MIRIAMS: Models Integration for Reliable Identification and Accurate Multiple Sclerosis segmentation

Multiple sclerosis lesion segmentation is a critical task in medical imaging since it aims to identify and delineate brain lesions. In this study, we propose a new method called Models Integration for Reliable Identification and Accurate Multiple Sclerosis Segmentation (MIRIAMS), which is a robust ensemble approach that integrates three distinct neural network architectures to enhance segmentation performance in 3D volumes, setting it apart from the majority of existing methods which operate in 2D. MIRIAMS consists of three different 3D networks: a traditional 3D U-Net, a 3D UNETR that leverages Vision Transformers, and a 3D Swin-UNETR, which employs Swin Vision Transformers. This combination allows the model to capture features with varying receptive fields, improving the overall lesion segmentation. Unlike previous studies, we conduct a per-patient evaluation, ensuring a more individualised and clinically relevant assessment of the segmentation performance. Our results demonstrate that the ensemble approach outperforms individual models, providing a promising, reliable and effective tool for multiple sclerosis lesion segmentation.