Využití neuronových sítí pro detekci a hodnocení lézí v bílé hmotě mozkové u pacientů s roztroušenou sklerózou s využitím interpretovatelných metod umělé inteligence

Abstract

Multiple sclerosis (MS) is a chronic disease of the central nervous system; magnetic resonance imaging (MRI) is key to monitoring disease progression and treatment efficacy. Assessing disease burden requires quantifying the number and volume of white matter lesions. However, manual lesion segmentation is time-consuming and prone to inter-rater variability. This thesis proposes an automated system for MS lesion segmentation from a single MRI scan using the nnU-Net v2 framework and a multi-architecture ensemble (CNN-3D, ResEncL-3D, ResEncL-2.5D). Fifteen models (3 architectures × 5 folds) were trained on a combined dataset of MS lesions (MSLesSeg-2024) and vascular white matter hyperintensities (WMH Challenge 2017). The validation-selected Best-2/arch ensemble (6 models) achieved a Dice coefficient of 0.718 on MSLesSeg-2024 and 0.803 on WMH-2017, comparable to the winners of both challenges in a methodologically defensible manner. A key contribution is the integration of explainable AI through three complementary methods (Grad-CAM, 3D RISE, Occlusion Sensitivity) enabling visual verification of the model's decision-making process. The work also documents important negative results: size-based labeling and TopK loss are counterproductive for this task.