Task-Agnostic Brain Representations: A Foundation Model for fMRI Using Masked Autoencoders

Motivation: Understanding brain activity is a key neuroscience challenge. While fMRI offers insights, its high dimensionalit may limit its use in modeling brain function. Goal(s): We propose a foundation model for ROI-based fMRI data, trained on resting-state data from HCP, to develop generalizable brain latent representations. Approach: Using a masked autoencoder with self-supervised learning, we train a transformer model on fMRI time series from the HCP dataset. The model encodes signals into a latent space and reconstructs masked segments, capturing key spatiotemporal features. Results: The model produced strong, transferable representations, achieving high performance in downstream tasks like classification across seven cognitive tasks. Impact: We foundation model for fMRI, trained on resting-state data from the HCP to develop generalizable brain representations. Using self-supervised learning, this task-agnostic model can be applied to various neuroscience tasks, including physiological prediction and brain decoding.