Federated Multimodal and Multiresolution Graph Integration for Connectional Brain Template Learning

Abstract

The connectional brain template (CBT) is an integrated graph that normalizes brain connectomes across individuals in a given population. A well-centered and representative CBT can offer a holistic understanding of the brain roadmap landscape. Catchy but rigorous graph neural network (GNN) architectures were tailored for CBT integration, however, ensuring the privacy in CBT learning from large-scale connectomic populations poses a significant challenge. Although prior work explored the use of federated learning in CBT integration, it fails to handle brain graphs at multiple resolutions. To address this, we propose a novel federated multi-modal multi-resolution graph integration framework (Fed2M), where each hospital is trained on a graph dataset from modality m and at resolution $r_m$ to generate a local CBT. By leveraging federated aggregation in a shared layer-wise manner across different hospital-specific GNNs, we can debias the CBT learning process towards its local dataset and force the CBT to move towards a global center derived from multiple private graph datasets without compromising privacy. Remarkably, the hospital-specific CBTs generated by Fed2M converge towards a shared global CBT, generated by aggregating learned mappings across heterogeneous federated integration GNNs (i.e., each hospital has access to a specific unimodal graph data at a specific resolution). To ensure the global centeredness of each hospital-specific CBT, we introduce a novel loss function that enables global centeredness across hospitals and enforces consistency among the generated CBTs. Our code is available at https://github.com/basiralab/Fed2M.