Segmentation basée sur l'apprentissage profond des régions cérébrales liées à la maladie de parkinson et des lymphomes en neuroimagerie

Deep learning has revolutionized medical image segmentation, leading to major improvements in automatic delineation of anatomical and pathological structures. However, significant challenges remain in improving segmentation performance, ensuring proper evaluation, and facilitating clinical translation. This thesis addresses these issues through four main contributions. The first two contributions are of methodological nature and aim, in particular, at dealing with cases where few annotated samples are available for training. The two others are experimental contributions, motivated by the fact that general purpose segmentation have emerged in the past years, including frameworks such as nnU-net and, even more recently, foundation models. In the first contribution, we propose a novel loss function based on projected pooling to introduce soft topological constraints. This method improves the delineation of the red nucleus in quantitative susceptibility mapping (QSM) when dealing with small training samples and is further validated on multiple segmentation tasks. Second, we introduce a frequency-domain disentanglement approach that separates high- and low-frequency components during training. This method enhances segmentation accuracy for specific tasks, particularly in Parkinson related regions segmentation, demonstrating the importance of frequency-specific learning in deep neural networks. Third, we conduct a systematic evaluation of foundation models, including SAM, MedSAM, and UniverSeg, in comparison with conventional supervised learning models. Our findings reveal that while foundation models show promise, they fail to outperform specialized deep learning models in complex medical segmentation tasks, reinforcing the need for task-specific approaches. Fourth, we train an open-source segmentation tool for Primary Central Nervous System Lymphoma (PCNSL) based on the nnU-net framework, which is validated across multiple external datasets. This tool demonstrates strong generalizability, improving the reproducibility and accessibility of automatic tumor segmentation in clinical practice. Our findings highlight key challenges in medical image segmentation. Achieving substantial improvements beyond state-of-the-art models remains difficult, and evaluation practices are often inadequate, with many studies lacking statistical rigor and comprehensive visual assessment. We advocate for greater transparency, improved validation strategies, and better interdisciplinary collaboration between AI researchers and medical professionals. The future of segmentation may involve interactive models that allow user guidance, making deep learning-based segmentation more flexible and clinically applicable.