Classification of Multiple Sclerosis Lesions in MRI Scans Using Vision Transformers

Multiple Sclerosis (MS) is a chronic inflammatory and demyelinating disease of the central nervous system, where early and accurate diagnosis via Magnetic Resonance Imaging (MRI) is crucial for effective patient management. While deep learning models, particularly Convolutional Neural Networks (CNNs), have shown promising results in medical image analysis, they often struggle to capture long-range dependencies and global contextual information, which are essential for identifying complex and scattered lesion patterns in MS. To address these limitations, this study proposes ViT-MS, a novel framework based on the Vision Transformer (ViT) architecture for the automatic classification of MS lesions in FLAIR MRI images. We utilized a dataset comprising ۶۰ MS patients, from which ۲۳,۰۴۰ image patches were extracted using a sliding window technique. To ensure robust evaluation and prevent data leakage, a strict patient-wise data splitting strategy was implemented. Furthermore, to handle the significant class imbalance between lesion and healthy tissue patches, we employed a weighted loss function combined with specialized sampling strategies. Experimental results demonstrate that the proposed ViT-MS model achieves an accuracy of ۹۶.۷%, an F۱-score of ۰.۹۶۷, and an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of ۰.۹۹۵. These results indicate a significant performance improvement over baseline CNN models such as ResNet-۵۰ and EfficientNet. The findings suggest that the self-attention mechanism in ViT effectively captures global contextual features, making it a powerful tool for assisting radiologists in the accurate and efficient detection of MS lesions.

کلیدواژه ها:

Multiple sclerosis (MS) ، MRI images ، deep learning ، Vision Transformer (ViT) ، automatic classification