantification in spectroscopy is traditionally based on model fitting. Prior knowledge, fitting boundaries and choice of algorithms represent crucial steps that introduce dependencies on users and/or fit packages. Deep Learning (DL) has introduced the possibility to speed up quantitation. However, questions arise in regards to how to access uncertainties. Here, a preliminary comparison of accuracy metrics between the two methods is explored using simulated datasets with known ground truth (GT). Methods: Datasets with 16 metabolites were simulated. Concentrations vary independently between 0 and twice a normal reference concentration from literature. A constant downscaled water reference is added at 0.5 ppm to ease quantitation. Macro-molecular background (MMBG), shim and SNR mimic in vivo acquisitions and were independently and uniformly varied (time domain water referenced SNR 5-40, shim 2-5 Hz, MMBG amplitude 1-2), Fig. 1 . A shallow Convolutional Neural Network (CNN) 1 uses spectrograms 2 calculated via an in-house script as input, Fig. 1 . Relative concentrations are provided as output. Referencing to an estimated water content, absolute concentrations can be evaluated. A simulated spectrum of a typical median acquisition scenario (SNR, shim) was taken as reference and fitted. Cramer Rao Lower Bounds (CRLBs) were used as uncertainty reference for the comparison. • CNN predicts well-represented singlets and characteristic metabolites of human brain spectra well whereas it is less accurate on predicting low-SNR coupled metabolites. However, CNN accuracy resembles for most cases CRLBs of a median spectrum, Fig. 2 . • Low concentrations tend to be overestimated, high concentrations underestimated. • CNN uncertainties scale proportionally with 1/SNR, i.e. noise. • Optimized spectrogram representations are found to be well suited for spectral quantification. Comparison with otherwise used onedimensional CNNs 3 is needed. • DL has recently shown comparable uncertainties to traditional fitting 4 , but is faster and user-friendlier. • Ideally, CNN or traditional estimations should be unbiased, thus return uniform distributions for uniform test data. However, our CNN predictions tend towards the mean of the training data, in particular for weakly represented metabolites. Intrinsic uncertainties can explain this phenomenon: predictions at the boundaries of the training range are folded back towards the mean value in case of strong uncertainty, given lack of knowledge outside the boundary. Stronger regularization, longer training and deeper network architecture do not reduce this effect substantially. • Active learning with training data stressing difficult cases may improve predictions. Conclusion and discussion: This study proposed a 1D-CNN model for grading meningiomas based on 1H-MRS. Our results indicated that 1H-MRS provides relevant features to the proposed deep learning model for meningioma grading. Future studies will be conducted to verify our results in a larger patient cohort. 1 Institute of Myology, Neuromuscular Investigation Center, Paris, FR Introduction: In skeletal muscle, the multi-exponential behavior of T2 relaxation seems to reflect histological compartmentation [1] . Nonetheless, in order to assess compartments' intrinsic properties from the 1D T2 spectra, knowledge of the compartmental exchange rates is needed. 2D relaxation-exchange methods could provide such information [2] . Inversion of 2D relaxation data into 2D spectra is an ill-posed problem. Although stable solutions can be obtained via regularized Inverse Laplace Transform (rILT) [3] , these are penalized by relatively low spectral resolution. In this work we propose a non-linear least squares (NLLS) approach and confront it to the conventional rILT method on simulated data. Methods: Analytical solutions of the Bloch-McConnell equations were used in order to simulate signals from a CPMG-storage-CPMG sequence ( Fig. 1 ) for 4 bi-compartment systems, characterized with distinct compartment sizes (m 0 a = 90 or 80%), residence times ( s a = 750 or 500 ms; equilibrium condition: m 0 a /s a = m 0 b /s b ) and fixed intrinsic T1 and T2 values (T 1a /T 1b = 1250/1650, T 2a /T 2b = 30/ 190 ms). The sequence (Fig. 1 ) was parametrized as follows: IES = 1 ms, n 2 = 100, n 1 = 1…100 ms in 16 logarithmically spaced steps, and TS = 10…500 ms in 15 logarithmically spaced steps. This resulted in a 16 9 100 2D relaxation data per TS value. Five replicates of each 15 9 16 9 100 data set were built by adding complex white noise (SNR = 104) and taking the absolute value. T2-T2 spectra were obtained from rILT [4] of each 2D relaxation data, and the relative fractions of each peak were calculated. Each 2D relaxation data was also fitted, via NLLS, to the following model: S n 1 ; n 2 ð Þ ¼ P 11 exp À n 1 þ n 2 ð ÞÂIES=T 21 ð Þ þ P 22 exp À n 1 þ n 2 ð ÞÂIES/T 22 ð Þ þ P 12  ðexpðÀ n 1 =T 21 þ n 2 =T 22 Þ Â IES ð Þ þ exp À n 1 =T 22 þ n 2 =T 21 Þ Â IES ð Þ ð Þ 1 ð Þ Where P 11 , P 22 , P 12 , T 21 , T 22 are the adjustable variables; P 11 and P 22 are the fractions of magnetization that did not exchange during TS (diagonal peaks in the 2D spectra), P 12 represents the fraction that did exchange (identic off-diagonal peaks in the 2D spectra), and T 21 and T 22 are the apparent T2 values. Finally the intrinsic parameters, m 0 a , T 2a , T 2b and s a were estimated by fitting the experimental curves for the peaks'' areas as functions of TS, obtained from both methods (rILT and NLLS), to a theoretical model derived from the analytical solutions of the Bloch-McConnell equations (see Fig. 2 ). The results obtained with the NLLS method showed much higher accuracy and precision then the conventional rILT method (see Table 1 ). Discussion: When one can reasonably assume the number of compartments in the compartmental model, a non-linear fit of 2D relaxation data to the proposed model (Eq. 1) provides more reliable results than the conventional rILT method. Hippocampus was segmented in T2-w images by automatic registration of a rat brain atlas as described in [2] and translated to T1-w images by affine registration. T1-w images were preprocessed (isotropic voxel interpolation, intensity normalization and discretization) and 107 intensity, shape, and texture radiomics features of the hippocampus were extracted by pyradiomics [3] . To reduce dimensionality, features with extremely high or low variance and highly correlated features were discarded, and the most relevant ones for classification were selected. Machine learning was applied to the selected features to classify into 4 groups: adult (\ 1 year-old) WT, adult Tg, aged ([ 1 year-old) WT and aged Tg, to differentiate early pathological from healthy brain and from advanced pathology. The dataset was split into training-validation (142 images) and test (52 images) sets. Machine learning was trained on the training-validation set with a k-fold (k = 10) cross-validation scheme and different models to find the most accurate. The accuracy of the selected model was assessed in the test set. Results: Six features resulted from the selection step (Table 1 ) and linear Discriminant Analysis was the selected model to classify. Figure 1 shows the confusion matrix in the test dataset. The classification between Tg and control animals was 69% accurate. It achieved 78% if only aged rats were considered, but it was lower (64%) in the no-aged animals. In the images correctly identified as Tg, the accuracy to classify between aged and adult Tg rats was 75%. Discussion: Radiomics has shown promising results in AD patients [4] but to the best of our knowledge, this is the first time it is applied to an animal model of AD. Although Tg aged rats, where frank pathology is expected, were more accurately identified, promising results were also obtained in early ages. Age was considered as indicator of disease stage, but more accurate individual assessment could result in better classification. Introduction: Meningiomas are the most frequent tumors of the central nervous system that account for 38% of primary brain tumors. Meningiomas are classified into benign (Grade I), atypical (Grade II), and malignant (Grade III) meningiomas based on their histological features [1] . Grade I, or low-grade meningiomas (LGMs) are associated with a better prognosis than Grade II/III, or high-grade meningiomas (HGMs), which have invasive characteristics and rapid progression [2] . Radiomics extract a large number of quantitative and minable features from medical images, and have potential for noninvasive meningioma grading. The aim of this study is to investigate the potential of radiomics features calculated from contrast-enhanced (CE) T1-weighted (T1-w) for grade prediction in meningiomas. Methods: Thirty-three low-grade (Grade I) and 43 high-grade (Grade II/III) meningioma patients (mean (± std) age = 50.36 ± 12.9 years, age range = [27, 80] , male/female: 30/46) were included in this study. Post-contrast (gadolinium) T1-w MRI data (TR/TE = 415/10 ms) were acquired using a 3 T clinical MR scanner (Siemens Healthcare, Erlangen, Germany). The workflow of the study is given in Fig. 1 . First, the CE tumor volumes were manually segmented using 3D Slicer. To minimize the effect of inhomogeneity between acquisitions, images were z-score normalized and resampled to the same resolution (1 9 1 9 1 mm). Then, a full set of 1132 features were extracted using PyRadiomics [3] . Feature selection followed by collinear feature elimination was conducted using Gini index and Pearson''s correlations (r [ ± 0.9). The relationships between radiomics features and meningioma grades were assessed using Mann-Whitney U test. Machine learning algorithms were applied to classify LGMs and HGMs using tenfold cross-validation with 50 repetitions. The feature elimination process resulted in nine predictive features for differentiating LGMs from HGMs (P \ 0.005 for all) (Fig. 2) . The best performance for grading meningiomas was obtained with a gradient boost classifier (loss function: deviance, learning rate: 0.1) yielding an accuracy of 0.76 ± 0.043 (0.82 ± 0.06 sensitivity and 0.69 ± 0.07 specificity). The results of this study indicated that conventional CE T1-w texture radiomics could predict meningioma grade. The heterogeneous signal enhancement of CE T1-w MRI indicates tumor heterogeneity due to intratumoral hemorrhage, cystic degeneration, and necrosis [4] . Thereby, texture analysis of CE T1-w MRI could provide quantitative information about the tumor signal intensity patterns that are difficult for the human eye to perceive. Introduction: Primary central nervous system lymphoma (PCNSL) is a rare form of highly-aggressive lymphoma. Magnetic Resonance Imaging (MRI) is widely performed for the diagnosis, staging and response assessment. Radiomics extracts quantitative features from medical images to improve diagnosis, prognostic and personalization of therapy in the oncology field. One of the limitations of radiomics in MRI concerns the reproducibility of the extracted features. MRI intensities are not standardized, and they strongly depend on the type of scanner, contrast sequence and acquisition parameters [1] . Herein, we report a study focused on the effect of signal intensity normalization on the reproducibility of radiomics features extracted from MRI scans acquired in different centres and with different scanners. Methods: Twenty-three patients scanned with T1 weighted (T1w) MRI in 3 different time points were selected from the PCNSL Institutional Database of San Raffaele Hospital (Milan). All images were pre-processed as follows: bias field correction, skull stripping, intensity normalization and voxels resampling [2] . Z-score, White-Stripe and Nyul [1] normalization methods were applied for MRI intensities normalization. In the region of the pons, where no pathological modifications were seen, a sphere of 1 cm of diameter was created. This choice was made to estimate the sole effect of normalization, thus avoiding any possible disease-specific signal intensity modification. Radiomic features were extracted using the Pyradiomics library; morphological features were not included in the intensity normalization analysis. The Interclass Correlation Coefficient (ICC) was calculated to evaluate the reproducibility of each feature for both original and normalized images [3] . Kruskal-Wallis test and its post hoc were used to compare the ICCs obtained with different normalization methods. Results: Figure 1 shows median ± quartiles of ICCs computed on both original and normalized images. The highest ICC was obtained for features extracted from images normalized by Z-score, with a significant improvement (p \ 10-9) of about 30% on average if compared to original images. Discussion: Intensity normalization in T1w MRI has a significant impact on the extracted features. Z-score method increases the reproducibility of radiomic features. A larger image study is required to confirm these preliminary results and to better investigate the effect of normalization for prognostic radiomics studies of PCNSL. S6.P10. Using transfer learning for IDH mutation prediction in gliomas using whole brain diffusion anisotropy indices layers of each network were fine-tuned for 100 epochs to volumetric whole-brain DAI maps, with Adam optimizer (learning rate = 10-4) using TensorFlow [8] . Google Colab GPU environment was used to train the networks. Results: The training and validation accuracy of each model are shown in Fig. 1 , and the performance on the test set is shown in Table 1 . In the test set, ResNet-34 trained for 100 epochs predicted IDH status best with 70% accuracy (AUC: 0.87). Discussion: This study demonstrates the potential of transfer learning with ResNet in predicting IDH mutation from whole-brain DAI maps. However, the relatively small cohort size led to overfitting. Furthermore, optimizing model depth may decrease the difference between training and validation accuracies. Introduction: Meningiomas are the most common intracranial brain tumor and accurate pre-treatment grading is crucial to guide treatment decisions [1] . Various deep learning models were tested to non-invasively predict meningioma grade using apparent diffusion coefficient (ADC) images coupled with other magnetic resonance imaging modalities [2] . In this study, a ResNet-50 model was trained to differentiate low-grade and high-grade meningiomas using diffusion anisotropy indices (DAIs). Methods: In this IRB approved study, 76 meningioma patients (46F/ 30 M, 50.74 ± 12.89 years of age, grade I: 34, grade II: 39, grade III: 3) with post-contrast T1-weighted MRI (T1-post, TE/TR = 10/589, 0.69 9 0.69 9 3.6 mm, 40 slices) and diffusion tensor imaging (DTI, TE/TR = 69/4900, 1.8 9 1.8 9 2.34 mm, 42 slices, b = 1000 s/ mm 2 , 30 directions) were included. The patients were stratified into low-grade (Grade I) and high-grade (Grade II and III) groups for classification. Enhancing tumor lesions were semi-automatically segmented [3] on T1-post MRI and registered to DTI space [4] . DAIs (ADC, fractional anisotropy (FA), relative anisotropy (RA)) were calculated from diffusion eigenvalue images [5] , then enhancing tumor regions were masked. Masked DAI images were cropped to include the tumor slice with the largest cross-section including the former and latter slices. These 3 slices in each DAI map were resized to (224, 224) and combined to create an RGB image as the input to the deep learning model. The cohort was stratified into training, validation, and test sets with a 80/20 split ratio between train-test and train-validation sets. A ResNet-50 network [6] was trained for 50 epochs using Tensorflow [7] . The model was compiled with Adam optimizer, whose learning rate was initialized as 10-4 and reduced when validation accuracy plateaued during training. Three ResNet-50 models were separately trained for ADC, FA, and RA. Results: Training and validation accuracy plots of the ResNet-50 model are given in Fig. 1 , and the model performance on the test set is shown in Table 1 . FA images performed best among all DAIs with a test accuracy of 63% and 0.76 test AUC. Discussion: This study demonstrates that deep learning can be used to differentiate between low-and high-grade meningiomas. However, the difference between training and test accuracy indicates overfitting and can be improved by increasing the cohort size. Introduction: Asymptomatic internal carotid artery stenosis (ICAS) is a major public health issue 1 and causes up to 15% of all strokes. 2 Often these strokes are located at the edge of vascular territories, i.e. in individual watershed areas (iWSAs), which are most vulnerable to hemodynamic dysfunction. 3 It is currently unclear, which of several hemodynamic MRI parameters are best suited to predict disease severity and whether hemodynamic changes within iWSA have a higher discriminative ability. 4 Identifying most sensitive parameters and volumes of interest (VOIs) can increase clinical applicability, provide deeper understanding of the pathology and point out most relevant parameters for further research. To this end, we aim to predict ICAS by applying a random forest classifier (RFC)5 to an extensive set of eight multi-modal MRI parameters within and outside iWSA in patients and healthy controls (HC). 6 Methods: Twenty-four asymptomatic unilateral ICAS-patients (70.6 ± 6.4 years) and 24 HC (70.4 ± 4.6 years) underwent multimodal MRI (Philips 3 T Ingenia) comprising breathhold-fMRI, mqBOLD, pseudo-continuous ASL and DSC-MRI. 6 Eight perfusion, oxygenation and microvascular parameters were included (see Fig. 1 ). Feature vectors were generated from mean parameter values inside and outside iWSA 3 in grey matter (GM) and white matter (WM) for each hemisphere and for ipsi-vs. contralateral differences. An RFC (MathWorks MATLAB 9.8) was trained for feature ranking using 96 features (12 VOIs 9 8 parameters). Importance scores and out-of-bag accuracies were calculated. 5 Results: Bootstrapped importance scores of an RFC trained with 96 features are shown in Fig. 1 . After eliminating correlated features, highest ranked features in order of decreasing importance scores are inter-hemispheric differences of TTP in WM, CBF in GM and ipsilateral CVR in GM, all inside iWSAs (Fig. 2) . Using only the highest ranked feature as input, the accuracy is 73.2%. It rises to 82.3 ± 2.8% and 87.7 ± 2.1% (area under the curve 0.88) for the three and seven most important features, respectively (Fig. 2, 3) . Actually, more than four features do not provide significant improvement. Whole hemisphere GM and WM VOIs, neglecting iWSAs, yielded a significantly lower accuracy (81.7 ± 2.3% for seven features, t-test, p \ 0.001). We successfully applied an RFC to determine most relevant parameters and VOIs to predict ICAS. Ranking order and increased accuracy of hemodynamic parameters within iWSAs are in line with 3, 6 . Few sensitive features, i. e. difference of TTP, CBF and ipsilateral CVR allow to detect ICAS patients. In conclusion, RFCs can identify most sensitive hemodynamic parameters to predict ICAS and may reveal most relevant features for diagnostics and further research in cerebrovascular diseases. Decomposition (CPD) ( Table 1 ). In order to compare the two approaches, a non-parametric Mann-Whitney U test was applied using a standard threshold of significance (p = 0.05) as shown by the numbers in Table 1 in bold. Discussion: In this study, a new approach for multiparametric blind source separation using the SDF paradigm was proposed. Higher levels of performances were observed using multilinear rank-(Lr, Lr, 1) compared to the more classical and more rigid CPD model, showing the capability of the GM tissue to discriminate MS clinical profiles. Single factorization of each morphological tensor was also investigated obtaining pour results, demonstrating the usefulness of the fusion approach. Introduction: Abnormal concentrations of tau, beta-amyloid (AB) proteins, and neurodegeneration are biomarkers of Alzheimer''s Disease (AD), but less is known how they interact. Therefore, we aimed to predict tau positivity (Tau ?) using magnetic resonance imaging (MRI) biomarkers in AB positive (AB ?) sample of AD and mild cognitive impairment (MCI) subjects. Methods: We selected an age-and gender-matched sample from the ADNI database: AD AB ? (n = 12, age = 74.2, 50% female), MCI AB ? (n = 26, age = 71.3 ± 5.8, 54% female) and Controls (CL) (n = 35, age = 71.2 ± 6.1, 43% female)-AB negative (AB-)subjects. Tau positivity was obtained using composites of rSUV computed from Flortaucipir PET for each Braak stage [1] and a global composite tau-rSUV (regions showing statically significant differences (p \ 0.05) in NL AB-vs (MCI ? AD) AB ?). The cutoff value of each composite was the one that best discriminates NL ABfrom (MCI ? AD) AB ? .i.e., the value that maximizes the chisquare statistic value with the highest significance. We used machine learning models (voting classifier) to predict Tau ? with demographic (age and gender), clinical (Geriatric Depression Score (GDS) and Mini-Mental State Exam (MMSE) score), and brain volumes from MRI. The best features were selected using Spearman correlation or chi-square statistics. Results: The voting classifier was able to classify Tau ? with accuracies above 70.0% using the cutoff values of Braak1, Braak34 [the respective areas, please see [1] and the composite tau-rSuv ( Table 1 ). The best performance was achieved using the composite Braak1. The right middletemporal, right parahippocampal, left accumbens area, left amygdala, left lateral ventricle and right accumbens volumes, and the GDS score were selected to classify the three rSUV composites. Discussion: Our results show that cortical and subcortical volume changes are potential biomarkers to predict tau-PET in an AB ? sample. The fact that only Braak1, Braak34 and composite tau-rSUV could predict tau-PET may be associated with the fact that the cutoffs values were obtained from CL AB-vs (MCI ? AD) AB ? , as MCI AB ? may present less atrophy in the brain than AD. The best three classifications shared volumes supporting the Braak et al. hypothesis, wherein the severity of brain atrophy increases with tau protein progression [2] . Further studies should be conducted to confirm that the selected regions are only characteristic of AD and MCI tau ? and AB ? patients. S6.P17. Prediction of head motion parameters by measuring extracranial magnetic field changes during multi-slice EPI acquisition *L. Bortolotti 1 , R. Bowtell 1 1 University of Nottingham, Sir Peter Mansfield Imaging Centre, Nottingham, GB Introduction: Previously, head motion parameters were predicted from NMR field camera measurements without concurrent image acquisition 1 . Here, prediction was performed by making measurements of extra-cranial field changes in the quiet periods of a standard multi-slice EPI acquisition. The results represent a step forward in integrating standard imaging with a marker-less motion monitoring technique. Methods: The field probes were sited between the transmit and receiver RF head coils using a customized probe holder. An MPT camera was also sited in the magnet bore to allow simultaneous measurements of head position and extracranial magnetic field to be made during the training phase (Fig. 1a) . Multi-slice EPI sequence was performed with a slice TR of 75 ms (48 slices, 3 mm isotropic resolution, TR = 3.6 s, TE = 20 ms). The field camera acquired field measurements every 150 ms (synchronised with every 2nd slice acquisition), with the 10 m s acquisition commencing 30 ms after the slice excitation, so limiting the effect of eddy currents from the applied imaging gradients. Measurements of extracranial magnetic field changes and head motion parameters were acquired with and without simultaneous scanning with the subject executing a range of different head movements (head shaking/nodding, foot wiggle, rest). Head motion parameters were predicted from the extra-cranial field measurements by using a 1-hidden layer delayed feedback neural network (NARX) 1 . Results: The magnetic field changes measured with and without simultaneous scanning have similar magnitude and temporal variation for comparable head movements (Fig. 2) . Prediction of motion parameters was successfully performed (Fig. 3) . The best results were produced when the training dataset was formed from the movement conditions (rest and foot wiggle) involving smaller head displacements. The Table (Fig. 3 ) compares predicted motion parameters to the actual values. Discussion: Head motion during EPI scan acquisition can be characterised in a 7 T MRI scanner by measuring the extra-cranial magnetic field changes using a field camera. Prediction accuracy was improved by forming a training set from data acquired with only small head movements. The acquisition of less than 5 min of training data is then required during which time head movement parameters need to be simultaneously measured using a different motion monitoring approach. Here, we used an MPT camera for concurrent measurements, but in the future, training could be based on analysis of recorded image data. This work represents a step towards the full development of a marker-less technique for head motion tracking that doesn''t not require modification of the image acquisition sequence. Introduction: The effect of aging on the volume, T1, T2*, and susceptibility of subcortical structures has been studied 1 . Although venous atlases gained interest recently 2,3 , venous assessment in subcortical regions is challenging due to surrounding iron-rich structures 2 . Further, MR-based assessment of aging-induced changes in the venous vasculature remains understudied. Here, the recently introduced vessel distance mapping (VDM) 4 along with ROI-based vessel densities was applied to assess if the subcortical venous vasculature declines with aging. Methods: The openly available ATAG study was re-used. It provides 7 T GRE and MP2RAGE data of 30 young, 14 middle-aged, and 9 elderly subjects 5, 6 . From the 0.5 mm isotropic GRE, a high-passed, QSM-based SWI was computed with QSMbox 7 . The high-pass minimizes enhancement of non-vessel structures. Vessel segmentation was done with OMELETTE 8 . Vessel densities were computed per ROI as the ratio of detected vessel volume to total ROI volume. From the segmentation, VDM computes for each non-vessel voxel the distance to the closest vessel using the Euclidian distance transform. The workflow is shown in Fig. 1 . To build age-specific atlases, the GRE data was registered nonlinearly to MNI space with ANTs 9 via the additionally provided MP2RAGE slab and whole brain volumes. For each age group, the respective ATAG atlas was used to compute group statistics of the VDM and vessel densities. Kolmogorow-Smirnow tests were used to detect statistically significant differences. The averages per age group are shown in Fig. 2 . The sparse vessel information is interpolated by VDM, yielding distance maps resembling the underlying structure (distances increasing from CSF to white matter) and providing non-zero values even if no vessel is detected within the ROI. Figure 3 shows the group results for VDM and vessel densities for subcortical regions. Only the striatum showed the expected aging-related decrease in vasculature as lower densities and large distances, respectively. Other regions show an unexpected U-shape with decreased vessel distances in the middle-aged group. Significant differences were found for VDM in the internal globus pallidus and for densities in the striatum (see Fig. 3 ). Discussion: Assessment of aging-induced changes in the subcortical vasculature is feasible with the presented pipeline. Except for the striatum, the results show an unexpected aging-vasculature-relation, requiring future investigations to assess if these are due to bias in the data acquisition/analysis or actual changes in the venous vasculature. Acknowledgment: This work was funded by the DFG-MA 9235/1-1 motion/eddy current correction/spatial warping to MNI152 atlas. Conversion of BOLD signals from time to power spectral density, detrended fluctuation analysis, and FD calculation was done with Matlab. Using normative age/sex matched controls, brain regional Z-score (ZFD) was calculated. ROI-based analysis was used to compare 91 regions from each brain with ZFD B -1.68 defined as clinically important. Overall disease burden (ODB) was defined as the sum of all regions with significantly lower Z-scores (i.e., \ 1.68 standard deviations lower than the mean). Personalized regional brain FD anomalies were compared against regional brain hippocampal volumes determined using Neuroquant [vi] . The most affected brain regions were the secondary somatosensory cortex (74%), Broca's area (63%), the amygdala (53%), insula (42%), and hippocampus (37%). For predicting ODB, significant metrics were the hippocampal volume (P \ 0.004), sex (P \ 0.04), and age (P \ 0.0005), with interaction terms of age:sex (P \ 0.02) and age:hippocampal volume (P \ 0.002). FD may thus be a useful metric in the assessment of clinically suspected AD. Future analysis will include neurocognitive scores and additional brain measures. Introduction: Time-lapse MRI provides a non-invasive tool for three-dimensional tracking of single labelled immune cells with whole-brain coverage in vivo. 1, 2 However, manual cell counting and pattern categorizing is time-consuming and elaborate. We aimed to improve analysis of time-lapse MRI data via an adapted cell tracking tool that links cell positions to trajectories allowing for automated motion pattern categorization. Methods: In vivo time-lapse MRI using a repeated T2* gradient echo sequence 1 of 15 mouse brains was performed on a 9.4 T Biospec (Bruker Biospin) small animal MRI. Monocytes were labelled in vivo by i.v. injection of Ferucarbotran (Resovist, Bayer AG) 24 h before MRI. First, detected cells were counted manually and were categorized in short-term short-range (1-2 consecutive time frames), long-term short-range ([ 2 consecutive time frames), or long-term long-range ([ 2 consecutive time frames and cell movement) motion patterns. Second, for automated categorization, cells were tagged in ImageJ and coordinates were linked to trajectories ( Fig. 1 ) using the corresponding part of the MATLAB implementation by D. Blair and E. Dufresne 3 of the IDL particle tracking code of J. Crocker, D. Grier, E. Weeks 4, 5 . Hereby, patterns were classified as before with a defined minimum travel distance of 1.2 pixels for a long-term long-range pattern. Mean velocity of moving cells was also calculated. Results: For comparison of manual categorization versus the automated tool, Bland-Altman analysis was performed resulting in a positive bias for the total number of detected cells (2.53 ± 7.23) as well as for the short-term short-range (2.33 ± 7.84) and long-term short-range (2.47 ± 3.50), and a negative bias for long-term longrange (-2.27 ± 3.08) motion pattern (Fig. 2) . Overall, reliability of the automated approach seems to be higher when fewer cells are tagged. Furthermore, the mean velocity of moving cells was determined to be 0.20 ± 0.06 lm/s (Fig. 3) , matching the speed of patrolling monocytes. 6 Discussion: The automated cell tracking tool allows for a reliable analysis regarding motion pattern categorization of iron-labelled monocytes in time-lapse MRI of mouse brains and can therefore be used to analyze data more efficiently. To further improve observer dependence of time-lapse MRI data analysis, we aim at developing a tool that automatically detects labelled cells too. Proceedings of 29th ISMRM Proc. 22nd ISMRM Proc. 29th ISMRM Proc. 21st ISMRM Grau 2020 Rational approximation of the Ernst equation for dual angle R1 mapping revisited: beyond the small flip-angle assumption References Numerical Recipes in C ? ? : The Art of Scientific Computing S6.O6. QC Flex: a flexible quality control tool for large MRI cohorts Nodal detection in head and neck cancer by USPIO enhanced MRI: Where are the USPIOs in the blood? Biological and clinical manifestations of juvenile Huntington's disease: a retrospective analysis ORYX-MRSI: A data analysis software for multislice 1H-MRSI Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI Simultaneous in vivo spectral editing and water suppression Hc and the measurement of neuronal pH in patients with epilepsy Elevated Hc and GABA in subject on isoniazid as assessed through 1H MRS at 7 T Decreased brain pH as a shared endophenotype of psychiatric disorders The in vivo neuron-to-astrocyte lactate shuttle in human brain Proceedings of 29th ISMRM, no. 4195. S3.P2. Free-breathing 3D alpha-mapping for ventilation quantification MoDL: Model-based deep learning architecture for inverse problems A Clinically Viable Vendor-Independent and Device-Agnostic Solution for Accelerated Cardiac MRI Reconstruction Groupwise Non-Rigid Registration with Deep Learning: An Affordable Solution Applied to 2D Cardiac Cine MRI Reconstruction Acknowledgements: FCT (SFRH/BD/120006/2016, PTDC/EMD-EMD/29686/2017) Acknowledgements: FCT(SFRH/BD/120006/2016 SS-ZTE patent P202030504 Spiral blurring correction with water-fat separation for magnetic resonance fingerprinting in the breast Inductive measurement and encoding of k-space trajectories in MR raw data Fast dynamic ventilation MRI of hyperpolarized Xe using spiral imaging References: 1. www.anaconda.com/state-of-data-science /project/dcm2bids Mutsaerts 2020 NeuroImage Göttler, JCBFM, 2020. 3: Kaczmarz, JCBFM, 2020. 4: Asllani, HBM, 2009. 5: Mutsaerts, NeuroImClin, 2014. 6: Mutsaerts, JCBFM, 2017. 7: Asllani, MRM S4.P12. How independent is the community structure of static and dynamic functional networks from the brain's underlying structure? Decoding subject-driven cognitive states with whole-brain connectivity patterns A cortical rat hemodynamic response function for improved detection of BOLD activation under common experimental conditions Qualitative sex differences in pain processing: emerging evidence of a biased literature Neurovascular coupling: fMRI of human brain during visual stimulation Free water DTI at high and standard spatial resolution -optimal parameters and fitting procedures Arribarat 1 , Y. Fave 1 , H. Gros-Dagnac 1 Concp Magn Reson Part B: Magn Reson Eng Kashyap 2021, ISMRM Glasser 2013 NI Proc. ISMRM 2021 #1784 References (DOI): 1 Assessment of the safety of magnetic resonance imaging for patients with an active implantable medical device A technique to evaluate mri-induced electric fields at the ends of practical implanted lead Technical specification ISO/TS 10974 www Low-field MRI: how low can we go? A fresh view on an old debate Calculation of the magnetic stray field of a uniaxial magnetic domain A Fast 0.5 T Prepolarizer Module for Preclinical Magnetic Resonance Imaging Research Project Summaries-American Brain Tumor Association Proc of the 25th ACM SIGKDD International Conference on Knowledge Discovery&Data Mining Department of Simulation, Imaging and Modelling for Biomedical Systems Magnetic Resonance Imaging Core Facility, Barcelona, ES Introduction: Animal models of Alzheimer''s disease (AD) are essential in the research of the pathology and potential treatments. TgF344-AD rats progressively show all AD pathological hallmarks This can provide a translational approach to identify imaging biomarkers and predict AD progression. Methods: 30 Tg344-AD rats and 34 WT littermates underwent MRI on a 7 T Bruker scan at different ages each from 3 to 18 months, resulting in 194 images References S6.P9. The effect of MRI signal intensity normalization for radiomics analysis on PCNSL patients *M Standardization of brain MR images across machines and protocols: bridging the gap for MRI-based radiomics Impact of image preprocessing methods on reproducibility of radiomic features in multimodal magnetic resonance imaging in glioblastoma ) before surgery. IDH mutation was identified using Sanger sequencing with an Applied Biosystems TM 3500 Series Genetic Analyzer. DAIs (apparent diffusion coefficient (ADC), fractional anisotropy (FA), relative anisotropy (RA)) were calculated from diffusion tensor eigenvalues[5]. For preprocessing first and last 5 slices of DAI maps were removed. Volumetric ADC, FA, and RA data were assigned to one of the 3 input channels of the deep learning model. No other preprocessing was performed. Data were split into training and test sets retaining 90% of data in the training set. Then the training set was further split into a training and validation set retaining 90% of the data in the training set Proc. of the IEEE CVPR Proc. of USENIX OSDI Proc. of the IEEE CVPR Proc. of USENIX OSDI References: 1: de Weerd, Stroke, 2010. 2: Kamel, Stroke, 2019. 3: Kaczmarz, Neuroradiology PET Imaging of Tau Deposition in the Aging Human Brain Potential pathways of abnormal tau and a-synuclein dissemination in sporadic Alzheimer''s and Parkinson''s diseases Abdominal obesity and metabolic syndrome Das Design der Machbarkeitsstudien für eine bundesweite Kohortenstudie in Deutschland Whole-Body MR Imaging in the German National Cohort: Rationale, Design, and Technical Background nnU-Net: a self -conFiguring method for deep learning-based medical image segmentation Methods: In this IRB-approved study subgroups: double negative (-), IDHonly (? -), TERTonly (-?), double positive (? ?) with 9 Discussion: This study demonstrates the potential of MGF parameters in NAWM for predicting tumor genotype Complexity analysis of resting state functional magnetic resonance imaging (rsfMRI) data as an early diagnostic tool for Alzheimer's disease (AD) References GB Introduction: Quantitative susceptibility mapping (QSM) at 7 T provides a powerful approach for characterising brain iron concentration at the sub-millimetre level1,2. Changes in tissue iron have been linked to a range of neurological conditions. In preparation for measurements of longitudinal changes in the hippocampal sub-fields in Alzheimer''s disease, we have characterised the reproducibility of high-resolution 7 T susceptibility measurements in the hippocampus Hippocampal segmentation involved applying the open-source, ASHS software5 to the T1-and T2-weighted images using a pipeline that combines multi-atlas6 label fusion and learning-based error correction. Regions of interest (ROIs) delineated using this approach are: Cornu ammonis (CA) areas: CA1, 2 and 3, hippocampal tail (TAIL), dentate gyrus (DG), subiculum (SUB) & entorhinal cortex (ERC). The average value of the susceptibility and volume of each of ROI was evaluated for each scan. Volumes of CA2 & CA3 were too low to produce reliable susceptibility values, so CA1-3 values were combined into a single CA measurement. Results: Figure 1 shows a susceptibility map, with overlaid hippocampus subfield classification for one subject Patient-specific connectivity pattern of epileptic network in frontal lobe epilepsy References 1 The Matlab Particle Tracking Code Repository Methods of Digital Video Microscopy for Colloidal Studies Particle tracking using IDL S198 Magn Reson Mater Phy S2.O9, S3.O3, S3.P1, S3.P2 S2.P25 S200 Magn Reson Mater Phy A. S4.P2, S4.P3 A. S6.P10, S6.P11, S6.P21 Ö zduman, K. S6.P10, S6.P11, S6.P21, S6.P7, S6 O4 Ö ztürk-Işık, E. S6.P10, S6.P11 Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations Discussion and conclusion: SANDI maps obtained from both magnitude and real data follow expected patterns with high f soma in GM and high f neurite in WM. Rician noise affects mostly f neurite which has slightly higher values when estimated from magnitude data, especially at lower SNR. The developed framework to identify the optimal clinical trajectory from a given set of trajectories and then estimating the power dissipation at the hotspots is quiet robust and generic. The framework is applicable to arbitrary structure with multiple electrode leads and helpful in optimizing the RF response with respect to the lead trajectories.Conclusions/discussion: We have demonstrated the advantages of prepolarization on long T1samples, with a home-made PMRI system, and the observed increase in signal-to-noise ratio agrees with our models for the electronics and time evolution of the magnetization. The prepolarization pulse fall times are * 10 ms and can be made even shorter (\ 300 us) with high power electronics under development, enabling hard tissue PMRI [3].Thanks to its demonstrated performance, wide applicability to lowfield MR, accessible documentation and friendly user community [9], MaRCoS is rapidly growing in scope, with parallel multi-device operation in development. It is already able to provide much of the functionality of a modern MR spectrometer at a fraction of the cost and without the drawbacks of proprietary systems. We acknowledge the generous contributions of Suma Anand, Hengjie Liu and Ivan Fomin. Acknowledgments: This study has been supported by TUBITAK 1001 grant 119S520. We showed that susceptibility distortion correction significantly improved the tSNR of pCASL-3D GRASE data across the whole brain, with impact on perfusion in some regions. Improvements in tSNR have been previously reported for BOLD signals, even with static correction 4 , but not for ASL. This step should hence be integrated in the ASL preprocessing pipeline, especially due to its intrinsically low SNR.S5.P1. Theoretical validation of simulated B1 for a cavity resonator at 300 MHz Introduction: A transceiver volume coil based on Mansfield 's cavity resonator design was previously developed and showed a better uniformity and performance compared to the standard birdcage [1] . The B 1 improvement of this resonator design was also demonstrated theoretically [2] . However, the study of RF coils importantly rely on numerical simulations which usually facilitate the computation of B 1 . These results motivated us to theoretically validate its numerical magnetic field. Method: B 1 numerical simulations of a cavity resonator with 6 rungs and 6 circular cavities were conducted at 7 T. This coil design is intended to be used in preclinical MRI. All numerical computations were performed with the commercial software tool COMSOL Multiphysics (COMSOL 3.2, Burlington, MA, USA). Coil was simulated in the quadrature-drive mode and tuned at 300 MHz. Figure 1 .a) shows an illustration of the coil design used in all simulations. We used the following analytical expression for the magnetic field of the cavity resonator, B 1 [3] :where l = 9 cm (length) and d = 9 cm (diameter) and i cav is the current. Results and discussion: B 1 numerical simulations were computed and and show in Fig. 1 .b). A really good uniformity and intensity of the magnetic field can be observed for this particular design. Then, a comparison plot was computed using the numerical data and theoretical calculations obtained with the equation above. Figure 2 shows the comparison plots. The simulated profile was taken along the white line in the B 1 simulation, as shown in Fig. 2 . A very good correspondence can be observed between numerical and theoretical computations. This supports the electromagnetic simulations of our transceiver resonator design. These results will allow us to reliably study the coil performance using different layouts. [1] . A straightforward theoretical frame of B 1 is most desirable to actually validate experimental results obtained with the popular birdcage coil. We derived an expression to compute B 1 of a birdcage coil at high frequency. Method: We derived an expression for B 1 based on the theoretical work reported in [1, 2] , together with the expression obtained by De Zanche et al. for a 4 rung birdcage coil [3] : (1,i,0) is a unit vector in the direction of the circularly polarized RF field generated by the volume coil, and n = r -1 (x,y,z) is the unit normal vector to the radial direction. Functions j(kr) represent the Bessel polynomials. l and d are the length and the diameter of the birdcage coil, respectively. Phantom images were acquired with a standard spin echo sequence. The acquisition parameters were used: TE/TR = 25 ms/900 ms, FOV = 40 mm 9 40 mm, matrix size = 256 9 256, slice thickness = 2 mm, NEX = 1, with a transceiver quadrature birdcage coil (4 rungs, length 0 12 cm and diameter = 7.7 cm) at 7 T [4] . Figure 1 .a) shows a photograph of the coil prototype. Results and discussion: Unlike other difficult formulas for B 1 , reported in the literature, our theoretical model is fairly easy to follow for the further developments of this type of volume coils. A comparison plot of the theoretical and experimental results is shown in Discussion: Overall, results suggest that both CPNs are good candidates to study whether they reach and accumulate in the tumor of an orthotopic and xenograft GBM model. : Images with good SNR could be obtained in times of a few minutes at 7 T with 5 mm resolution in subjects loading with D 2 O to * 1.5% concentration (* 100 times natural abundance). Similar levels of signal enhancement were found over the 6-week loading period. T 1 -values for deuterium in HOD were significantly shorter than corresponding values for 1 H in H 2 O [5] , as expected due to the effect of quadrupolar 2 H relaxation. 2 H and 1 H, T 2 * values were similar, but the quadrupolar 2 H relaxation did not show strong sensitivity to iron content in deep grey matter (Fig. 2 ).• Evaluating and respecting fitting uncertainties is equally crucial for DL and traditional approaches. Introduction: Meningiomas are the most common adult brain tumors [1] . The 5-year recurrence rate in grade I meningiomas is 5% while this rate is 40% in grade II (atypical) meningiomas [2, 3] . Furthermore, grade III (anaplastic) meningiomas have the worst overall survival rate of up to 2 years [3] . Therefore, tumor grading plays a significant role in the prognosis and treatment planning of meningiomas [4] . This study aims to develop a one-dimensional convolutional neural network (1D-CNN) to determine the grade of meningiomas based on single-voxel proton magnetic resonance spectroscopic (1H-MRS) data. Methods: Fifty-seven meningioma patients (15 M/42F, mean age: 52.55 ± 13.22 years, range: 18-80 years, 20 WHO grade I, 37 higher-grade (34 WHO grade II and 3 WHO grade III)) were included in this study. The patients were scanned at pre-surgery time points using a 3 T clinical MR scanner (Siemens Healthcare, Erlangen, Germany) with a 32-channel head coil. 1H-MRS was obtained from a manually placed region of interest with no necrosis, edema, and hemorrhage, using a Point Resolved Spectroscopy (PRESS) sequence (TR/TE = 2000/30 ms). Fitted spectra obtained by LCModel [5] were used in a 1D-CNN model to classify the grades of meningiomas. Preprocessing steps including L2 normalization, smoothing (Savitzky-Golay filter (window size = 11, order = 2), Yeo-Johnson power transformation, and min-max normalization were applied on the input data before classification [6] . Optuna [7] was employed for hyperparameter tuning of the models with 50 trials. ADASYN was used to overcome the imbalanced dataset problem. Introduction: Shape-based radiomic features, like sphericity index, have been associated with overall patient survival [1] . However, 3D representation of tumors might lack natural look and smoothness (staircase artifact) due to the anisotropic voxel sizes in MRI, which might potentially compromise the morphological measurements of segmented tumors. This study aims to assess the effect of staircase reduction methods on shape features of T2-weighted MRI of meningiomas. Methods: Fourteen radiomic shape-based features were extracted from T2-weighted MRI of 19 patients with meningiomas. All segmentations (0.21 9 0.21 9 3.6 mm (N:13), 0.26 9 0.26 9 6.5 mm (N:2), 0.24 9 0.24 9 6.5 mm (N:1), 0.43 9 0.43 9 3.6 mm (N:3)) were resampled to the same resolution of 1 9 1 9 1 mm. The first method (M1) directly extracted shape-based radiomic features, whereas the second method (M2) applied median smoothing with kernel size set to the original slice thickness (Fig. 1 ). On the other hand, the third strategy (M3) used morphological contour interpolation [2] followed by median smoothing. Then, several shape features were calculated using PyRadiomics [3] . A Friedman test followed by Wilcoxon post-hoc test was used to determine differences in extracted features. For Friedman test, the cut-off P value was 0.0036 due to Bonferroni multiple comparisons, and for Wilcoxon post-hoc test a P value of 0.05 was considered statistically significant. The influence of smoothing and interpolation strategies was assessed by intraclass correlation coefficients (ICC) and confidence intervals (CI).Results: M1 had the highest surface area and surface to volume ratio (P \ 0.01 for all) (Fig. 2 ). In contrast, M1 had the lowest sphericity index (P \ 0.0001 for all). M1 resulted in higher maximum 2D diameter column and maximum 3D diameter than both M2 and M3 (P \ 0.01 for all). On the other hand, M2 had a significantly lower maximum 2D diameter slice, mesh volume, and voxel volume than the other approaches (P \ 0.05 for all). M1 showed a significantly higher minor axis length than M2 (P \ 0.001). The remaining five measures did not significantly differ between the methods (P [ 0.004). The stability of features was excellent for 13/14 radiomic features (ICC [ 0.9), but not for the sphericity index (ICC = 0.82, CI = 0.17, 0.95) (Fig. 3 ). The shape features calculated after M1, M2, M3 preprocessing approaches were similar in intensities except for the sphericity index, which was highly affected by the pre-processing approach. Our results indicated that the shape-based radiomic features were generally stable across different smoothing and interpolation strategies. Introduction: Machine learning (ML) has been recognized as an emerging novel technology for building an automatic image reconstruction pipeline from large data sets. However, most of the current general-purpose ML frameworks are not available to clinical applications, which are limited by the scarcity of data and the lengthy learning times. Recently, the generative interpolation network has been found to learn the feature with fewer data. This can save time and deliver high throughput. Previously, we demonstrated a feasible model for reconstructing non-Cartesian MRI. In this study, we examine the new PyNUFFT based network [1] for static BLADE MRI [2] . Method: The comparative study was approved by the Institutional Review Board (IRB). A total number of 32 in vivo T2 weighted brain MRIs were obtained from the DICOM database for evaluation. The k-space of the BLADE sequence includes multiple overlapping rectangular k-space patches, which cover the circular region in k-space while sharing the central k-space. No patient data were needed to build the network, and the complex-valued back-propagation was performed by quasi-orthogonal complex random bases generated in the image space and k-space. An interpolation sparse matrix in k-space and the image domain were obtained from a Python package. We used Numpy and Scipy to implement the network. With the coordinate format (COO) and a batch number of 3, the learning time on a single core CPU is fast and the total learning time for 7500 epochs was 14 h. The reconstruction results of 2500, 5000, and 7500 epochs were compared against conjugate gradient (CG) and density compensation (DC) methods. Two board-certified radiologists evaluated the subjective image quality by means of a five-point score.Results: The structural similarity metric (SSIM) of the five reconstruction methods ( Fig. 1 (A) -(E)) ranged from 0.993 to 0.999, which were significantly different (p \ 0.001).The SSIM of DC was the lowest (0.993 ± 0.0038), while the four other methods appeared to perform similarly (0.999 ± 0.0002). The SSIM of the CG method was significantly superior to the learning-based methods (p \ 0.001). Two radiologists reported that all five reconstruction methods were visually identical (see Fig. 1 ). The subjective image quality scores were: noise level (4.67 ± 0.54), tissue contrast ( Introduction: MS is an autoimmune inflammatory disease characterized by demyelination and neurodegeneration processes, leading to cognitive and physical impairments [1] . Brain morphological connectivity represents a new method for characterizing the brain networks which can be obtained directly from the anatomical T1w image by measuring the Gray Matter (GM) features [2] . The goal of this work is to exploit the multiparametric GM information obtained from three morphological features of the cortical GM such as: thickness, curvature and area. The multi-features connectome data were combined together using tensor-based Structured Data Fusion (SDF) technique [3] for the unsupervised classification of MS clinical profiles.Methods: A tensor-based multilinear rank-(Lr, Lr, 1) [4] SDF application was used to jointly factorize three GM morphological tensors (Thickness, Curvature and Area) describing characteristics of GM tissue (Fig. 1) . 90 patients, distributed in four clinical profiles (12 CIS, 30 RRMS, 28 SPMS, 20 PPMS) underwent an MR examination protocol composed of a 3-dimensional T1-weighted MPRAGE sequence with repetition time/echo time/time for inversion (TR/TE/ TI) = 1970/3.93/1100 ms, flip angle = 15°, voxel size = 1 9 1 9 1. The parcellation task was performed using the FSAverage atlas [5] yielding 68 brain regions in total. Afterward, each voxel was classified into four classes [WM, cortical GM, sub-cortical GM, cerebrospinal fluid (CSF)] and the three morphometric features were calculated. The connectome data were obtained comparing all GM regions yielding a total of three connectomes for each patient which were stacked together obtaining three final tensors (Fig. 2) . The multilinear rank-factorization (Lr, Lr, 1) was then applied to combine together the three tensors, obtaining a latent representation of MS patients (factor components) which was then exploited for classification using k-means clustering algorithm (Fig. 1) .Results: Good levels of performances were observed using our approach compared to the more classical Canonical Polyadic nnU-Net was trained using n = 30 training cases in a fivefold crossvalidation scheme yielding an ensemble of five separate U-Nets. Each model was used to segment n = 11,191 data sets from the GNC. As the models were randomly initialized and different training cases were used during the training, resulting segmentations are slightly different. Taking advantage, the mean pairwise Dice scores of different ensemble members can be used as a measure for segmentation uncertainty per fat compartment 5 . Statistical measures for outliers are applied to identify faulty segmentations. Results: Figure 1 shows the boxplot of the resulting uncertainty scores for both fat compartments. A total of n = 217 candidates was identified, reducing datasets for manual inspection by about 98%. After manual inspection, n = 15 (about 7%) of the candidates had to be excluded (see Fig. 2 ). Figure 3 shows an example of a misidentified candidate.Discussion: Due to quality control of the GNC data, imaging errors were expected to be rarely found. The application of mean pairwise Dice score as a value of segmentation uncertainty significantly reduced the number of data sets for manual inspection. In a small number of data sets, imaging errors in the regions of interest could be identified, and the data sets are excluded from further analysis. Mean pairwise Dice score is a feasible measure for the identification of imaging errors in large cohorts with very low computational overhead. Repeatability measurements of biomarkers derived from whole-body diffusion weighted imaging (WBDWI) to assess response to treatment on patients with metastatic bone disease using a fully automated software solution The Royal Marsden NHS Foundation Trust, Sutton, GB Introduction: Whole-Body Diffusion Weighted Imaging (WBDWI) demonstrates high sensitivity for detecting metastatic bone disease in patients with advanced prostate cancer (APC) [1] . WBDWI is a quantitative technique that provides non-invasive measurement of the tumour diffusion volume (tDV) and the global Apparent Diffusion Coefficient (gADC), a surrogate of tumour cellularity. However, manual delineation of regions of interest (ROIs) within these datasets is too cumbersome to be clinically viable. Therefore, we have assessed the repeatability of quantitative ADC-based biomarkers derived from automated delineation of metastatic bone lesions on WBDWI.Methods: Nine APC patients were scanned twice using WBDWI prior to initiation of anti-cancer treatment. WBDWI may vary depending on acquisition date, imaging protocol or scanner manufacturer. Therefore, signal on b = 900 s/mm 2 images was normalized using a Noise-Corrected, Exponentially Diffusion Weighted MRI (niceDWI), which synthesizes new contrast using a weighted combination of voxel-wise ADC and ADC uncertainty,, estimated using a recently developed algorithm [2] :with a c = 20,000 s/mm 2 , and b c = 900 s/mm 2 . Signal images were used as inputs to an atlas-based pipeline to automatically delineate ROIs in all repeatability studies. ROIs were transferred onto estimated ADC maps, from which 5 first-order global ADC (gADC) histogram parameters (mean, median, variance, skew and kurtosis) and one size/shape-based parameter (logarithm of tumour diffusion volume, tDV) were derived. Repeatability of these parameters was calculated using Bland-Altman analysis. The intraclass correlation coefficient (ICC), Coefficient of Variation (CoV), repeatability coefficient (RC), and percentage repeatability (%RC) were calculated for each of the WBDWI parameters. Results: Table 1 summarises the values estimated for CoV, RC and %RC. Mean/median/variance gADC and log-tDV showed CoV lower than 10% and ICC higher than 0.9. Figure 1 shows the Bland-Altman plots for each of the WBDWI parameters investigated. Figure 2 demonstrates the best performing case-study of automatic ROI delineation.Discussion: The repeatability of WBDWI parameters derived from an automated software solution showed high repeatability for mean/median/variance metrics derived from histograms of gADC and log-tDV. A previous study indicated a variation of 17% for Median gADC and from -50% to ? 26% for log-tDV after treatment [3] . A variation of this magnitude would be outside the limits of agreement derived in this study. Therefore, DWI-based biomarkers derived from an automated segmentation algorithm with signal normalization using the niceDWI technique may be able to assess successful response to treatment in clinical practice. Introduction: Prior studies proposed that the longitudinal relaxation rate R 1 depends on the orientation of the nerve fibers with respect to the main magnetic field based on molecular dynamics (MD) simulations [1, 2] . The MD simulations proposed that orientation dependency may be mainly attributed to anisotropic dipole-dipole interactions due to structural polarization of water molecules occurring within the myelin sheath [2] . Further, the hydrogen nuclei in the lipid membrane of the myelin sheath reveal a structural conFiguration, which lead to a spatially and temporally asymmetric spin environment causing anisotropic dipole-dipole interactions with neighboring protons [2] . This would result in an orientation dependent relaxation rate R 1 observed in the MD simulations. R 1 orientation dependency has only been examined in a few studies but may be crucial for the extraction of more precise tissue features [1, 2] . The goal of this study was to investigate the behavior of the orientation dependency of R 1 in in situ post mortem conditions. Methods: 13 deceased subjects were examined and stored in a cooling chamber at 4°C prior to study procedures. The in situ brain temperature was measured transethmoidally prior to the MRI scan. In order to determine R 1 orientation dependency, post mortem in situ MRI scans were conducted at 3 T. R 1 was acquired using an inversion-recovery spin-echo sequence [6 TIs: 30-1200 ms, TE/TR: 12/7060 ms, 40 slices, slice thickness 4 mm, in-plane resolution 1 9 1 mm 2 ] and subsequent fitting of a mono-exponential decay signal model. The orientation of white matter fibers were computed using diffusion-weighted single-shot echo-planar imaging [64 isotropically distributed diffusion directions, TE/TR: 109/18700 ms, 100 slices, isotropic resolution of 1.8 mm 3 , b = 2000 s/mm 2 ]. Results: The Figure reveals the averaged post mortem orientation dependent R 1 over all subjects. R 1 relaxation rate increases for an increasing fiber angle with respect to the main magnetic field between fiber angles of h = 0-50°, while R 1 slightly decreases for the fiber angles between h = 55-90°. Discussion: In situ post mortem R 1 reveals a similar orientation dependency compared to the MD simulations of Schyboll et al. [1, 2] . The behavior found in this study of the post mortem R 1 orientation dependency reflects the orientation dependency of the averaged correlation time (reaches maximum at h = 50°) proposed by Schyboll et al. [1] . These post mortem findings would enable to investigate the biophysical meaning of the orientation dependency of R 1 based on brain tissue histology. Introduction: In glioma patients, alterations of diffusion anisotropy indices (DAIs; apparent diffusion coefficient (ADC), fractional (FA), relative (RA) anisotropy) in NAWM may detect differences in varying tumor genotype [1] . However, multimodal Gaussian fitting (MGF) parameters may better model the competing effects of infiltrating cells, in contrast to standard summary statistics [2] . This study aims to investigate the predictive power of MGF parameters of DAIs and diffusion tensor eigenvalues (DTEs; E1, E2, E3) in the NAWM for glioma genotype prediction. Alzheimer's Disease (AD) is a neurodegenerative disorder associated with amyloid (Ab) deposition in the brain leading to cognitive impairment. While the diagnosis is typically based on clinical findings, imaging plays a key supportive role [i] . A tool for early diagnosis could improve understanding and provide better patient outcome. Biological time-varying signals can be classified as statistical fractals [ii] . The brain regional fractal dimension (FD), or temporal complexity, can be determined using resting state functional magnetic resonance imaging (rsfMRI). Once the FD is known, the Hurst exponent can be determined. In AD there is reduced time domain complexity [iii] . Our hypothesis is that AD impacts specific brain regions creating a unique signature, evident in H frequency domain variants, which could serve as a metric for early diagnosis/establishment of disease severity. Specifically, rsfMRI may serve as a tool for the early diagnosis of AD and could improve our understanding of the underlying pathophysiology. Patients with clinically suspected AD had a 3 T MRI with 32-channel head coil. Imaging included high resolution T1-weighted and resting state BOLD data. MR images of all patients was reviewed by a neuroradiologist to remove any patients from this data base, whose findings might also impair cognitive function (like amyloid angiopathy/strokes/…). Post-scan analysis with FSL [iv,v] Introduction: Epilepsy is a disease characterized by abnormal structural and functional properties of brain networks. Structural and functional connectivity (SC and FC) based on magnetic resonance imaging (MRI) and their network metrics can be used to detect features of functional integration and segregation and identify certain patterns of local networks [1] . We aim to study the behaviour of connectivity matrices/metrics to recognize hypo or hyperconnected areas that could be associated with epileptogenic zones. Methods: This study was approved by the Institutional Review Board, and it comprised 10 patients (3 women, 7 men; mean age 31 ± 10 years; range 21-50 years) with refractory idiopathic epilepsy. In addition to a standard protocol, resting-state functional MRI (rs-fMRI) and diffusion-weighted images (DWI) in 64 directions were acquired on a 3.0 T Signa PET/MRI system (General Electric HC, Fig. 1 ). FC was obtained using the CONN toolbox (Matlab R2019b), to estimate seed-based connectivity measures. On the other hand, SC was quantified with probabilistic tractography based on the ball and stick model, using FSL Software. Network measures, including betweenness centrality (BC), cluster coefficient (CC), degree, local efficiency (LE), strength, and edges distance (ED), were calculated for SC and FC with the Brain Connectivity Toolbox. Significant differences in the connections and metrics between a single patient and a group of healthy controls (n = 70 [2] ) were found using a Crawford-Howell modified t test. Each p value was then transformed to its multiplicative inverse (h value), obtaining contrast matrices and metrics (Fig. 2 ). Clusters of 4 or more regions with significant differences (above a threshold of an h value) were identified and compared with the location of the epileptogenic zone according to the patient''s clinical diagnosis. Sensitivity and specificity were calculated for different thresholds (ROC curves), and the threshold cut-off value was quantified for each matrix and metric to achieve the best performance.Results: Results are summarised in Figs. 1 and 3. SC, FC and SC ? FC matrices showed high sensitivity and specificity values. The specificity of FC is higher than that of SC. The ED and CC metrics showed better performance than LE, BC, degree and strength.Discussion: The analysis of contrast FC and SC matrices, together with the ED and CC metrics, could be a possible method to implement in the evaluation of individual patients. Sensitivity and specificity values were high because the considered epileptogenic zones were up to 8 cm 2 ; in the future, we aim to add information provided by EEG and PET to reduce the areas up to 2 cm 2 . Introduction: T 2 mapping allows early detection of osteoarthritis(OA) in knee cartilage. OA is characterized by loss of collagen integrity and increased water content resulting in longer T 2 relaxation times [1] . Previous studies have reported mean population T 2 values for healthy and OA cartilage of 35 and 40 ms, respectively [2] . Due to time constraints, clinical T 2 mapping is usually performed with a multi-spin-echo (MSE) sequence and mono-exponential fitting.Recently, dictionary-based methods have enabled more accurate T 2 estimates by matching the MSE data to a dictionary of pre-computed echo modulation curves (EMC) [3] . Our aim is to compare different MSE protocols for knee cartilage T 2 mapping, varying the refocusing pulse as it impacts RF power tissue deposition (SAR). Protocols were applied to the ISMRM phantom and initial tests were performed in vivo. Methods: Several MSE sequences were acquired in the ISMRM phantom [4] (Fig. 1a) on a 3 T Philips Achieva, with refocusing flip angles (FA) of [90, 120:5:145, 180]8, 908 excitation pulse, echo train length = 10, FOV = 210 9 210 mm 2 , in-plane reconstructed resolution = 0.29 9 0.29 mm 2 and 2.5 mm slice thickness. An inter echo spacing of 5.8 ms (the smallest allowed by the software) was chosen to potentiate the detection of short T 2 values [5] . To create the EMCs, T 1 was fixed at 1000 ms (a prior simulation study revealed very low sensitivity to T 1 ); transmit field B 1? = 0.6:0.01:1.4 and T 2 = [1:0.2:10, 10:0.5:50, 50:1:100]ms values were considered [3] . T 2 maps were generated for each protocol with both methods (EMC and mono-exponential fit) and compared to the ISMRM phantom values. The best candidate protocol was applied in vivo in a preliminary validation test in an asymptomatic subject. Results: The T 2 map and mean values estimated with the EMC and mono-exponential fit methods for the different protocols in the phantom vials with T 2 values close to those previously reported for the knee cartilage are shown in Figs. 1b and 2. T 2 mapping of the cartilage is shown in Fig. 3 .Discussion: Higher accuracy of the EMC method was confirmed at both short and long T 2 values. For short T 2 values (vials 8, 9 and 13) the use of a 908 refocusing FA resulted in poorer accuracy, while for longer T 2 (vial 7) more accurate estimates were obtained for the tested protocols. By using a refocusing FA of at least 1208, accurate estimates could be obtained. However, to limit SAR, the use of a 120/1258 refocusing FA may be preferred in vivo. Regarding the in vivo preliminary test, the T 2 value of the cartilage agreed with previous studies, as the EMC fit (27 ± 13 ms) has lower values compared to the mono-exponential fit (38 ± 14 ms). Future work will be to evaluate the in vivo reproducibility.