Template-Based Federated Multiview Domain Alignment for Predicting Heterogeneous Brain Graph Evolution Trajectories from Baseline

Emircan Gündogdu, Islem Rekik

Abstract: Predicting the brain graph (or connectome) evolution trajectory can aid in the early diagnosis of neurological disorders or even prior to onset. However, when dealing with heterogeneous datasets collected from various hospitals, each representing a unique domain (e.g., functional connectivity), the task of prediction becomes even more challenging within a federated learning (FL) context. To the best of our knowledge, no FL method was designed to predict brain graph evolution trajectory with hospitals trained on separate data domains, which presents one of the complex non-IID and data heterogeneity cases that existing FL methods struggle to solve in general. To address this issue, we introduce the concept of template-based domain alignment, where we leverage a prior universal connectional brain template (CBT), encoding shared connectivity patterns across the available domains. Specifically, we propose a template-based federated multiview domain alignment (TAF-GNN). Our TAF-GNN architecture consists of two parts: (1) template-based alignment where we align each distinct domain (i.e., hospital) to the prior universal CBT trajectory domain by using our proposed graph-based domain aligner network, and (2) GNN-based trajectory federation where we train a 4D graph neural network (GNN) for each hospital on its CBT-aligned brain trajectories. Our results on both real and simulated longitudinal connectivity datasets demonstrate that our TAF-GNN significantly outperforms other architectures with different alignment methods in both centralized and federated learning paradigms. Our TAF-GNN code is available on GitHub at https://github.com/basiralab/TAF-GNN.