Deep Learning of Automatic Encoder Based on Attention for ADHD Classification of Brain MRI

Attention Deficit Hyperactivity Disorder (ADHD) is one of the most common psychiatric disorders. For neurodevelopmental disorders, an increasing number of studies have combined magnetic resonance imaging (MRI) and deep learning to diagnose and explore ADHD. In this paper, we proposed an encoder network based on 3D attention to automatically diagnose ADHD. Our experiments were carried out based on functional Magnetic Resonance Imaging (fMRI) images from ADHD-200 data. The model implemented fine-tuning techniques based on self-encoder, which reduced the heterogeneity of fMRI data from different sites and increased the generalization ability of the model. The model introduced attention mechanisms to achieve better classification performance with an accuracy of 76.4%, while it identified important brain regions for ADHD classification. Our method could provide new insights into the understanding the diagnosis of ADHD.