key: cord-0257756-gstoid7n authors: Cali, R. J.; Freeman, H. J.; Billot, B.; Barra, M. E.; Fischer, D.; Sanders, W. R.; Huang, S. Y.; Conklin, J.; Fischl, B.; Iglesias, J. E.; Edlow, B. L. title: Synthesis of High-Resolution Research-Quality MRI Data from Clinical MRI Data in Patients with COVID-19 date: 2021-11-26 journal: nan DOI: 10.1101/2021.11.25.21266090 sha: c353dadc61d33b3a4b1adef5c462955f7a450b7e doc_id: 257756 cord_uid: gstoid7n Pathophysiological mechanisms of neurological disorders in patients with coronavirus disease 2019 (COVID-19) are poorly understood, partly because of a lack of high-resolution neuroimaging data. We applied SynthSR, a convolutional neural network that synthesizes high-resolution isotropic research-quality data from thick-slice clinical MRI data, to a cohort of 11 patients with severe COVID-19. SynthSR successfully synthesized T1-weighted MPRAGE data at 1 mm spatial resolution for all 11 patients, each of whom had at least one brain lesion. Correlations between volumetric measures derived from synthesized and acquired MPRAGE data were strong for the cortical grey matter, subcortical grey matter, brainstem, hippocampus, and hemispheric white matter (r=0.84 to 0.96, p[≤]0.001), but absent for the cerebellar white matter and corpus callosum (r=0.04 to 0.17, p>0.61). SynthSR creates an opportunity to quantitatively study clinical MRI scans and elucidate the pathophysiology of neurological disorders in patients with COVID-19, including those with focal lesions. Neurological manifestations of coronavirus disease 2019 (COVID-19) have been recognized since the initial days of the pandemic [1] . Recent international, multi-center studies suggest that up to 80% of hospitalized patients with COVID-19 experience neurological symptoms [2] , and up to 20% have altered levels of consciousness [3] . The pathogenesis of COVID-19 neurological disorders remains unclear, partly because few neuroimaging studies have been performed to investigate their associated alterations in brain structure and function [4] [5] [6] . Given the logistical challenge of enrolling patients with COVID-19 in neuroimaging studies, and the potential risk of research staff being exposed to severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the vast majority of MRI data worldwide have been acquired for clinical purposes [7] , with large slice thickness or inter-slice spacing (i.e., 5 to 6 mm) to decrease scan time (henceforth referred to as "clinical scans"). This lack of highresolution, research-quality MRI data has hindered efforts to elucidate the pathogenesis, prognosis, and natural history of recovery from COVID-19-related neurological disorders. We recently developed a joint super-resolution and synthesis method for generating research-quality 1 mm isotropic T1-weighted Magnetization Prepared Rapid Acquisition Gradient Echo (T1 MPRAGE) MRI data [8] . The method, termed SynthSR, involves training a convolutional neural network (CNN) with highly diverse synthetic images generated from segmentations, which makes the technique robust to changes in MR contrast, resolution, and acquisition direction of the clinical scans. SynthSR has not been tested in patients with COVID-19, nor has it been tested in patients with focal brain All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted November 26, 2021. ; https://doi.org/10.1101/2021.11.25.21266090 doi: medRxiv preprint lesions, which are common in patients with severe COVID-19 [7] . In this retrospective study of 11 patients with neurological disorders related to severe COVID-19, we were uniquely positioned to validate SynthSR, because all patients had a clinical MRI scan that included an isotropic research-quality T1-weighted image to compare the synthesized T1 MPRAGE against. To determine the feasibility of applying SynthSR to this patient population, we tested the hypothesis that quantitative cortical and subcortical volumetric measures derived from synthesized MPRAGE images strongly correlate (r≥0.75) with measures derived from acquired, "gold-standard" MPRAGE images. We retrospectively identified patients admitted to the Neurosciences Intensive Care Unit (ICU) at our institution during the initial COVID-19 surge on the East Coast of the United States (March to June 2020) and the second COVID-19 surge (December 2020 to January 2021). Patients were included if they met the following criteria: 1) positive PCR test for SARS-CoV-2; 2) clinical diagnosis of severe COVID-19 (i.e., respiratory failure requiring mechanical ventilation); 3) clinical brain MRI performed on the 3 Tesla Skyra MRI scanner (Siemens Healthineers, Erlangen, Germany) located in our Neurosciences ICU; 4) clinical brain MRI included T2-weighted (T2) fluid-attenuated inversion recovery (FLAIR), T2, or T1 sequences (at least two of these three sequences); 5) clinical brain MRI included 1 mm isotropic T1 MPRAGE or multi-echo MPRAGE (MEMPRAGE) to provide an acquired, "gold-standard" dataset. This MRI All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted November 26, 2021. ; https://doi.org/10.1101/2021.11.25.21266090 doi: medRxiv preprint scanner was designated for imaging patients with COVID-19 who were critically ill and therefore at high risk of having brain lesions [7] . Patients with clinical MRI performed on MRI scanners located elsewhere in the hospital were excluded to ensure consistency in the T1 MPRAGE/MEMPRAGE acquisitions and to facilitate testing SynthSR's performance characteristics in patients with brain lesions, which has not previously been assessed [8] . All acquired (clinical) sequences from the clinical MRI scans were evaluated for brain lesions by two board-certified neuroradiologists (S.Y.H. and J.C.). Initial disagreements were resolved by consensus review of the images. The purpose of these evaluations was to determine if SynthSR successfully synthesizes T1 MPRAGE data in patients with lesions without overfitting (generating anatomical data that is not actually there) or underfitting (failing to produce data that otherwise should be present). We also aimed to determine whether SynthSR's performance characteristics differ for patients with and without lesions, in terms of grey matter segmentation, white matter segmentation, and measurements of tissue volumes. To generate high-resolution synthetic images, we selected two clinical sequences with large inter-slice spacing for each subject (Supplementary Table 1 ). We prioritized T2 FLAIR and T1 sequences as input images for SynthSR ( Figure 1A ). In the absence of a T2 FLAIR or T1 sequence, we used a T2 sequence. The two clinical sequences were rigidly co-registered and entered into a 2-channel SynthSR model to All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted November 26, 2021. ; generate a single, high-resolution synthesized T1 MPRAGE image for each subject. The high-resolution synthesized and acquired MPRAGEs were then processed using 'recon-all', a FreeSurfer function (https://surfer.nmr.mgh.harvard.edu) [9] , to generate a full-brain segmentation and volumetric statistics for both datasets ( Figure 1B) . We adapted a previously published QA scale [10] to rate the anatomic accuracy of cortical surfaces and subcortical segmentations of the acquired and synthesized T1 Table 2 inaccurate synthetic anatomy. The rater was thus not blinded to the identity of a dataset as acquired or synthesized, but the rater was blinded to the lesion ratings performed by the neuroradiologists and the cortical volumetric data. We used Spearman's Correlation Coefficient to test for associations between the acquired and synthesized volumetric measurements in 11 prespecified regions of interest. Correlations were considered significant at a Bonferroni-corrected p value <0.0045 (0.05/11). All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. SynthSR successfully generated synthesized T1 MPRAGE data for all 11 patients. QA ratings revealed that for 7 of 11 patients, the quality of the FreeSurfer segmentations for the acquired and synthesized data was similar (Supplementary Table 3 ). For the other four patients, there were variable differences in quality of the FreeSurfer segmentations, with the acquired data quality outperforming the synthesized data quality in two, and synthesized data quality outperforming acquired data quality in two. Where differences were observed, the synthesized datasets led to more errors in subcortical segmentation, whereas the acquired datasets contained more errors in the cortical surfaces. One error unique to the synthesized data was observed in patient 2, All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Volumetric results for all 11 regions are provided in Table 2 . Given that all 11 patients had brain lesions, it was not possible to determine if the correlation coefficients differed between patients with and without lesions. Using a newly developed synthesis method based on a convolutional neural network (SynthSR), we demonstrate that high-resolution research-quality MRI data can be synthesized from clinical MRI data in patients with COVID-19. In eight of 11 prespecified anatomic regions, the quantitative volumetric measurements derived from the synthesized MPRAGE dataset strongly correlated with those derived from the acquired MPRAGE dataset, suggesting that synthesized MPRAGE data can be used to investigate clinical-anatomic correlations in future studies of COVID-19-related neurological disorders. Crucially, SynthSR successfully generated 1 mm isotropic MPRAGE datasets in the presence of hemorrhagic or ischemic lesions in all 11 patients, suggesting that SynthSR may be generalizable to patients with COVID-19 who have a broad spectrum of brain lesions [7] . The potential applications of SynthSR to COVID-19-related neurological disorders are myriad. Clinical MRI scans that were previously only amenable to qualitative analysis can now be tested for associations between cortical and subcortical volumetric measures and clinical syndromes. In future work, SynthSR could also be used in patients with multiple clinical MRI scans to assess for longitudinal volumetric changes that are associated with development of chronic "long-COVID" neurological symptoms. By combining SynthSR with lesion-mapping techniques [10] , it may also be possible to identify the neuroanatomic distribution of COVID-19-related brain lesions and their associated neurological syndromes. Importantly, for SynthSR to reach its full potential, its performance characteristics will need to be improved in subcortical regions such as the cerebellar white matter and cerebellar cortex. In summary, the SynthSR tool, which we distribute to the academic community at github.com/BBillot/SynthSR and https://surfer.nmr.mgh.harvard.edu/fswiki/SynthSR, can be used to support international collaboration in the study of COVID-19-related neurological disorders. Synthesis of high-resolution images from clinical MRI data creates an opportunity for quantitative volumetric analysis of clinical MRI scans acquired in patients with COVID-19 that otherwise would have insufficient resolution for such analysis. All rights reserved. No reuse allowed without permission. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted November 26, 2021. ; https://doi.org/10.1101/2021.11.25.21266090 doi: medRxiv preprint preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted November 26, 2021. ; preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this this version posted November 26, 2021. ; https://doi.org/10.1101/2021.11.25.21266090 doi: medRxiv preprint (Table 2) , this qualitative visual comparison of the acquired and synthesized datasets reveals similarities in several anatomic regions (e.g., subcortical grey matter, brainstem, cortical grey matter, and hemispheric white matter), and differences in other regions (e.g., cerebellar white matter). Neurologic Manifestations of Hospitalized Patients With Coronavirus Disease Global Incidence of Neurological Manifestations Among Patients Hospitalized With COVID-19-A Report for the GCS-NeuroCOVID Consortium and the ENERGY Consortium Neurologic manifestations in hospitalized patients with COVID-19: The ALBACOVID registry Intact Brain Network Function in an Unresponsive Patient with COVID-19 Neuroanatomical substrates of generalized brain dysfunction in COVID-19 Neuroprognostication of Consciousness Recovery in a Patient with COVID-19 Related Encephalitis: Preliminary Findings from a Multimodal Approach Common Data Elements for COVID-19 Neuroimaging: A GCS-NeuroCOVID Proposal Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast Optimizing the accuracy of cortical volumetric analysis in traumatic brain injury. MethodsX