Encin, A., Gilmore, A., Rokem, A., Dickie, E., Glatard, T.
First systematic benchmark of four structural MRI foundation models on sex classification, brain age prediction, and Parkinson's disease classification.
Foundation models pre-trained on large neuroimaging datasets may help overcome limited sample sizes in mental health studies, but their generalization across diverse clinical populations is unclear.
Evaluated AnatCL, BrainIAC, 3D-Neuro-SimCLR, and SwinBrain on T1-weighted MRI from PPMI, HBN, and NKI for sex classification, brain age prediction, and Parkinson's disease classification, comparing against FreeSurfer features and untrained CNN baselines.
3D-Neuro-SimCLR showed the most consistent performance overall, though all models failed to classify early-stage Parkinson's disease above chance. Untrained CNNs achieved comparable or better performance than FreeSurfer in multiple instances, establishing them as computationally efficient reference models.