Deep Learning-Based Multiple Sclerosis Lesion Segmentation at 7 Tesla: Comparison of Single- and Multi-Contrast MRI Inputs

Abstract

Automated lesion segmentation can support rapid and reproducible quantitative assessment and long-term monitoring of multiple sclerosis. Ultra-high-field 7T MRI provides increased spatial resolution and lesion visibility but introduces challenges for automated analysis, including field inhomogeneity, susceptibility-related artifacts, and spatially varying image contrast. Although deep learning methods perform well at conventional field strengths, their application to 7T MRI remains underexplored. We evaluated nnUNet using FLAIR alone, MP2RAGE alone, and combined MP2RAGE+FLAIR. The dataset included 115 longitudinal 7T MRI scans from 57 patients with MS, with 20 scans from 11 independent patients reserved for testing. DeepMedic was included as a comparative architecture. On the test set, nnUNet achieved mean DSC values of 0.72, 0.71, and 0.66 for combined, MP2RAGE-only, and FLAIR-only inputs, respectively. No significant DSC differences were observed between nnUNet and DeepMedic. For both architectures, adding FLAIR to MP2RAGE did not show a statistically significant additional benefit in the present test cohort, whereas MP2RAGE-containing inputs significantly outperformed FLAIR-only input. The findings of the present study provide preliminary evidence suggesting that the observed voxel-level segmentation performance was primarily associated with the inclusion of MP2RAGE. However, lesion-level performance depended on lesion size, indicating further investigation of contrast-specific effects on small-lesion detection.