Template-aware network-collaboration model with hierarchical attention for major depressive disorder detection and biomarker discovery

Abstract

Major Depressive Disorder (MDD) is a complex psychiatric illness characterized by distributed brain network alterations. Resting-state functional Magnetic Resonance Imaging (fMRI) offers a non-invasive probe of these functional disruptions, but its high-dimensional and subtle patterns necessitate computational decoding. While graph-based deep learning methods have been applied to model brain connectivity, they often fail to capture the distinct interactions within and between predefined functional subnetworks. We introduce the Template-aware Network-collaboration Model with Hierarchical Attention (TNA) to address these limitations. The framework constructs a multi-scale graph that explicitly models both intra-subnetwork and inter-subnetwork dynamics through a dedicated collaborative convolution, then employs a lightweight, hierarchical attention mechanism to adaptively weight neurobiologically plausible connections, and lastly derive compact and discriminative representations for classification and biomarker discovery. Experiments on the Rest-meta-MDD dataset show that TNA achieves competitive diagnostic accuracy. Beyond classification, the model provides an efficient and interpretable tool for pinpointing aberrant functional circuits, thereby contributing to the identification of potential network-based biomarkers and offering insights into the network pathophysiology of MDD. Our code is available at: https://github.com/Bald-creator/TNA.