Deep Learning-Based Detection of Multiple Sclerosis from Multi-View MRI Slices

Multiple sclerosis (MS) is a chronic neurological disorder whereby the myelin sheath that links neurons in the central nervous system is gradually destroyed leading to various functional weaknesses. Although magnetic resonance imaging (MRI) is the indisputable method of mapping MS lesions, the process of scans analysis can be highly problematic: patterns of lesions may change significantly, and the results of scans analysis may be subjected to divergence depending on the observer. In the present study, we suggest an autonomous MS detection pipeline based on deep learning utilising multi-view MRI slices. MS Axial, MS Sagittal, Control Axial, and Control Sagittal were the four categories that were taken considerations. Transfer learning models including VGG16, ResNet50, InceptionV3, and DenseNet121 were evaluated, along with classical classifiers trained on deep features. Furthermore, U-Net-based segmentation was trained on pseudo lesion masks to assess its utility for lesion localization. Experimental results demonstrate the feasibility of automated MS detection, with numerous evaluation metrics for segmentation.