Federated Multi-Site Brain Disease Diagnosis via Structure-Aware Personalization

Federated learning has emerged as a promising paradigm for multi-site collaborative brain disorder diagnosis, preserving data privacy. However, the effectiveness is fundamentally constrained by the non-independent and non-identical distribution (non-IID) of neuroimaging data across sites. To alleviate this critical challenge, we present a personalized federated learning framework specifically designed to mitigate domain shift caused by site-specific heterogeneity in brain network data. Our framework integrates a structure-aware scaling gate to model topological biases through local and global graph features, and a parameter decoupling strategy to balance global generalization and local personalization. Extensive experiments on multi-site datasets (ABIDE and Schizophrenia) demonstrate that our method consistently surpasses both conventional federated learning approaches and state-of-the-art personalized federated methods.