Autism Spectrum Disorder Diagnosis Using The Graph Neural Network

Autism Spectrum Disorder(ASD)is a neurodevelopmental disorder,and its effective identification would be beneficial for medical diagnosis and treatment.Recently,geometric deep learning methods,such as Graph Convolutional Neural Networks(GCN),have been shown to provide a general solution for disease prediction using medical imaging.This paper focus specifically on the application of geometric deep learning methods in predicting autism.In order to integrate more information from subjects,this paper constructs population graphs,where each node encodes the feature information of each subject,and the edges encode the relationship information between subjects.To effectively define the correlation between subjects,this chapter explores various methods for constructing population graphs: Phenotype-Edge(P-Edge),s MRI-Edge(s-Edge),phenotype combined with f MRI-Edge(PF-Edge).To capture the correlation between subjects,this paper first introduces a graph attention network(GAT)that can learn the correlation between node features,and proposes several GAT-based autism prediction model methods based on the constructed population graph: p-GAT,s-GAT,and pf-GAT.Experimental results show that the p-GAT method achieves an average accuracy of 71.6% on the ABIDE dataset,which is 2.1% higher than previous GCN-based autism prediction methods.Moreover,due to the introduction of attention mechanism,GAT-based methods have certain interpretability,which facilitates interpretable analysis in autism prediction tasks.However,by comparing p-GAT,s-GAT,and pf-GAT methods,this paper verifies that the GAT method that ignores edge information lacks certain generalization ability.To improve the model’s generalization ability,this paper further introduces relation-aware attention on top of the GAT method and proposes autism prediction methods based on relation graph attention network(RGAT): p-RGAT,s-RGAT,and pf-RGAT.RGAT-based methods combine node-aware attention and relation-aware attention in a simple additive way,improving the model’s generalization ability.Experimental results show that p-RGAT,s-RGAT,and pf-RGAT have stable prediction accuracies of around 70% for autism.To combine relation-aware attention and node-aware attention in a more rational way,this paper proposes a new graph neural network,the relation graph dual-attention network(RGDAT).After calculating node-aware attention and relation-aware attention separately,RGDAT introduces attention mechanism again to distinguish the impact of node-aware attention and relation-aware attention on overall attention.Experimental results show that RGDAT-based autism prediction models: p-RGDAT,s-RGDAT,and pf-RGDAT,can achieve a prediction accuracy of around 72% for autism,with pfRGDAT achieving an average accuracy of 73.9%.The results demonstrate that the RGDAT method proposed in this paper improves prediction accuracy while maintaining the model’s generalization ability.