AN INTERPRETABLE MRI-BASED DEEP LEARNING MODEL TO SUPPORT THE DIAGNOSIS OF MAJOR DEPRESSIVE DISORDER

Major depressive disorder (MDD) is a highly prevalent psychiatric disease which results from a complex and not completely understood interaction of genetics and environmental factors. Moreover, it is associated with an increased risk of cognitive deficits, diabetes mellitus, heart disease and stroke, affecting the quality of life of patients. The diagnosis of MDD mainly depends on criteria from the Diagnostic and Statistical Manual of Mental Disorders 5th edition (DSM-5) and treatment response, which could be subjective and susceptible. As a robust complement to clinical neurobehavior-based detection, computer-aided diagnosis based on neuroimaging data holds the promise of objective diagnosis and prognosis of brain disorders. Magnetic resonance imaging (MRI) is part of the neuroimaging research and represents a potential diagnostic modality for MDD. In recent years, Convolutional Neural Networks (CNNs), which are the most established among Deep Learning (DL) models, have shown convincing results in the diagnostic classification of brain disorders. However, few studies focused on DL research related to 3D structural MRI (sMRI) data of MDD patients. Therefore, the main objective of this work is to propose a framework for the automated identification of MDD using a CNN model, called 3D-DenseNet, based on sMRI data. The dataset includes T1-weighted sMRI scans collected from 282 MDD subjects and 251 matched normal controls (NC) recruited from the Faculty of Psychology of the Southwest University, an imaging site of the REST-meta-MDD consortium. Three different 3D-DenseNets, with 169, 201 and 264 layers respectively, are trained to classify the preprocessed sMRI scans into two categories: MDD patients (positive) and NC (negative). The dataset is split into training and validation sets using stratified 5-fold cross-validation. Four metrics are used to evaluate the CNNs’ performance on validation sets in the task of MDD detection, including Area Under the ROC Curve, Accuracy, Precision and Recall. Results, averaged on the 5 folds, are presented. Besides, an interpretability module, consisting of gradient-weighted class activation mapping (Grad-CAM), is used to provide insight into the most relevant brain regions for classification, possibly confirming their role as neuroimaging biomarkers of MDD. Although its performances can be improved, this framework shows potential to be employed in a clinical setting to support the diagnosis of MDD.