Segmentation automatique du système vasculaire pour améliorer le système d'aide à la décision basé sur l'IA pour les maladies artérielles pérophériques

Peripheral artery disease (PAD) affects over 230 million people worldwide. Generally caused by atherosclerosis, it is characterized by the narrowing or occlusion of the arteries in the lower limbs. Owing to its frequent asymptomatic presentation, PAD is often diagnosed at advanced stages, increasing the risk of cardiovascular complications and amputation. As such, PAD is associated with high mortality and morbidity, representing a major public health concern. Computed Tomography Angiography (CTA) is commonly used to assess arterial lesions and anatomy, guiding revascularization strategies. However, current manual analysis of CTA is time-consuming and operator-dependent, underscoring the need for automated tools to support clinical decision-making. This thesis aims to develop an AI-based system for the comprehensive and automated assessment of PAD, facilitating personalized treatment planning. The main challenge addressed is the segmentation of small, tortuous arteries with frequent occlusions, as well as calcification plaques and arterial stents. The segmentation of these structures in the lower limbs should enable the extraction of precise anatomical and clinical features to provide clinicians with the necessary information to guide the preoperative planning for PAD.To overcome this segmentation challenge, we propose two methodological and one clinical contributions. (1) SoftMorph is a framework that converts any binary morphological operation into a differentiable probabilistic counterpart, enabling its integration into neural networks either as a final layer or within the loss function. Probabilistic filters are defined as the expectation of the binary filter over all possible binary configurations and expressed as a multi-linear polynomial derived from its truth table. For intractable cases, approximations are obtained via quasi-probabilistic operators by applying various fuzzy logic operators to convert the Boolean expression defining the morphological operation, preserving the original filter's complexity. Experiments demonstrated improvements in topological preservation for the segmentation of tubular structures. (2) Regional Hausdorff Distance losses are developed, a family of loss functions to improve boundary precision in segmented structures, particularly relevant in pathological contexts. The method relies on a fully differentiable erosion-based distance function to produce differentiable computation of the maximum, modified, and averaged regional Hausdorff Distances. These loss functions achieved state-of-the-art performance without requiring any auxiliary losses for the training of segmentation networks across multiple modalities. (3) Finally, these innovations are applied in a real clinical context for the segmentation of lower-limb arteries, stents and calcification plaques. PADSET, an in-house CTA dataset of PAD patients, is curated and annotated to provide ground truth masks of each structure. Then, deep-learning-based automatic segmentation methods are explored, along with the application of the two previous technical contributions to achieve high-performance automatic segmentation on the PADSET dataset. Additionally, the approach involves automatically identifying key arterial branches to extract the precise locations of clinically relevant features.The automatic tool reduces inter-observer variability, and supports pre-surgical planning for the treatment of PAD.