key: cord-267055-xscwk74r authors: Chassagnon, Guillaume; Vakalopoulou, Maria; Battistella, Enzo; Christodoulidis, Stergios; Hoang-Thi, Trieu-Nghi; Dangeard, Severine; Deutsch, Eric; Andre, Fabrice; Guillo, Enora; Halm, Nara; Hajj, Stefany El; Bompard, Florian; Neveu, Sophie; Hani, Chahinez; Saab, Ines; Campredon, Aliénor; Koulakian, Hasmik; Bennani, Souhail; Freche, Gael; Barat, Maxime; Lombard, Aurelien; Fournier, Laure; Monnier, Hippolyte; Grand, Téodor; Gregory, Jules; Nguyen, Yann; Khalil, Antoine; Mahdjoub, Elyas; Brillet, Pierre-Yves; Tran Ba, Stéphane; Bousson, Valérie; Mekki, Ahmed; Carlier, Robert-Yves; Revel, Marie-Pierre; Paragios, Nikos title: AI-Driven quantification, staging and outcome prediction of COVID-19 pneumonia date: 2020-10-15 journal: Med Image Anal DOI: 10.1016/j.media.2020.101860 sha: doc_id: 267055 cord_uid: xscwk74r Coronavirus disease 2019 (COVID-19) emerged in 2019 and disseminated around the world rapidly. Computed tomography (CT) imaging has been proven to be an important tool for screening, disease quantification and staging. The latter is of extreme importance for organizational anticipation (availability of intensive care unit beds, patient management planning) as well as to accelerate drug development through rapid, reproducible and quantified assessment of treatment response. Even if currently there are no specific guidelines for the staging of the patients, CT together with some clinical and biological biomarkers are used. In this study, we collected a multi-center cohort and we investigated the use of medical imaging and artificial intelligence for disease quantification, staging and outcome prediction. Our approach relies on automatic deep learning-based disease quantification using an ensemble of architectures, and a data-driven consensus for the staging and outcome prediction of the patients fusing imaging biomarkers with clinical and biological attributes. Highly promising results on multiple external/independent evaluation cohorts as well as comparisons with expert human readers demonstrate the potentials of our approach. • A Covid-19-specific holistic, highly compact multi-omics signature integrating imaging/clinical/ biological data and associated comorbidities for automatic patient staging is presented and evaluated. • Short and Long-term prognosis for clinical resources optimization offering alternative/complementary means to facilitate triage for Covid-19 • Clinically-relevant quantification and staging tool validated by comparison with clinical experts is reported. Coronavirus disease 2019 (COVID-19) emerged in 2019 and disseminated around the world rapidly. Computed tomography (CT) imaging has been proven to be an important tool for screening, disease quantification and staging. The latter is of extreme importance for organizational anticipation (availability of intensive care unit beds, patient management planning) as well as to accelerate drug development through rapid, reproducible and quantified assessment of treatment response. Even if currently there are no specific guidelines for the staging of the patients, CT together with some clinical and biological biomarkers are used. In this study, we collected a multi-center cohort and we investigated the use of medical imaging and artificial intelligence for disease quantification, staging and outcome prediction. Our approach relies on automatic deep learning-based disease quantification using an ensemble of architectures, and a datadriven consensus for the staging and outcome prediction of the patients fusing imaging biomarkers with clinical and biological attributes. Highly promising results on multiple external/independent evaluation cohorts as well as comparisons with expert human readers demonstrate the potentials of our approach. c 2020 Elsevier B. V. All rights reserved. Preprint submitted to Medical Image Analysis October 13, 2020 China (Zhu et al., 2020) caused by the SARS-Cov-2 virus, and it could lead to respiratory failure due to severe viral pneumonia . The disease spread worldwide leading the World Health Organization to declare it as a pandemic in March 2020. One of the important actions to handle the pandemic is the fast and robust use of imaging along with clinical and biological comorbidities for the quantification and staging of patients upon their hospital admission. Being able to identify patients that need intubation upon admission is very important and essential for the management of a hospital's resources and the most optimal management of patients. Moreover, a robust staging of the patients could also facilitate proper selection of patients for different treatments, reducing the unnecessary use of the hospital's intensive care units. To the best of our knowledge, currently the staging of the patients is mainly based on clinical and biological biomarkers such as age, sex and other comorbidities Li et al., 2020a; Yuan et al., 2020; Tang et al., 2020; Onder et al., 2020; Guo et al., 2020; Terpos et al., 2020) , while the role of imaging is mainly focusing on an estimation of the disease extent from CT scans. This estimation is mainly done by medical experts and hence suffers from inter-and intra-observer variability. In this study, we investigated an automatic method ( To the best of our knowledge this is among a few systematic efforts to quantify disease extent, to discover low dimensional and interpretable imaging biomarkers and to integrate them to clinical variables into short and long term prognosis of COVID-19 patients. The paper is organized as follows: we first review related work mainly focusing on interstitial lung diseases (ILDs) diseases, which is followed by a description of all the components and implementation details of our method. We then present the acquired multi-center dataset, the evaluation setting, and the results of our experiments. Furthermore, we discuss in detail similarities and differences of our method with other recently proposed methods for quantification and staging of COVID-19. Lastly, we present possible directions for future research. In this section, we provide a short review of previous studies on quantification of ILDs since COVID-19 and ILDs share a lot of similarities due to their diffuse pathological manifestations, such as ground glass opacities, band consolidations, and reticulations. Furthermore, we elaborate on studies that tackle severity or treatment response for such types of disease. There are numerous studies proposed the last years on automatic quantification of ILD diseases using CT scans. The main goal of these studies is to develop models that are able to identify one or more types of different pathological lung tissue in ILD cases (such as ground glass, consolidation, honeycombing, etc) and successfully separate them from the healthy tissue. Initial efforts were mainly based on classification schemes. In particular, small patches including only a single tissue type were extracted and described using a number of handcrafted features focusing mainly on texture, then these features were used to train different machine learning classifiers (Gangeh et al., 2010; Huber et al., 2012) . Following recent advances in deep learning and especially the success of convolutional neural networks (CNNs), researchers have recently employed such tools also in thoracic imaging tasks (Chassagnon et al., 2019) , with ILD quantification being among them. The main advantage of CNNs is their ability to generate features automatically from the input, and create meaningful representations for the studied per time problems. In particular, a patch-based framework using a convolutional architecture is presented in for the automatic quantification of 5 different ILD patterns. Similarly, in (Gao et al., 2018) a patch-based approach is adapted to classify them in 6 different ILD patterns. Even if the method reported higher performance than other methods based on handcrafted features, the use of patches, besides being time consuming and ineffi-cient, does not exploit the texture of the entire lung. Many of the already proposed CNNs have further been adapted to perform the task of semantic segmentation in an endto-end fashion instead of only image classification. Semantic segmentation refers to the task of infering a class for each of the pixels of an image instead of a single class for an image. Such models can be found in literature both in 2D (Badrinarayanan et al., 2017; Ronneberger et al., 2015) and 3D (Ç içek et al., 2016) and have also been used for ILD quantification. The authors of present the coupling of 2D fully convolutional networks with deformable registration for the automatic quantification of systemic sclerosis disease. Moreover, in (Anthimopoulos et al., 2018) Staging of patients with ILDs is very important as it could greatly help clinicians with their daily practice, while choosing treatment options (Kolb and Collard, 2014) . There have been a number of studies recently that try to identify and extract biomarkers from CT scans and associate them with the severity and treatment of ILD patients. These biomarkers are usually enhanced with clinical and physiological information to provide a scoring system as survival predictor. Among the variety of biomarkers, disease extent is one of the most powerful ones providing strong associations with severity and mortality (Cottin and Brown, 2019; Tomassetti et al., 2015) . Visual scoring of the disease extent on CT can be time-consuming (Robbie et al., 2017) highlighting the need for tools for automatic disease quantification. Moreover, except the disease extent, the location of the disease is also very important for the staging. In (Depeursinge et al., 2015; Christe et al., 2019) the quantification of the disease is performed on different lung regions providing descriptive information about the severity of the ILD patients. A variety of works report that radiomics, quantitative features extracted from the images, provide valuable information about the severity and response to treatment for different diseases including cancer (Sun et al., 2018) . These features could also provide very good tools for monitoring disease progression and therapeutic response (Wu et al., 2019) . In particular, in (Bocchino et al., 2019) intensity-based characteristics such as skewness and kurtosis were used together with disease extent to distinguish between systemic sclerosis patients with and without ILD diseases. Moreover, in (Lafata et al., 2019) a variety of image radiomics and their relationship with the pulmonary function were investigated. Their results indicate that high-throughput radiomics data extracted from the lungs may be associated with pulmonary function as measured by common PFT metrics. In this section, we describe our AI driven scheme for the quantification of CT scans for patients suffering from COVID-19 pneumonia. Furthermore, we provide a method for the automatic selection and combination of multi-modal variables towards a holistic signature designed for the COVID-19 triage. On the basis of this interpretable, clinically relevant signature we develop advanced machine learning techniques integrating multi-modal data for severity assessment and short/long term outcome prediction. Our method endows robustness, good generalization properties, explainability and establishes causality with known clinical COVID-19 confounding factors. In the following parts of this section, we provide details for all the different components of the system. Segmentation of the heart and breast were extracted by using the software ART-Plan (TheraPanacea, Paris, France). ART-Plan is a CE-marked solution for automatic annotation of organs, harnessing a combination of anatomically preserving and deep learning concepts. The segmentation of lungs was also performed using ART-Plan software, but the models used were re-trained using COVID-19 patients in order to address proper segmentation of diseased lungs. In particular, the existing lung models, providing segmentation of left and right lungs, were re- For the registration of the CT scans to the templates, an elastic registration framework based on Markov Random Fields was used, providing the optimal displacements for each template (Ferrante et al., 2017) . In particular, the registration is performed by a non-linear transformer T , corresponding to the operator that optimizes in the continuous domain Ω the following energy, where ρ j corresponds to the different similarity metrics (sum of absolute difference, normalised cross correlation, etc) used to compare the source 3D volume to the target anatomy, w j are linear constraints factorizing the importance of the different metric functions and ψp¨q is a penalty function acting on the spatial derivatives of the transformation. Concerning the details of the architecture, in our experiments each C i consists of a SegNet (Badrinarayanan et al., 2017) based architecture. More specifically, for the CovidE2D models the CT scans were separated on the axial view. Each network included 5 convolutional blocks, each one containing two Conv-BN-ReLU layer successions. Maxpooling layers were also distributed at the end of each convolutional block for the encoding part. Upsampling operators were used on the decoding part to restore the spatial resolution of the slices together with the same successions of layers. To fully exploit the 3D nature of our dataset, the second component of our proposed CovidENet is based on a 3D fully convolutional network similar to 3D-UNet (Ç içek et al., 2016) . In order to train this model, 3D sub-volumes of the CT scan that fully included without any downsampling either the left or right lung were extracted. The corresponding sub-volumes were also extracted from the ground truth annotation masks. To this end, we trained the model with the CT scan sub-volume as input and the annotation as target. As far as the architecture is concerned, the model consisted of five blocks with a down-sampling operation applied every two consequent Conv3D-BN-ReLU layers. Additionally, five decoding blocks were utilized for the decoding path, were at each block a transpose convolution was performed in order to up-sample the input. Skip connections were also employed between the encoding and decoding paths. The dimensions of the input that corresponded to the spatial dimensions of the CT scan and consequently the spatial dimensions of the features maps were not bound to some fixed dimension in order to feed the entire left/ right lung volumes. As such, 3D volumes of arbitrary spatial dimensions could be fed to the network and thus the batch size was fixed to 1. In order to combine disease extent with disease characteristics and patients commodities, we investigate a variety of imaging characteristics extracted using disease, cardiac and lung segmentations. These imaging characteristics (radiomics) were then combined with meaningful clinical and biological indicators that have been reported to be associated prognosis of COVID-19. Patient charts were reviewed to assess short term (4 days after the chest CT) and long term prognosis (31 days after the chest CT). For the staging task, patients were divided in 2 groups: those who died, or required mechanical ventilation either at the initial or at a subsequent admission as severe cases (S), and the rest as non-severe cases (NS). For the prognosis task, three distinct subpopulations were defined: those who had a short term negative (SD = short-term deceased) outcome (deceased within 4 days after admission), those who didn't re- cover (LD= long-term deceased) within 31 days after the chest CT (either died after day 4 or still intubated at day 31) and those who recovered (LR= long-term recovered). The last two groups formed the short intubated (SI) group of patients. Radiomics features were extracted from the CT scans using the previously described segmentations of the disease, lung and heart. As a preprocessing step, all images were resampled by cubic interpolation to obtain isometric voxels with sizes of 1mm. Subsequently, disease, lung and heart masks were used Using all the calculated attributes (clinical, biological, imaging) we constructed a high dimensional space of size 543, including clinical/biological variables. A min-max normalization of the attributes was performed by calculating the minimum and maximum values for the training and validation cohorts. The same values were also applied on the test set. To prevent overfitting and discover the most informative and robust attributes for the staging and prognosis of the patients we propose a robust biomarker selection process. Feature selection is very important for classification tasks and has been used widely in literature especially for radiomics (Sun et al., 2018) . First, the training data set was subdivided into training and validation on the principle of 80%-20% maintaining the distribution of classes between the two subsets identical to the observed one. To perform features selection, we have created 100 subdivisions on this basis and evaluated variety of classical machine learning -using the entire feature space -classifiers such as Regarding imaging features, we identified the following features as more important for the prognosis of the COVID-19 patients. These features include both first and second order statistics together with some shape features. Length on the Axis. The selected disease area features capture both disease extent and disease textural heterogeneity. Disease textural heterogeneity is associated with lesions, the presence of which generates imaging pattern more complex than pure ground glass opacities usually found in mild disease. The selected lung features capture the dispersion and heterogeneity of lung densities, both of which may reflect the presence of an underlying airway disease such as emphysema but also the presence of sub-radiological disease. Lastly, the selected heart features can be seen as a surrogate for cardiomegaly and coronary calcifications. The staging/prognosis component was addressed using an ensemble learning approach. Similarly to the biomarker extraction, the training data set was subdivided into training and validation sets on the principle of 80%-20%. This subdivision was performed such that the distribution of classes between the two subsets was identical to the observed one. We have used 10-fold cross validation on this basis and evaluated the average performance of the following supervised clas- To perform the short-term deceased (SD), long-term Deceased (LD), long term recovered (LR) classification task, a SD/SI (SI: intubated at 4 days) classifier and a LD/LR clas-sifier was applied in a hierarchical way, performing first the short-term staging and then the long-term prognosis for patients classified as in need of mechanical ventilation support. More specifically, a majority voting method was applied to classify patients into SD and SI cases. Then, another hierarchical structure was applied on the cases predicted as SI only to classify them into the ones who didn't recover within 31`days of mechanical ventilation (LD) and the ones who recovered with 30 days on mechanical ventilation (LR). In order to train all the models, each CT scan was normalized by cropping the Hounsfield units in the range r´1024, 300s. A variety of hyperparameters including loss functions, learning rates, optimizers had been tested and in this section we report the ones with the best performance for each compo- This retrospective multi-center study was approved by our In- (Table 3 ). In addition to the CT examination -when available -patient sex, age, and body mass index (BMI), blood pressure and diabetes, lymphocyte count, CRP level and D-dimer level were also collected (Table 3) . For short-term outcome assessment, patients were divided into 2 groups: those who died or were intubated in the 4 days following the CT scan composed the severe short-term outcome subgroup, while the others composed the non-severe short-term outcome subgroup. For long-term outcome, medical records were reviewed from May 7th to May 10th, 2020 to determine if patients died or had been intubated during the period of at least one month following the CT examination. The data associated with each patient (holistic profiling), as well as the corresponding outcomes both in terms of severity assessment as well as in terms of final outcome and readers assessment will be made publicly available. imaging (Reader C ) were asked to perform a triage (severe versus non-severe cases) and for the severe cases (short-term deceased versus short-term intubated) prognosis process to predict the short-term outcome. The dice similarity score (DSC) was calculated to assess the similarity between the 2 manual segmentations of each CT exam of the test dataset and between manual and automated segmentations. The Hausdorff distance (HD) was also calcu- For the stratification of the dataset into the different categories, classic machine learning metrics, namely balanced accuracy, weighted precision, and weighted specificity and sensitivity were utilized. The evaluation of CovidENet together with its components and the comparison with the 2 independent experts is summarised in Table 4 p= 0.352). As shown in Figure 5 correlation to disease extent from manual segmentations was better when using Covi-dENet (r " 0.94, p ă 0.001) compared to Covid3D (r " 0.71, p ă 0.001) or Covid2D (r " 0.92, p ă 0.001) which oversegmented the disease. Examples of disease segmentations are presented in Figure 6 . One can observe that the segmentations provided by CovidENet are very close to the ones generated by the experts. In particular, the algorithm detects the diseased regions even in the case that they are relatively small capturing all the different opacities of COVID-19 such as ground glass and consolidation. The holistic COVID-19 pneumonia signature is presented in (Table 5) S/NS outcomes had a balanced accuracy of 70% (vs 67% for human readers consensus), a weighted precision of 81% (vs 78%), a weighted sensitivity of 64% (vs 70%) and specificity of 77% (vs 64%) and outperformed the consensus of human readers (Figure 7 , Table 6 ). Our method successfully predicted 81% of the severe/critical cases opposed to only 61% for the consensus reader. The superiorty of our approach is also indicated by the higher AUC reported (0.76), in comparison with the one achieved by the different readers (0.69). Severe cases as depicted in Figure 7 referred to diabetic men, with higher level of volume/heterogeneity of disease and C-reactive protein levels. Moreover, as indicated in Figure 7 the non-uniformity on GLRLM for both lung and disease together with the disease extent seems to contribute considerable to the classification of the patients to NS versus S cases. The COVID-19 pneumonia pandemic spiked hospitalizations, while exerting extreme pressure on intensive care units. In the absence of a cure, staging and prognosis is crucial for clinical decision-making for resource management and experimental outcome assessment, in a pandemic context. Our objective was to predict patient outcomes prior to mechanical ventilation support. The proposed ensemble classifier aiming to predict the SD/(LD or LR) had a balanced accuracy of 88% (vs 81% for human readers consensus), a weighted precision of 94% (vs 87%), a weighted sensitivity of 94% (vs 88%) and specificity of 81% (vs 75%) and outperformed consensus of human readers (Table 6 ). Our method for prognosis of SD/ LD/ LR had a balanced accuracy of 71%, a weighted precision of 77%, a weighted sensitivity of 74% and specificity of 82% to provide full prognosis (Figure 8 ). Concerning the performance of our method for the classification of LD and LR patients (Ta-ble 7), our ensemble classifier reports a balance accuracy of 69%, a weighted precision of 76% a weighted sensitivity of 74% and a weighted specificity of 65%. As indicated also in Figure 8 the performance of our method reach an AUC of 0.86 for the SD, a 0.86 for the LR and 0.76 for the LD classes. Moreover, the age, HBP and lung non uniformity on the GLSZM seems to associate better for this task. Moreover, in order to assess the impact of each feature category on the implemented models we performed an ablation study by successively removing one category of features from the 6 categories defined for each classification task. Results are presented in Table 8 . The feature categories were identified as follows: a) D0: disease extent, b) D1: disease variables that are shape/geometry related, c) D2: disease variables that are tissue/texture, d) O1: heart/lungs variables that are shape/geometry related, e) O2: heart/lungs variables that are tissue/texture, f) B1: age, gender, biologi- cal/obesity/diabetes/fat/high blood pressure. One can observe that the Clinical Only category contributes a lot for the separation of SD/LD/LR while for the NS/S cases their contribution is marginal, in contrary to the other imaging characteristics. AI-enhanced imaging, clinical and biological information proved capable to identify patients with severe short/long-term outcomes, bolstering healthcare resources under the extreme pressure of the current COVID-19 pandemic. The information obtained from our AI staging and prognosis could be used as an additional element at admission to assist decision making. Variety of studies have reported the use of deep learning for the diagnosis and quantification of COVID-19 with CT scans. In particular, studies have already reported on deep learning diagnosing COVID-19 pneumonia on chest CTs. In (Li et al., 2020b ) the authors proposed the use of a deep learning architec-ture based on ResNet50 for the diagnosis of COVID-19 reporting very high performances, while they investigated the attention maps produced from their network. A very similar method is presented in (Mei et al., 2020) reporting the use of deep learning on COVID-19 diagnosis. Moreover, in the authors propose the use of a UNet architecture for the quan- Table 6 . ity to regress the proposed scores. Finally, recently (Tilborghs et al., 2020) presents a comparable study of deep learning based methods for the automatic quantification of COVID-19. Assessing the severity of COVID-19 patients is also a very quickly evolving topic in the medical community with some methods being currently under review. Extracting valuable information from the imaging using recent advances is very important and could potentially facilitate the clinical practice. Starting with, disease extent is known to be associated with severity (Li et al., 2020a; Yuan et al., 2020) erogeneous lesions than pure ground glass opacities observable in mild cases. In (Li et al., 2020c) , the authors proposed the use of Siamese networks for the severity assessment of COVID-19 directly from CT scans. In (Bai et al., 2020) , the authors proposed a deep learning pipeline based on LSTMs using 2D CT slices and a fusion of imaging and clinical information to as- (Lassau et al., 2020) proposed the assessment of severity using a deep learning tool achieving an AUC of 0.79 on a completely independent cohort, with low sensitivity. Again, even if we can not perform a direct comparison our method reports similar performance in a completely independent cohort, while it is based on interpretable features extracted from different regions. Finally, in (He et al., 2020) a 2D deep learning based approach using a multi-task learning is presented in order to separate COVID-19 patients to severe and non severe cases. To the best of our knowledge this study is the first to have developed a robust, holistic COVID-19 multi-omics signature for disease staging and prognosis demonstrating an equivalent/superior-to-human-reader performance on a multicentric data set. Our approach complied appropriate data collection and methodological testing requirements beyond the existing literature (Mei et al., 2020 heterogeneity of lung densities, reflecting the presence of an underlying airway disease such as emphysema and the presence of sub-radiological disease. Among clinical variables, a higher CRP level, lymphopenia and a higher prevalence of hypertension and diabetes were associated with a poorer outcome, consistent with previous reports Guo et al., 2020; Terpos et al., 2020) . Interestingly, age was less predictive of disease severity than of poor outcome in severe patients. This is linked to the fewer therapeutic possibilities for these generally more fragile patients. Lastly, the average body mass index (BMI) in both non-severe and severe groups corresponded to overweight. Despite being correlated with BMI, the fat ratio measured on the CT scanner was only weakly associated with outcome. Several studies have reported obesity to be associated with severe outcomes Chaganti et al., 2020) and an editorial described the measurement of anthropometric characteristics as crucial to better estimate the risk of complications (Stefan et al., 2020) . However a meta-analysis showed that whereas being associated with an increased risk of COVID-19 pneumonia, obesity was paradoxically associated with reduced pneumonia mortality (Wynants et al., 2020) . Overall, the combination of clinical, biological and imaging features demonstrates their complementary value for staging and prognosis. In conclusion, we show that the combination of chest CT and artificial intelligence can provide tools for fast, accurate and precise disease extent quantification as well as the identifica- Tensorflow: Large-scale machine learning on heterogeneous distributed systems Lung pattern classification for interstitial lung diseases using a deep convolutional neural network Semantic segmentation of pathological lung tissue with dilated fully convolutional networks Segnet: A deep convolutional encoder-decoder architecture for image segmentation Classification of interstitial lung abnormality patterns with an ensemble of deep convolutional neural networks Performance of a new quantitative computed tomography index for interstitial lung disease assessment in systemic sclerosis Quantification of tomographic patterns associated with covid-19 from chest ct Deep learning: definition and perspectives for thoracic imaging Computer-aided diagnosis of pulmonary fibrosis using deep learning and ct images Interstitial lung disease associated with systemic sclerosis (ssc-ild) Automated classification of usual interstitial pneumonia using regional volumetric texture analysis in high-resolution ct Deformable registration through learning of context-specific metric aggregation A texton-based approach for the classification of lung parenchyma in ct images Holistic classification of ct attenuation patterns for interstitial lung diseases via deep convolutional neural networks Diabetes is a risk factor for the progression and prognosis of covid-19 Synergistic learning of lung lobe segmentation and hierarchical multi-instance classification for automated severity assessment of covid-19 in ct images Serial quantitative chest ct assessment of covid-19: Deep-learning approach. Radiology: Cardiothoracic Imaging 2 Texture feature ranking with relevance learning to classify interstitial lung disease patterns Staging of idiopathic pulmonary fibrosis: past, present and future An exploratory radiomics approach to quantifying pulmonary function in ct images Aibased multi-modal integration of clinical characteristics, lab tests and chest cts improves covid-19 outcome prediction of hospitalized patients Ct image visual quantitative evaluation and clinical classification of coronavirus disease (covid-19) Artificial intelligence distinguishes covid-19 from community acquired pneumonia on chest ct Automated assessment of covid-19 pulmonary disease severity on chest radiographs using convolutional siamese neural networks Artificial intelligence-enabled rapid diagnosis of patients with covid-19 Case-fatality rate and characteristics of patients dying in relation to covid-19 in italy Scikitlearn: Machine learning in python Evaluating disease severity in idiopathic pulmonary fibrosis U-net: Convolutional networks for biomedical image segmentation Obesity and impaired metabolic health in patients with covid-19 A radiomics approach to assess tumour-infiltrating cd8 cells and response to anti-pd-1 or anti-pd-l1 immunotherapy: an imaging biomarker, retrospective multicohort study Abnormal coagulation parameters are associated with poor prognosis in patients with novel coronavirus pneumonia Hematological findings and complications of covid-19 Comparative study of deep learning methods for the automatic segmentation of lung, lesion and lesion type in ct scans of covid-19 patients Staging systems and disease severity assessment in interstitial lung diseases. Current opinion in pulmonary medicine Atlasnet: Multi-atlas non-linear deep networks for medical image segmentation Computational radiomics system to decode the radiographic phenotype Computed tomographic biomarkers in idiopathic pulmonary fibrosis. the future of quantitative analysis Prediction models for diagnosis and prognosis of covid-19 infection: systematic review and critical appraisal Association of radiologic findings with mortality of patients infected with 2019 novel coronavirus in wuhan, china Clinical course and risk factors for mortality of adult inpatients with covid-19 in wuhan, china: a retrospective cohort study A novel coronavirus from patients with pneumonia in china Writing -Review & Editing, Eric Deutsch : Validation, Writing -Review & Editing, Fabrice Andre: Validation, Writing -Review & Editing, Enora Guillo: Data Curation, Writing -Review & Editing, Nara Halm: Data Curation, Writing -Review & Editing, Stefany El Hajj: Data Curation We thank Mihir Sahasrabudhe and Norbert Bus for their valu- The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.