An Efficient Attention-Based Network for Screening Major Depressive Disorder with sMRI

Major Depressive Disorder (MDD) is a prevalent psychiatric disorder that adversely affects the quality of life of those affected. Artificial intelligence (AI) has emerged as a promising tool for computer-aided diagnosis of MDD based on magnetic resonance imaging (MRI) technology. In this study, we propose a novel deep learning framework that utilizes the SlowFast network to analyze structural MRI (sMRI) for screening MDD. We incorporate a multi-scale feature fusion mechanism that integrates attentional mechanisms with path information to enhance the channel attentional features of gray matter images. Furthermore, we explore the correlation between regional features of sMRI slices and the spatial information our classification model focuses on. We evaluated our proposed method on MDD datasets generated from multiple research centers with varying scanning parameters and scanner field strengths, achieving a classification accuracy of 93.12%. Our method outperforms existing sMRI-based unimodal methods and provides clinicians with a reliable method for MDD diagnosis. During the investigation of regional spatial dependency, an impressive classification accuracy of 94.38% was achieved. By introducing the attention mechanism in the multi-scale feature fusion process, we can effectively focus on channel features, achieve more accurate MDD classification, and provide a more precise and reliable method for sMRI-based MDD diagnosis.