Diagnosis of Major Depression Disorder Using 3D Convolutional Neural Networks



Alipoury MM¹ ; Tavakoli H² ; Masoudnia S³ ; Ameri A³ ; Zadeh MRN³ Show Affiliations
Source: Frontiers in Biomedical Technologies Published:2021

Abstract

Major Depressive Disorder (MDD) is a significant cause of morbidity and unproductivity worldwide. Due to heterogeneity of MDD characteristics, diagnosis based on clinical questionnaires would not be adequate. Previous Magnetic Resonance Imaging (MRI) studies have reported that MDD would be accompanied with changes in cortical and subcortical Gray Matter (GM) structures including the hippocampus, amygdala, anterior cingulate cortex, caudate nucleus, and dorsolateral prefrontal cortex. Hand-crafted Morphological attributes can be identified by Voxel-Based Morphometry (VBM) method. Convolutional Neural Network (CNN) can facilitate the identification of morphological changes in structural MRI for clinical practice. We propose a combination of classic feature extraction and three-Dimensional (3D) CNN model that can extract deep-learned features automatically from cortical hand-crafted VBM attributes. This combination is done by giving the VBM gray matter features as input to model which is able to extract heterogeneous changes. By overcoming the underlining heterogeneity and effectively detecting the abnormalities of GM, accuracy of 86% and specificity of 83% and sensitivity of 89% was achieved for the evaluating of classification between the MDD vs the Healthy Control (HC) using REST-meta-MDD Data Sharing Consortium. Our algorithms have the potential to provide an unbiased, and non-invasive assessment of MDD that may allow more efficient treatments. Copyright © 2021 Tehran University of Medical Sciences.