Abstract
Thalamic atrophy is a sensitive imaging marker of neurodegeneration in multiple sclerosis (MS) and related disorders, though thalamus segmentation remains method-dependent. Quantitative magnetic resonance imaging (qMRI) may enhance thalamic boundary contrast, particularly in the context of deep learning. We benchmarked thalamic segmentations from two atlas-constrained algorithms, FreeSurfer and FIRST, and two deep learning algorithms, DBSegment and MindGlide (an MS-trained model), against ground truth (GT) labels, tested whether quantitative R1 maps improve performance, and evaluated clinical validity cross-sectionally and longitudinally. We generated thalamus masks using each algorithm from T1-weighted data in a single-scanner cohort (baseline n = 321; 1-year follow-up n = 234) including patients with MS/related disorders and healthy controls. Using MindGlide, we also produced FLAIR- and R1-based masks and ensembles. Manual GT labels were obtained for 50 MS patients using T1w and FLAIR scans. For voxel-wise GT agreement, DBSegment yielded the highest Dice-similarity coefficient; atlas-constrained methods showed the highest sensitivity but lowest precision, while MindGlide balanced both. Volumetrically, MindGlide showed the most accurate estimates; DBSegment and FreeSurfer showed proportional bias, and both atlas-constrained methods overestimated thalamic volumes. Adding R1 input to MindGlide produced modest or no gains in GT agreement. Additionally, MindGlide volumes were most consistently associated with disability and cognitive scores cross-sectionally, and longitudinally showed the largest effects between thalamic volume change and EDSS worsening. Incorporating R1 maps offered no cross-sectional benefit but strengthened longitudinal associations. Higher-resolution qMRI and multi-contrast deep learning architectures may further enhance thalamic segmentation and monitoring in neuroinflammatory diseases.