Researchers' New AI Model Identifies Depression Across MRI Sites with 73.6% Accuracy

A new arXiv paper from Zhanpeng Zheng and colleagues tackles one of neuroimaging's hardest problems: identifying major depressive disorder (MDD) from resting-state fMRI when data comes from multiple clinical sites. Different MRI scanners and protocols introduce distribution shifts, making it difficult to generalize a model trained on one hospital to another. The researchers built a multi-source, multi-view graph domain adaptation framework that models three complementary views of functional connectivity simultaneously: Pearson correlation, sparse representation, and Granger causality graphs. Each view is encoded by a view-specific graph attention network, and a dual-stream adaptive fusion module captures pairwise interactions between views. Lightweight hyperbolic residual encoding then refines the representation with curvature-aware geometry, preserving disease-relevant structure better than flat Euclidean spaces.

To align source and target domains, the framework uses class-wise Cauchy-Schwarz alignment alongside adversarial learning, information maximization, and confidence-aware pseudo-labeling—all in an unsupervised setting for the target sites. Evaluated across seven unlabeled target domains, the method achieved a mean accuracy of 73.60% and an AUC of 71.90%, outperforming existing approaches that treat multi-view learning and domain adaptation separately. The authors note that few prior studies jointly model multiple FC views under multi-source unsupervised domain adaptation. Their unified approach not only handles heterogeneous acquisition conditions but also maintains cross-view consistency. The source code is publicly available, inviting replication and clinical follow-up. For AI-driven mental health diagnostics, this represents a step toward models that can be trusted across different hospitals and scanner types.