Classification and Biomarker Exploration of Autism Spectrum Disorders Based on Recurrent Attention Model

Abstract

It has become a mainstream to realize the accurate classification of diseases by using deep learning. However, due to the opaque learning and diagnosis mechanism of these models, it is difficult for doctors to believe their diagnosis results and obtain more useful information from these models. According to the brain structural Magnetic Resonance Imaging (sMRI) data of autistic patients, this article proposes an Autism Spectrum Disorders (ASD) classification model based on Recurrent Attention Model (RAM). On this basis, a detailed data analysis is made according to the focus area of the model. To solve the poor convergence of the Policy Gradient (PG) algorithm used in traditional RAM, a Deep Deterministic Policy Gradient Recurrent Attention Model (DDPG-RAM) model and a priority experience replay algorithm based on Gaussian sampling are proposed. The internal information such as Time Difference (TD) error and reward value of training samples is fully utilized. Experimental results show that the algorithm improves the classification accuracy, convergence and stability. Finally, according to the attention block and attention center from the output results of the model, the subcortical tissue of autism group and control group were analyzed. Attention block mainly concentrated in the right thalamus and right caudate nucleus, while the attention center mainly focused on the right white matter and right cerebral cortex. There were significant differences in volume and voxel values of the above regions between the autistic group and the control group. Most of these regions are related to behavioral decision-making, self-learning, communication and some cognitive functions, which provides the possibility that brain structural abnormalities lead to behavioral abnormalities in autistic patients. The model proposed in this article can provide a new method for doctors to find biomarkers.