Méthodes de réseaux de neurones sur graphes pour la classification des formes cliniques de sclérose en plaques et l'estimation du handicap

Deep Learning (DL), an important branch of Artificial Intelligence (AI), has recently gained interest with the quick development of computing power of hardware, leading to very large applications in all domains of our society from social networks, autonomous vehicles, to nature language processing. Analysis of medical images constitutes a major application of deep learning based on Artificial Neural Networks (ANN), assisting physicians with accurate diagnosis and disease follow-up. More recently, Graph Neural Networks (GNN) have shown interesting capabilities for analyzing non-Euclidean data structures and small datasets as found in medical imaging. Multiple Sclerosis (MS) is a chronic inflammatory and neurodegenerative disease of the central nervous system leading to progressive physical and/or cognitive disability. MS patients undergo different clinical courses, 85% of them start from relapsing-remitting followed by secondary progressive phases, while 15% of the patients start from the onset by primary progressive phase of the disease. Even though there is no cure for MS, early intervention can reduce symptoms, prevent relapses and hence significantly improve patients’ quality of life. Therefore, identifying the clinical form and the disease progression is a helpful clinical task. Magnetic resonance imaging (MRI) is an essential examination tool for MS patients’ diagnosis and follow-up. MRI is a non-invasive diagnostic technique that provides high-contrast anatomical images of the brain, enabling physicians to detect the main pathological signs such as inflammatory lesions of the white matter (WM) and gray matter (GM) degenerative atrophy and decide on therapeutic approaches. This pathological process can be characterized by models of the brain’s morphological, structural and functional connectivity. In this thesis, we focus our research on characterizing the pathological processes of MS patients, by using the morphological connectivity graphs. Generated from the characteristics of the GM cortical thickness, measured on T1w-MRI images acquired in clinical routine within OFSEP (French Observatory of Multiple Sclerosis), these graphs are obtained after parcellation of the cortical GM into brain regions according to different anatomical atlases. The GM thickness data are then synthesized in the form of a characteristic vector of each region. These data allow the generation of morphological connectivity matrices, defined by calculating the similarity between the feature vectors of GM cortical thickness for each brain region. The Graph Convolutional Network (GCN) can then be applied to classify the different clinical forms of MS. Since gender and age are correlated to the disease progression, we propose two methods to investigate the impact of these two confounding factors by either including or excluding these demographic data. These data can either be directly included in the GCN as a complementary layer of convolution, or removed from the original graph data by regressing out their impact. We also study other GNN architectures such as the graph attention network to enable adaptive focus on the importance of graph nodes’ neighbors. Graph isomorphic network is also applied as it focuses more on graph structure and presumably achieves maximal expressiveness among all GNNs. These GNNs are further applied to Expanded Disease Status Scale (EDSS) estimation, the score used in clinic to evaluate disability of MS patients. Finally, a region interpretability study is conducted on the graph dataset to define the most relevant regions in MS analysis and their alignment with clinical findings.