fMRI Foundation Models

These models pretrain on large collections of resting-state and task fMRI to learn representations that can be adapted to new participants, datasets, or prediction tasks. Approaches include masked reconstruction, contrastive and cross-atlas alignment, generative diffusion, temporal-dynamics modeling, and multimodal prediction from audio, video, or language; inputs range from voxels and cortical maps to parcellated connectivity. They are evaluated on tasks such as cognitive-state decoding, trait and clinical prediction, brain-response encoding, and transfer across scanners or populations. Results show that performance depends strongly on representation and pretraining objective, while scaling does not guarantee better outcomes; robust out-of-domain evaluation and biologically reliable biomarkers remain open challenges.