Artificial Intelligence analysis of lesion dynamics and brain volume in patients with multiple sclerosis
Multiple Sclerosis and Related Disorders, 2026
Abstract
Background
Quantitative MRI markers increasingly complement conventional clinical assessment in multiple sclerosis (MS). Artificial intelligence (AI)–based volumetry enables standardized evaluation of lesion burden and brain atrophy in routine care.
Objective
To examine the association between AI-derived volumetric measures and disability, assess whether annual brain volume loss (ABVL) and lesion dynamics predict atrophy, and descriptively compare proxy radiological phenotype groups with established clinical phenotypes in a real-world MS cohort.
Material and Methods
This retrospective study included 888 MRI examinations from 455 patients with MS (2016–2020). Longitudinal analyses were performed in 234 patients with ≥2 scans (667 MRIs), and an early MS cohort comprised 302 patients (580 scans). Automated segmentation (mdbrain® v3.4.0) provided lesion metrics and age/sex/skull-volume-adjusted brain volumes from routine 3D FLAIR and native T1-weighted sequences acquired under real-world clinical conditions. Pathological atrophy was defined as a normative z-score < –2. Generalized estimating equations (GEE) evaluated predictors of atrophy.
Results
Proxy radiological phenotype groups were defined as lesion-led (43.2%), cortex-led (35.4%), and NAWM-led (21.4%); clinical phenotypes included RRMS (82%), SPMS (14%), and PPMS (4%). EDSS correlated with lesion volume (ρ=0.28, p<.001) and total brain volume (ρ=–0.32, p<.001). In 433 longitudinal intervals, 11.8% showed mdbrain-defined atrophy. Higher EDSS (OR 1.53, 95% CI 1.28–1.83, p<.001) and longer follow-up (OR 2.24, 95% CI 1.36–3.70, p=.001) independently predicted atrophy; ABVL showed only borderline significance (p=.071). Lesion dynamics were not independently predictive of atrophy (p>.60). ABVL alone showed low discriminative value (AUC 0.571), whereas EDSS + interval length achieved AUC 0.766.
Conclusion
Gray-matter–predominant atrophy correlated more strongly with disability than lesion burden and frequently occurred in the absence of new lesions, indicating lesion-independent neurodegenerative processes that were observed across the defined proxy radiological phenotype groups. AI-based quantitative MRI offers reproducible atrophy assessment in real-world practice and may support quantitative MRI-based monitoring frameworks that include brain volume loss and facilitate detection of subclinical progression.
