Deep learning for 3D delineation of multiple sclerosis lesions in magnetic resonance imaging

Multiple sclerosis (MS) is a neurodegenerative disease that affects the central nervous system. The immune system damages the myelin sheath that surrounds nerve fibers, leading to the formation of lesions known as plaques. These plaques evolve from an initial inflammatory phase to a chronic stage, where they take on a scar-like appearance, hence the term “sclerosis.” Depending on the number, size, and location of the lesions, the disease can manifest in different forms, ranging from mild to severe. A rapid and accurate diagnosis is crucial, as it enables the selection of the most effective therapy for each patient. One of the most powerful tools for both diagnosing and monitoring the progression of the disease is Magnetic Resonance Imaging (MRI). In this context, there is a growing need for automatic and reliable methods for lesion segmentation that can support clinicians by reducing the limitations of manual segmentation, such as inter-operator variability and the considerable time required. Several approaches based on convolutional neural networks have been proposed in the literature, although they often rely on multiple MRI sequences and large datasets, which are not always available. The aim of this thesis is to develop an automatic method for the segmentation of MS lesions based exclusively on FLAIR images. For this purpose, a 3D U-Net architecture with attention mechanism was implemented and trained using a voxel-wise classification approach. The goal is to provide a robust and efficient tool that can support clinicians in the diagnosis, monitoring, and characterization of MS, offering a faster and accurate alternative to manual segmentation. The proposed method was evaluated on the publicly available dataset by Lesjak et al., which includes MRI scans of 30 MS patients. After excluding two cases due to inconsistent dimensions, FLAIR images from 28 patients were used. A preprocessing pipeline was applied, including bias field correction, resampling to isotropic resolution, and intensity normalization. The 3D U-Net with attention mechanism was trained using 3D patches extracted from FLAIR scans, with a balanced sampling strategy between lesion and non-lesion voxels. Model performance was assessed through leave-one-subject-out cross-validation, a strategy particularly suited for small datasets. The proposed model achieved promising results, with mean values of dice score coefficient of 70.02%, sensitivity of 68.28%, intersection over union of 55.48%, and Precision of 75.11%. In some cases, the dice exceeded 80%, highlighting the robustness of the method despite the small dataset and the use of a single imaging modality. Performance variability was observed in patients with very small lesion volumes, suggesting the need for further strategies to address data imbalance. In conclusion, the developed approach demonstrated efficiency and generalization capability, reaching results comparable to or better than other methods relying only on FLAIR images. By reducing preprocessing and computational demands, the workflow improves cost-effectiveness and clinical applicability. Nevertheless, limitations remain, mainly related to the variability in cases with low lesion load and the computational effort required by leave-one-subject-out cross-validation. Future work should focus on expanding the dataset, exploring alternative validation strategies, and investigating techniques to mitigate data imbalance, in order to further enhance the model’s performance and reliability.