Toward Integrating Federated Learning With Split Learning via Spatio-Temporal Graph Framework for Brain Disease Prediction

Junbin Mao, Jin Liu, Xu Tian, Yi Pan, Emanuele Trucco, Hanhe Lin

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

Abstract

Functional Magnetic Resonance Imaging (fMRI) is used for extracting blood oxygen signals from brain regions to map brain functional connectivity for brain disease prediction. Despite its effectiveness, fMRI has not been widely used: on the one hand, collecting and labeling the data is time-consuming and costly, which limits the amount of valid data collected at a single healthcare site; on the other hand, integrating data from multiple sites is challenging due to data privacy restrictions. To address these issues, we propose a novel, integrated Federated learning and Split learning Spatio-temporal Graph framework (F G). Specifically, we introduce federated learning and split learning techniques to split a spatio-temporal model into a client temporal model and a server spatial model. In the client temporal model, we propose a time-aware mechanism to focus on changes in brain functional states and use an InceptionTime model to extract information about changes in the brain states of each subject. In the server spatial model, we propose a united graph convolutional network to integrate multiple graph convolutional networks. Integrating federated learning and split learning, F G can utilize multi-site fMRI data without violating data privacy protection and reduce the risk of overfitting as it is capable of learning from limited training data sets. Moreover, it boosts the extraction of spatio-temporal features of fMRI using spatio-temporal graph networks. Experiments on ABIDE and ADHD200 datasets demonstrate that our proposed method outperforms state-of-the-art methods. In addition, we explore biomarkers associated with brain disease prediction using community discovery algorithms using intermediate results of F G. The source code is available at https://github.com/yutian0315/FS2G.
BibTeX
@article{mao2024toward,
  author = {Junbin Mao and Jin Liu and Xu Tian and Yi Pan and Emanuele Trucco and Hanhe Lin},
  title = {{Toward Integrating Federated Learning With Split Learning via Spatio-Temporal Graph Framework for Brain Disease Prediction}},
  journal = {IEEE Transactions on Medical Imaging},
  year = {2024},
  doi = {10.1109/tmi.2024.3493195},
}
Figure