Three-dimensional convolutional autoencoder extracts features of structural brain images with a diagnostic label-free approach: Application to schizophrenia datasets

Abstract

There has been increasing interest in performing psychiatric brain imaging studies using deep learning. However, most studies in this field disregard three-dimensional (3D) spatial information and targeted disease discrimination, without considering the genetic and clinical heterogeneity of psychiatric disorders. The purpose of this study was to investigate the efficacy of a 3D convolutional autoencoder (CAE) for extracting features related to psychiatric disorders without diagnostic labels. The network was trained using a Kyoto University dataset including 82 patients with schizophrenia (SZ) and 90 healthy subjects (HS), and was evaluated using Center for Biomedical Research Excellence (COBRE) datasets including 71 SZ patients and 71 HS. The proposed 3D-CAEs were successfully reconstructed into high-resolution 3D structural magnetic resonance imaging (MRI) scans with sufficiently low errors. In addition, the features extracted using 3D-CAE retained the relevant clinical information. We explored the appropriate hyper parameter range of 3D-CAE, and it was suggested that a model with eight convolution layers might be relevant to extract features for predicting the dose of medication and symptom severity in schizophrenia.