Encin, A.; Gilmore, A.; Rokem, A.; Dickie, E.; Glatard, T.

Foundation models pre-trained on large neuroimaging datasets offer a promising approach to overcome the limited sample sizes typical of mental health imaging studies, yet their generalization across diverse clinical populations remains unclear. We present the first systematic benchmark of four publicly available structural MRI foundation models - AnatCL, BrainIAC, 3D-Neuro-SimCLR, and SwinBrain - on tasks relevant to mental health research. Using T1-weighted MRI from Parkinson's Progression Markers Initiative (PPMI), Healthy Brain Network (HBN), and Nathan Kline Institute (NKI), we evaluate these models on sex classification, brain age prediction, and Parkinson's disease (PD) classification, benchmarking against models trained from FreeSurfer-derived cortical thickness and cortical surface area features, as well as an untrained CNN baseline. Although some individual foundation models outperformed FreeSurfer on particular tasks and datasets, 3D-Neuro-SimCLR demonstrated the most consistent performance overall, with the notable exception of HBN sex classification, and all models failed to classify early-stage Parkinson's disease above chance. Notably, untrained CNNs achieved performance comparable to or exceeding FreeSurfer in multiple instances, establishing them as computationally efficient reference models. The cross-model feature correlation analysis reveals that foundation model representations correlate differently with traditional cortical measurements. These findings position structural MRI foundation models, particularly 3D-Neuro-SimCLR and AnatCL, as promising avenues to boost the performance of neuroimaging predictive models in mental health.