DDN-Net: Deep Residual Shrinkage Denoising Networks with Channel-Wise Adaptively Soft Thresholds for Automated Major Depressive Disorder Identification

Abstract

Major Depressive Disorder (MDD) is a severe mental illness that poses significant challenges to society and families. Recently, using rs-fMRI, several graph-based methods have been proposed for MDD diagnosis. However, these methods encode the entire braingraph directly, without considering the subgraph structure of braingraph and the noise in rs-fMRI. In light of the two deficiencies, we proposed a Deep Residual Shrinkage Denoising Network (DRSD) with channel-shared soft thresholds to denoise the rs-fMRI based braingraphs. Meanwhile, to better preserve the subgraph-structure information of braingraph, we proposed a sub-braingraph normalization method and a sub-braingraph level convolutional network, which involve S-BFS and motif-matching in seven functional brain regions. Comprehensive experiments performed on rest-metamdd dataset show that the performance of DDN-Net surpasses many state-of-the-art depression diagnosis methods with an accuracy of 72.43%, which offers the potential for auxiliary diagnosis of depression in clinical settings.