Functional State Analysis Of ASD Brain Based On Dynamic Graph Convolutional Neural Network

Autism Spectrum Disorder(ASD)as a category of mental disorders,has long lacked a clear brain pathology and exhibit abnormalities in functional connectivity,involving numerous brain regions.In this thesis,we propose a dynamic map convolutional neural network model to effectively represent brain information,and investigate the different functional connectivity and potential biomarkers in ASD patients.functional connectivity and potential biomarkers in ASD patients.The main contents are as follows.To address the representing brain information without considering its topology in dynamic brain connectivity studies,graph neural network approach is used to build a Brain Graph.Considering Regions of Interest(ROIs)at different locations or subjects with specific identities as graph nodes,a new graph convolutional layer was selected to aggregate the features of different nodes using edge features that indicate the degree of correlation between two nodes.Given that fixed graph edges cannot effectively capture node information,we proposed a dynamic edge convolution layer,which extracts the top-K node information and edge features by calculating the Euclidean distance between node features,and then aggregates them according to the graph convolutional layer.Based on this,the node aggregation matrix is embedded by maximum and mean pooling to improve the clustering efficiency and representation accuracy.In addition,the proposed model is validated in simulated data.To explore the dynamic connectivity patterns in ASD,we obtained graph embedding of the resting state functional magnetic resonance imaging(rs-f MRI)data of 136 ASD and 152 controls using a dynamic graph convolutional neural network model.On the graph embedding information,we performed hard clustering and fuzzy meta-state analysis to obtain states and meta-states.Meanwhile,we calculated the occurrence frequency of each functional state,four graph theory metrics,functional connection strength and four meta-state metrics.After that,the two-sample test and FDR correction were used to determine whether there were significant group differences in functional states and meta-state metrics.Subsequently,the correlations between the measures and phenotypic information were analyzed.The results show that our algorithm outperforms other competing methods and identifies significant biomarkers present in patients with ASD that are partially different from previous ones,as well as features such as diminished kinetics of some meta-state indicators.