DL-Based Lesion Segmentation and Novel Biomarkers for Early EDSS Prediction in Multiple Sclerosis

: This study introduces an integrated pipeline for the early prediction of disability progression in multiple sclerosis (MS), targeting the low EDSS range (0–2.5). It combines multimodal data (MRI, age, clinical features) with a novel deep learning model called MSU-AttNet, which employs multi-scale dilated attention mechanisms to achieve accurate segmentation of MS lesions of varying sizes and locations. The model delivers Dice scores of 0.75 on a private dataset, 0.71 on the ISBI 2015 challenge, and 0.83 when tested on acute ischemic stroke lesions, demonstrating strong generalization. The pipeline also includes automatic brain lobe segmentation, standard MRI feature extraction, and two innovative EDSS-driven biomarkers: the EDSS-Topographic Lesion Proximity Score (ETPS), which quantifies strategic periventricular lesion burden, and the EDSS-Weighted Regional Lesion Index (ERLI), which integrates anatomically weighted lesion load adjusted for disease duration. Among several machine learning models evaluated, LightGBM, using these novel biomarkers together with conventional MRI features and clinical/demographic variables, provides the best performance with an RMSE of 0.51 and an MAE of 0.41, highlighting their key contribution to improving early disability risk assessment in MS.