Research On Sex Recognition Of Brain Network Using Graph Convolutional Network

The study of sex differences in the human brain is a hot direction in brain science research.Studying sex differences of brain networks is of great significance both for exploring the process of life development and for providing specific treatments for sex-related neurological diseases.In recent years,due to the significant progress achieved by the graph convolutional network(GCN)in the analysis of irregular graph data,more and more studies have used GCN in the research of brain networks.In this thesis,GCN is used for sex recognition of brain network to solve the shortcomings of existing methods.The main contributions include:(1)There are two problems for the current studies.One is that they only exploit a single brain atlas template to construct brain networks while ignoring the cerebellar network that contains important sex differences information,and the other is that they use a fixed-scale GCN operator for modeling while lacking attention to the receptive field of the GCN filter.To address these two problems,we design a two-branch multi-scale Chebyshev GCN(TMGCN)model.In this thesis,two complementary brain network atlas templates are used to construct a cerebrum network and a cerebellum network,which are served as the input to the two branches of the model.Additionally,the Chebyshev operator combined with different orders is used to capture the information of neighboring nodes within different ranges of the central node.Finally,the cerebral branch and the cerebellar branch are fused through a trainable weighted fusion layer for sex recognition.Furthermore,a trainable topk pooling layer is built into the model to find discriminative brain regions related to recognition.The results show that TMGCN could effectively integrate GCNs of different orders and fuse the cerebrum and cerebellum networks to complete sex recognition,and the significant brain regions found in the posterior hemisphere and posterior vermis of the cerebellum provides new brain sex differences evidence from the discovery of resting-state functional cerebellum networks using GCN for the first time.(2)Considering that existing GCN-based studies have ignored the effective combination of different spectral and spatial GCN algorithms for modeling and the functional patterns construction of brain networks are often limited to correlations between brain regions,this thesis expands the strategy of two-branch fusion to the combination of spectral domain and spatial domain GCN algorithm.A two-branch model(Spectral-Spatial Domain GCN,SSGCN)is designed that fuses spectral domain Chebyshev GCN and spatial domain Transformer GCN,where Transformer GCN has never been applied to brain network analysis as far as we know.According to the operation characteristics of spectral domain and spatial domain GCN,two cerebrum networks with different functional patterns are constructed for each subject using functional connectivity reflecting the correlation of brain regions and time series reflecting the dynamic changes of cerebral blood oxygen respectively as the input of the spectral branch and the spatial branch of the model.And then the features extracted by the spectral domain and the spatial domain GCN from two functional pattern brain networks are weighted and fused through end-to-end training to complete the final sex recognition.The experimental results show that the SSGCN could fuse the spectral and spatial domain GCN well and effectively utilize the brain network features of different functional modes,and the sex recognition performance outperforms several existing methods.To sum up,this thesis aims to design new GCN models to predict the sex of brain topological networks,tackles the problem in the research on sex recognition of brain network from a more comprehensive perspective,and provides new ideas for the study of sex differences in brain networks;and in terms of computer-aided diagnosis,this thesis may provide new evidence from the discovery of GCN in resting-state functional brain networks for further study of sex differences in the brain networks.