Detecting Major Depressive Disorder by Graph Neural Network Exploiting Resting-State Functional MRI

Major Depressive Disorder (MDD) has raised concern worldwide because of its prevalence and ambiguous neuropathophysiology. Resting-state functional MRI (rs-fMRI) is an applicable tool for measuring abnormal brain functional connectivity in MDD. However, effective method for early diagnosis and treatment for MDD is still lacking. In this study, we propose a three-stage classification framework to analyze rs-fMRI data for the diagnosis of MDD. We first apply self-supervised pretraining on developed graph encoder, incorporating triplet relationship among input subjects, to enable higher ability to learn robust and discriminative graph representations. Then, supervised classification is performed utilizing the pretrained encoder. Specifically, to better model subjects’ brain as functional connectivity network, our developed graph encoder consists of following modules: non-linear feature transformation, graph isomorphism convolution, topk pooling and hierarchical readout. Afterwards, ensemble learning is implemented to further boost model’s performance. Finally, we identify salient ROIs by investigating pooling scores learned by topk pooling layers, which implies brain areas potentially related to MDD and equips our model with fair interpretability. Experimental results on Rest-meta-MDD, a large-scale multisite dataset, suggest the efficacy of our method.