Self-Supervised Multi-Modal Learning for Accurate MRI Multiple Sclerosis Segmentation

Multiple sclerosis causes progressive brain morphology alterations through myelin sheath damage, leading to diverse symptoms depending on the affected brain areas. Magnetic resonance imaging (MRI) enables visualization and quantification of corresponding brain lesions, with different MRI pulse sequences as T1-weighted, T2-weighted, and FLAIR providing complementary insights on lesion characteristics. Automated lesion segmentation, typically achieved through supervised deep learning, has accelerated their quantification, with multi-modal approaches recently showing promising results. However, it requires extensive annotations, which are expensive to obtain and prone to errors. In this context, Self-Supervised Learning (SSL) offers a promising alternative to exploit unannotated data, reducing the need for annotations and improving model robustness. This study introduces a multi-modal self-supervised pre-training approach using a generative pretext task to enhance MS lesion segmentation. Our method reaches state-of-the-art performance on two open MS challenge datasets, highlighting the potential of multi-modal SSL multiple sclerosis lesion segmentation.