Automated Segmentation of Multiple Sclerosis Lesions using Deep Learning

This research introduces a pioneering Multiple Sclerosis (MS) lesion segmentation approach, employing a multi-scale UNet framework. The model architecture, comprised of Contraction and Expansive Paths, adeptly captures intricate details within input images. Leveraging a dataset of 60 patients MRI scans, manually segmented across three sequences, the study’s preprocessing involves slicing 3D data into 2D images, standardizing dimensions and enhancing image quality via sharpening and intensity normalization. Notably, the multi-scale UNet model achieved outstanding results, showcasing a remarkable Dice Similarity Coefficient (DSC) of 99.78% for lesion segmentation. Although the model showed decreased loss values over the training phase. The suggested approach outperformed existing algorithms like CNN, Acu-Net and ALLNet, according to a rigorous comparison analysis that looked at a number of different parameters. This study improves how doctors make decisions by making a big advance in automatically finding MS lesions. Notwithstanding challenges in conducting direct comparisons because of differences in datasets, the methodology’s proven robustness and efficacy point to its potential to revolutionize MS diagnosis and treatment.