Continual Learning for fMRI-Based Brain Disorder Diagnosis via Functional Connectivity Matrices Generative Replay

Researchers Qianyu Chen and Shujian Yu have introduced a groundbreaking continual learning framework designed to overcome a major hurdle in medical AI: catastrophic forgetting. Current AI models for diagnosing brain disorders from functional magnetic resonance imaging (fMRI) are typically trained on data from a single site or require simultaneous access to all multi-site data, which is unrealistic in clinical practice where patient information arrives sequentially from different institutions. This leads to models that fail to generalize and forget previously learned knowledge when trained on new data.

Their solution, detailed in a paper accepted by CVPR 2026, is the first framework specifically built for fMRI-based diagnosis across heterogeneous clinical sites. The core innovation is a structure-aware variational autoencoder (VAE) that synthesizes high-quality, realistic functional connectivity (FC) matrices—representations of neural interactions—for both patient and control groups. This generative model serves as a backbone for a multi-level knowledge distillation strategy, which aligns predictions and graph representations between new, real patient data and the AI-generated "replayed" samples from past tasks.

To maximize efficiency, the team incorporated a hierarchical contextual bandit scheme for adaptive replay sampling, intelligently deciding which past data to revisit. In experiments on multi-site datasets for major depressive disorder (MDD), schizophrenia (SZ), and autism spectrum disorder (ASD), their generative model significantly enhanced data augmentation quality. The overall framework substantially outperformed existing continual learning methods in preserving diagnostic accuracy and mitigating catastrophic forgetting as it learned from new clinical sites over time.