Research On Classification Method Of Autism Spectrum Disorder Based On Functional Brain Network And Graph Neural Network

The accurate diagnosis of autism spectrum disorder,a common mental disorder in children,has always been an important task in clinical practice.In recent years,science and technology have promoted the rapid development of artificial intelligence.Modeling brain imaging data to diagnose brain diseases and analyze pathogenesis have become the focus of researchers.Methods based on functional brain network and graph neural network has shown powerful performance in brain disease diagnosis.Combining brain image data and prior knowledge to construct the network structure can help understand the operation mechanism of the brain and improve the accuracy of auxiliary diagnosis of diseases.However,there are still challenges in constructing "ideal" functional brain network based on extracted data.Moreover,it remains unclear whether and to what extent the nonEuclivian structure of different brain networks affects the performance of graph neural network-based disease classification.In view of the above problems,the main work of this thesis is as follows:1.In this study,the resting-state functional magnetic resonance imaging data were preprocessed and the average time series of brain regions were extracted after a series of operations such as slice timing correction,head movement correction and skull separation.The time series will be taken as the main data for subsequent research.2.In order to better encode the connections between brain regions,this thesis proposed a new brain network construction method named Pearson’s Correlation-based Spatial Constraints Representation to estimate the brain network structure.Based on Pearson correlation,this method introduced the prior knowledge that brain regions with similar spatial distances are more likely to share similar connection topologies.It not only takes the distance between different brain regions into account,but also attempts to model highorder correlations(e.g.,the correlation among different edges or correlations’ correlation).3.In order to evaluate various brain network structural models,the two-sample t-test and Lasso algorithm are first used for feature selection,and then the most commonly used classification method—SVM is used to classify the selected features.Experimental results show that the proposed brain network construction method achieves the accuracy of67.28%,the AUC area of 0.7070,the sensitivity of 60.98% and the specificity of 72.34%.The results showed that spatial information may be an important biological mechanism affecting brain diseases.4.In view of the fact that graph data can contain more information,this study uses two variants of graph neural network — graph convolutional network and graph attentional network for auxiliary diagnosis of autism.During the experiment,the graph structure was constructed by taking the subjects as nodes and the selection of sites and gender as the edge relations between nodes.The results showed that the graph convolutional network achieved the accuracy of 69.50%,the AUC area of 0.75,the sensitivity of 68.24% and the specificity of 71.06%.The graph attention network achieved the accuracy of 70.83%,the AUC area of 0.73,the sensitivity of 69.37% and the specificity of 70.48%.5.In view of individual differences between subjects,the structure calculated by the brain network construction method is used as the graph structure for auxiliary diagnosis in the graph attention network,and the batch normalization optimization network structure is used.The experimental results show that the accuracy of 72.40%,the AUC area of 0.7580,the sensitivity of 71.15% and the specificity of 75.00% are achieved when the functional brain network construction method proposed in this study is graph structure.In this thesis,extensive experiments on functional brain network construction methods and classification framework were conducted on the ABIDE I dataset(n=871).The results demonstrated the superiority of Pearson’s Correlation-based Spatial Constraints Representation method,and different brain network has certain influence on the performance of the classification results based on graph attention network framework.This will help facilitate patient-control separation,and provide a promising solution for future disease diagnosis based on the functional brain network and graph neural network framework.