Multiple sclerosis is a chronic neurological disease that disrupts normal nerve signal transmission within the central nervous system, resulting in a wide range of physical and cognitive impairments. Early and accurate diagnosis is essential for effective treatment and disease management. Automated detection and precise segmentation of MS lesions from brain magnetic resonance imaging scans play a crucial role in supporting radiologists during clinical decision-making. This study provides a comprehensive evaluation of deep learning–based UNet architectures for MS lesion detection and segmentation. Five models are examined: the standard UNet, DenseUNet, Attention UNet, Residual UNet, and Swin-UNet, which integrates a Transformer-based encoder to enhance contextual feature representation. The models are trained and evaluated on four benchmark MRI datasets namely ISBI-2015, Mendeley, MICCAI-2016, and MICCAI-2021 covering diverse imaging conditions and lesion characteristics. Performance is assessed using established segmentation measures including Dice Similarity Coefficient, Jaccard Index, Sensitivity, and Precision. Additionally, the best-performing UNet variant is compared with other state of the art deep learning models to further validate its effectiveness. Experimental results demonstrate that incorporating dense connections, attention mechanisms, residual learning, and Swin Transformer modules significantly enhances the accuracy and robustness of MS lesion segmentation. These findings underscore the potential of advanced UNet based frameworks as reliable tools for assisting clinicians in the early and accurate diagnosis of multiple sclerosis.
Jain et al. (2026) studied this question.