Deep factor regression for computer-aided analysis of major depressive disorders with structural MRI data
Major depressive disorder (MDD) is a prevalent and debilitating psychiatric mood disorder that affects millions of people worldwide. Conventional methods for MDD severity diagnosis usually rely on neuropsychological assessments that are subjective and susceptible. Recently studies have shown that structural MRI (sMRI) can provide objective biomarkers for MDD severity diagnosis. However, current MRI-based methods generally rely on hand-crafted imaging features and cannot explicitly identify MDD-associated depression symptoms, thus failing to increase our understanding of clinical and cognitive staging of MDD. In this paper, we first employ five depression symptom factors to quantitatively measure MDD grade from different aspects. Then, we design an end-to-end deep factor regression network (DFRN) to predict these factors directly from 3D T1-weighted sMRI scans. To uncover the contributions of different brain regions, we generate attention maps to uncover the implicit attention of the learned DFRN models. Experimental results on 116 MDD subjects show that the predictions for all five factors are positively correlated with ground-truth values. Attention maps also highlight the most informative brain regions for each factor.
