Study On Deep-learning Based Diagnosis For Children With Autism Spectrum Disorder

Autism Spectrum Disorder(ASD)is a serious neurodevelopmental disorder,which contains three core symptoms: social interaction disorder,social communication disorder,stereotyping and repetitive behavior.Diagnosis for ASD mainly depends on experienced physicians,including comprehensive psychiatric examination and necessary psychological assessment,which is not only time-consuming,but also susceptible to subjective factors.Recently,many studies have begun to use biometric signals such as magnetic resonance imaging(MRI)or electroencephalogram(EEG)to assist the diagnosis of ASD,which enhances the objectivity of diagnosis,but the acquisition of these signals is limited by precise medical equipment and strict indoor environment.In order to promote the convenience of ASD children’s diagnosis,inspired by the abnormal gaze pattern of ASD children,this paper propose two deep learning-based ASD children’s video classification algorithms without strict restrictions.(1)In the first stage,136 ASD children and 136 Typically Developing(TD)primary school children’s gaze behavior videos were collected,which are called basic-dataset.An ASD children’s video classification algorithm based on eye tracking is proposed,which achieves 92.6% accuracy on basic-dataset.(2)In the second stage,135 videos of TD kindergarten children were added.In order to maintain the samples balance,135 videos were extracted from the basic-dataset of ASD children and TD primary school children,and merged into extended-dataset,resulting in 405 videos totally.Aiming at the shortcomings of the first stage algorithm,the algorithm of ASD video classification based on gaze estimation is proposed this stage.With the algorithm,AttentionGazeNet,a convolutional neural network with attention mechanism,is proposed for the first time in the field of gaze estimation.After using this network to acquire accurate gaze estimation on extended-dataset,the method of histogram is applied to calculating the features of both screen’s gaze area and head posture.Finally,the features are fed into Support Vector Machine(SVM)for three-classification.Experimental results show the 87.9% classification accuracy on extended-dataset.And for ASD children,the sensitivity and specificity is 88.1% and 90.7% respectively.This paper provides an objective and effective diagnosis method to address the difficult of ASD children’s diagnosis,hoping to provide help for the early diagnosis of ASD children in underdeveloped areas.