The Research On Feature Enhancement Methods For Brain Networks Based On Graph Convolution

The brain network based on resting-state functional magnetic resonance imaging(rsf MRI)is an undirected graph that includes regions of interest(ROIs)in the brain and the topological structure among brain regions,it can explore the pathogenesis of brain diseases by fully utilizing the features of brain regions and the interaction patterns among brain ROIs.The graph convolutional network(GCN)has demonstrated superior performance in learning brain functional connectivity representations due to its powerful non-Euclidean graph data modeling capability.However,given the complexity of the brain,we find that the application of GCN in brain networks still faces many challenges.For example,how to use GCN to enhance the features of brain networks and improve diagnostic performance of brain diseases,based on considerations of low-order and high-order properties,dynamics,and information complementarity between different levels of networks,is an urgent problem to be solved.To this end,the following work is carried out based on GCN from these characteristics of brain networks:(1)A GCN architecture based on self-attention pooling was proposed to enhance the features of low and high-order static brain networks.Firstly,the entire rs-f MRI time series was divided into multiple sub-windows using a "sliding window" strategy to increase the sample size.Secondly,for each sub-window,low-order static brain network and high-order static brain network were constructed respectively.Finally,these two static brain networks were fed into a graph convolutional model with self-attention pooling to achieve feature enhancement.Since low-order static brain network and high-order static brain network can reflect different static brain network information from different levels,we fuse the features of low and high-order static brain networks from multiple perspectives to further enhance the features of brain networks.The highlights of this method are: 1)Using the "sliding window" method to expand the sample size,which not only effectively alleviates the problem of "highdimensional small samples",avoids overfitting,but also avoids the problem of inconsistent parameters that may occur when collecting experimental data from multiple sources,which is beneficial to improve the performance of GCN.2)Considering both node features and edge information of brain networks,effective feature enhancement was achieved.In addition,the fusion of low-order and high-order static brain networks helps to enhance the expression ability of discriminative features and improve the accuracy of brain disease classification.(2)A feature enhancement method based on GCN was proposed to achieve joint learning of low-order dynamic brain network and high-order dynamic brain network.Firstly,the entire rs-f MRI time series was divided into multiple sub-windows using a "sliding window" method.For each sub-window,low-order dynamic brain network and high-order dynamic brain network were constructed respectively.Secondly,their joint learning was realized by maintaining the consistency between the adjacency matrix of low-order dynamic brain network and the node feature matrix of high-order dynamic brain network.Finally,the selfattention mechanism was introduced to consider the features of both nodes and edges in brain networks.The highlights of this method are: 1)It can achieve feature enhancement for high-order dynamic brain network.2)Joint learning of low and high-order dynamic brain networks was realized while considering the dynamic nature of brain networks and the complementary information of different levels of brain networks.3)Combining self-attention mechanism,not only node features of low and high-order dynamic brain networks but also the influence of edge weights on node features were considered,which is beneficial to capture more comprehensive,disease-related,and discriminative feature information.In this thesis,we apply the two proposed methods to dichotomous experiments in normal subjects and patients with autism spectrum disorders.The experimental results show that the two methods achieve better results compared to relevant control methods,validating the validity of the methods.Furthermore,our proposed method can also be extended to electroencephalogram data to enable the diagnosis of highly heterogeneous neurodevelopmental disorders,and a topic we will focus on in future work.