Multimodal MRI-Based Whole-Brain Assessment in Patients In Anoxoischemic Coma by Using 3D Convolutional Neural Networks.

BACKGROUND: There is an unfulfilled need to find the best way to automatically capture, analyze, organize, and merge structural and functional brain magnetic resonance imaging (MRI) data to ultimately extract relevant signals that can assist the medical decision process at the bedside of patients in postanoxic coma. We aimed to develop and validate a deep learning model to leverage multimodal 3D MRI whole-brain times series for an early evaluation of brain damages related to anoxoischemic coma. METHODS: This proof-of-concept, prospective, cohort study was undertaken at the intensive care unit affiliated with the University Hospital (Toulouse, France), between March 2018 and May 2020. All patients were scanned in coma state at least 2 days (4 ± 2 days) after cardiac arrest. Over the same period, age-matched healthy volunteers were recruited and included. Brain MRI quantification encompassed both "functional data" from regions of interest (precuneus and posterior cingulate cortex) with whole-brain functional connectivity analysis and "structural data" (gray matter volume, T1-weighted, fractional anisotropy, and mean diffusivity). A specifically designed 3D convolutional neuronal network (CNN) was created to allow conscious state discrimination (coma vs. controls) by using raw MRI indices as the input. A voxel-wise visualization method based on the study of convolutional filters was applied to support CNN outcome. The Ethics Committee of the University Teaching Hospital of Toulouse, France (2018-A31) approved the study and informed consent was obtained from all participants. RESULTS: The final cohort consisted of 29 patients in postanoxic coma and 34 healthy volunteers. Coma patients were successfully discerned from controls by using 3D CNN in combination with different MR indices. The best accuracy was achieved by functional MRI data, in particular with resting-state functional MRI of the posterior cingulate cortex, with an accuracy of 0.96 (range 0.94-0.98) on the test set from 10-time repeated tenfold cross-validation. Even more satisfactory performances were achieved through the majority voting strategy, which was able to compensate for mistakes from single MR indices. Visualization maps allowed us to identify the most relevant regions for each MRI index, notably regions previously described as possibly being involved in consciousness emergence. Interestingly, a posteriori analysis of misclassified patients indicated that they may present some common functional MRI traits with controls, which suggests further favorable outcomes. CONCLUSIONS: A fully automated identification of clinically relevant signals from complex multimodal neuroimaging data is a major research topic that may bring a radical paradigm shift in the neuroprognostication of patients with severe brain injury. We report for the first time a successful discrimination between patients in postanoxic coma patients from people serving as controls by using 3D CNN whole-brain structural and functional MRI data. Clinical Trial Number http://ClinicalTrials.gov (No. NCT03482115).

Investigating the discrimination ability of 3D convolutional neural networks applied to altered brain MRI parametric maps.

Convolutional neural networks (CNNs) are gradually being recognized in the neuroimaging community as a powerful tool for image analysis. Despite their outstanding performances, some aspects of CNN functioning are still not fully understood by human operators. We postulated that the interpretability of CNNs applied to neuroimaging data could be improved by investigating their behavior when they are fed data with known characteristics. We analyzed the ability of 3D CNNs to discriminate between original and altered whole-brain parametric maps derived from diffusion-weighted magnetic resonance imaging. The alteration consisted in linearly changing the voxel intensity of either one (monoregion) or two (biregion) anatomical regions in each brain volume, but without mimicking any neuropathology. Performing ten-fold cross-validation and using a hold-out set for testing, we assessed the CNNs' discrimination ability according to the intensity of the altered regions, comparing the latter's size and relative position. Monoregion CNNs showed that the larger the modified region, the smaller the intensity increase needed to achieve good performances. Biregion CNNs systematically outperformed monoregion CNNs, but could only detect one of the two target regions when tested on the corresponding monoregion images. Exploiting prior information on training data allowed for a better understanding of CNN behavior, especially when altered regions were combined. This can inform about the complexity of CNN pattern retrieval and elucidate misclassified examples, particularly relevant for pathological data. The proposed analytical approach may serve to gain insights into CNN behavior and guide the design of enhanced detection systems exploiting our prior knowledge.