Towards a general-purpose foundation model for fMRI analysis

Functional Magnetic Resonance Imaging (fMRI) is crucial for studying brain function and diagnosing neurological disorders. However, existing methods for fMRI analysis suffer from reproducibility and transferability challenges, due to complex pre-processing pipelines and task-specific model designs. In this work, we introduce a Neuroimaging Foundation Model with Spatial-Temporal Optimized and Rpresentation Modeling (NeuroSTORM) that learns generalizable representations directly from 4D fMRI volumes, and achieves efficient pre-trained knowledge transfer across diverse downstream applications. Specifically, NeuroSTORM is pre-trained on a remarkable 28.65 million fMRI frames (>$9,000 hours) collected from over 50,000 subjects, spanning multiple centers and covering ages 5 to 100. NeuroSTORM employs a shifted scanning strategy based on a Mamba backbone, enabling efficient direct processing of 4D fMRI volumes. Moreover, we introduce a spatial-temporal optimized pre-training strategy coupled with a task-specific prompt tuning technique, learning transferable fMRI features for efficient downstream adaptation. Experimental results show that NeuroSTORM consistently outperforms existing methods across five diverse downstream tasks: age and gender prediction, phenotype prediction, disease diagnosis, fMRI(-to-image) retrieval, and task-based fMRI (tfMRI) state classification. To further validate the clinical utility of NeuroSTORM, we evaluate it on two clinical datasets comprising patients with 17 different diagnoses from hospitals in the United States and Australia. NeuroSTORM maintains high relevance in predicting psychological/cognitive phenotypes and achieves the best disease diagnosis performance among all existing methods. In summary, NeuroSTORM offers a standardized, open-source foundation model to enhance reproducibility and transferability in fMRI analysis for clinical applications.