A generalizable structural brain MRI foundation model built through dual-priority federated pretraining

TL;DR - BrainFedFM is a structural brain MRI foundation model pretrained federatively on 164,707 scans across 42 sites, avoiding the need to pool sensitive raw images. Its dual-priority training approach achieved leading average performance across diverse neuroimaging tasks and improved robustness for underrepresented populations.

• Spatial-priority masking emphasizes informative anatomical regions during local site training.
• Site-priority server aggregation weights complementary contributions beyond cohort size alone.
• Evaluation covered 20 datasets and 17 classification, regression, and segmentation tasks.
• BrainFedFM ranked first on average among seven models, including four centralized foundation models, with especially consistent gains in classification and regression.