Metadata-Driven Federated Learning of Connectional Brain Templates in Non-IID Multi-Domain Scenarios
IEEE Transactions on Medical Imaging
|November 10, 2025
Summary
MetaFedCBT addresses data heterogeneity in federated learning for brain connectivity. This novel framework improves holistic brain template construction by predicting missing data, enhancing multi-domain analysis.
Area of Science:
- Neuroscience
- Machine Learning
- Data Science
Background:
- Federated learning enables collaborative estimation of connectional brain templates (CBTs) across multiple data domains (e.g., hospitals) while preserving data privacy.
- Existing federated CBT methods struggle with the non-independent and identically distributed (non-IID) nature of multi-domain brain connectivity data, leading to suboptimal performance.
- This data heterogeneity, or non-IID issue, limits the accuracy and generalizability of learned CBTs.
Purpose of the Study:
- To develop a novel federated learning framework, MetaFedCBT, designed to overcome the non-IID challenge in multi-domain connectional brain template learning.
- To enhance the representation ability of federated CBT by effectively handling the heterogeneity of brain connectivities across different data sources.
- To enable privacy-preserving holistic CBT learning through metadata-driven generation of brain connectivities for unseen domains.
Main Methods:
- Proposed a metadata-driven federated learning framework (MetaFedCBT) incorporating a metadata regressor and local-global network residual weights.
- Developed a metadata-driven connectivity generator to predict brain connectivities of unseen domains based on learned metadata.
- Implemented an iterative update mechanism for predicted metadata and brain connectivities to progressively approximate unseen domains during federated learning.
Main Results:
- MetaFedCBT effectively addresses the non-IID issue by generating informative brain connectivities, crucial for privacy-preserving holistic CBT learning.
- Experiments on multi-view morphological brain networks demonstrated MetaFedCBT's superiority over existing federated CBT learning models.
- The proposed framework significantly advanced the state-of-the-art performance in multi-domain CBT estimation.
Conclusions:
- MetaFedCBT offers a robust solution for federated connectional brain template learning in the presence of non-IID data.
- The metadata-driven approach enhances the accuracy and utility of holistic brain representations derived from decentralized data.
- This work advances the field of neuroimaging analysis by enabling more effective and privacy-preserving collaborative learning across diverse datasets.
