AUTOMATED WHITE MATTER LESION SEGMENTATION USING DEEP LEARNING: PERFORMANCE BENCHMARKS IN MULTIPLE SCLEROSIS AND CEREBROVASCULAR DISEASE

Abstract

White matter lesions visible on T2-FLAIR MRI serve as primary diagnostic and monitoring biomarkers in both multiple sclerosis and cerebrovascular disease, yet their manual delineation by neuroradiologists is slow, resource-intensive, and variable between raters. Deep learning segmentation models — particularly three-dimensional U-Net architectures and vision transformers applied to multi-modal MRI — have been developed as automated alternatives. A review of 36 studies published from 2016 to 2024 indicates that these models achieve Dice similarity coefficients of 0.72–0.89 for MS lesion segmentation and 0.68–0.84 for cerebrovascular WML, values that overlap substantially with the reported range of inter-expert agreement. Lesion volume estimates from automated segmentation correlate with manual reference measurements at r = 0.91–0.98. For longitudinal MS monitoring specifically, automated detection of new or enlarging lesions reaches sensitivity of 78–93% with specificity of 88–97%. These capabilities have practical implications for MRI services operating under neuroradiological workforce constraints, including those in Uzbekistan where specialist neuroimaging interpretation is concentrated in Tashkent.