Automatic Autism Diagnosis Model Based on a 3D Convolutional Neural Network and ReHo

Abstract:
Autism is a pervasive developmental disorder that combines specific pathologies in cognitive function, language function, and interpersonal and social communication, resulting in significant difficulties in social adjustment. It has been shown that there is an abnormality in the local consistency of the brain in autistic individuals compared to normal individuals. This study proposes a deep learning classification framework based on a ReHo brain map with a three-dimensional (3D) convolutional neural network to differentiate autistic individuals from normal individuals. In this study, after downloading the data of 1,112 subjects from the Preprocessed Connectomes Project (PCP) public dataset, its preprocessing methods included removing the first four volumes and slice timing correction, motion realignment, and intensity normalization. Finally, 3D ReHo data were obtained for 1,035 subjects (505 ASD patients and 530 normal subjects). For the characteristics of 3D ReHo data, instead of converting 3D brain map data into three multi-channel two dimensional (2D) images for classification using 2D CNN, this paper directly used a 3D convolutional neural network to receive 3D brain map data, which not only preserves the integrity of the data but also can derive local spatial features. In the end, a test loss of 0.0086 and a test accuracy of 1.0000 were obtained. Compared with traditional machine learning algorithms, this study dramatically improves classification accuracy and eliminates the need for complex feature selection manually. Moreover, due to the high accuracy of this study, it may also provide a reference for computer-aided diagnosis of ASD in the future to help clinicians in the initial screening of autism.