Functional Magnetic Resonance Imaging (fMRI) is a powerful tool for mapping brain functional connectivity by extracting blood oxygen signals from various brain regions. However, its clinical utility is often limited by the high cost of data collection and stringent privacy regulations that prevent the pooling of data across multiple healthcare sites. To address these challenges, a new study published in IEEE Transactions on Medical Imaging introduces FS²G, an integrated Federated Learning (FL) and Split Learning (SL) framework designed for robust brain disease prediction.
The Challenge: Data Silos and Privacy
Deep learning models for fMRI analysis typically require large, diverse datasets to generalize well. In practice, individual hospitals often have limited samples, and privacy laws like GDPR or HIPAA restrict the sharing of raw patient data. While Federated Learning allows for decentralized training, running complex spatio-temporal models entirely on local client hardware can lead to overfitting and high computational costs.
Methodology: The FS²G Framework
The FS²G framework partitions a spatio-temporal graph model into two components: a client-side temporal model and a server-side spatial model.
1. Client-Side Temporal Modeling
Local clients process fMRI time-series data using a novel Time-Aware Mechanism (TAM). Unlike standard attention mechanisms, TAM treats time points as nodes in a graph, allowing the model to capture periodic patterns and transitions in brain states. Features are then extracted via an InceptionTime module. By performing this initial processing locally, the framework ensures that raw patient data never leaves the hospital.
2. Server-Side Spatial Modeling
The server receives latent activations from the clients and constructs a "population graph" where each node represents a subject. To analyze this graph, the authors developed UniteGCN, which integrates multiple graph convolutional networks (GraphConv, SAGEConv, and TransformerConv) using Cross-Layer Attention (CLA). This approach allows the model to learn complex relationships between subjects across different sites.
Experimental Evaluation and Performance
The FS²G framework was evaluated on the ABIDE (Autism Brain Imaging Data Exchange) and ADHD200 datasets, demonstrating superior performance compared to existing state-of-the-art distributed learning methods:
- Accuracy: FS²G achieved 72.91% on ABIDE and 72.50% on ADHD200, surpassing other distributed methods like FedNI and SplitAvg.
- Robustness: The model showed significantly lower variance in performance across different sites, suggesting better stability in the face of noisy, multi-site fMRI data.
- Efficiency: By splitting the model, FS²G reduces the local parameter count, mitigating the risk of overfitting on small local datasets.
Clinical Interpretability and Biomarkers
A major contribution of this work is the use of a community discovery algorithm, BrainCD, to identify biomarkers. By analyzing the intermediate results of the FS²G framework, the researchers identified clusters of brain regions—such as the Lingual Gyrus and Fusiform Gyrus—that are highly relevant to Autism diagnosis. This interpretability provides clinicians with valuable insights into the underlying neural mechanisms of the diseases being predicted.
Conclusion
FS²G successfully bridges the gap between data privacy and model performance in medical imaging. By combining the strengths of Federated and Split Learning with advanced spatio-temporal graph networks, it provides a scalable solution for multi-site collaborative research. The source code for the project is available at https://github.com/yutian0315/FS2G.
Summary updated