Sequence-level machine learning assessment of MRI protocols in classifying multiple Sclerosis lesions

Abstract

Accurate classification of Multiple Sclerosis (MS) lesion activity is vital for disease management, yet gadoliniumenhanced MRI raises safety and cost concerns. This study evaluates contrast-free MRI sequences-T1-weighted (T1W), T2-weighted (T2W), Fluid-Attenuated Inversion Recovery (FLAIR), Diffusion-Weighted Imaging (DWI), and Susceptibility-Weighted Imaging (SWI)-using machine learning to optimize MS lesion classification. Retrospective data from 31 MS patients (187 lesions; 39 active) were collected from November 2023 to February 2024 using a 1.5 T S Magnetom scanner. Approximately 7500 radiomic features were extracted, reduced to 214-108 via Spearman correlation and sequential forward selection. LightGBM models were trained on single, pairwise, and multi-sequence combinations, evaluated on a 28-lesion test set using AUC-ROC, sensitivity, specificity, and precision-recall AUC. FLAIR achieved the highest single-sequence AUC-ROC (0.83, 95 CI: 0.77-0.89), while the FLAIR + T2W + SWI combination reached 0.88 (95 CI: 0.83-0.93), rivaling the fivesequence model (0.89, p = 0.31). Texture (52.3 ) and wavelet (31.8 ) features dominated, with robust performance under noise (AUC 0.86). This contrast-free FLAIR + T2W + SWI protocol offers accuracy comparable to gadolinium-based methods, supporting safer, cost-effective MS imaging, though multi-center validation is needed due to the small test set.