Major Depressive Disorder Classification with 3D CNNs and Grad-CAM Visualization on Structural Magnetic Resonance Images

Major Depressive Disorder (MDD) is a highly prevalent psychiatric disease with a complex etiology. Currently, the diagnosis of MDD is still mainly based on patients’ reporting and subjective judgement of clinicians. Biomarkers derived from structural Magnetic Resonance Imaging (sMRI) have pointed out several brain regions that can be discriminative of the disease. In this context, Deep Learning (DL) can provide end-to-end models to support disease diagnosis and classification from image analysis. In this work, three different 3D-DenseNet architectures are trained and validated on sMRI scans of a large cohort of 282 MDD subjects and 251 matched normal controls (NC) obtained from the REST-meta-MDD project. Results show that the DL models allowed reaching AUC of 0.63 and Recall of 0.88 in discriminating MDD and NC, confirming that 3D networks can fully mine the spatial information of sMRI. Furthermore, to visualize the most salient regions of the brain which contribute to the classification process, Grad-CAM is applied to the models. Heatmaps illustrate how the model with the highest performances in terms of AUC and Recall focuses on specific brain areas which have been confirmed to be anatomical biomarkers for MDD, namely the temporal lobe, the pons and the cingulate gyrus.