Research On Graphic Convolutional Neural Network Method For Brain Network Classification And Its Extension

Recently,artificial intelligence has developed rapidly.The diagnosis and analysis of brain diseases based on brain image data has attracted the attention of many doctors and artificial intelligence researchers.From the brain image data,the corresponding brain network can be constructed.Much of the research has focused on subgraph mining techniques.However,the potential structure of the brain network is very complex.It is impossible to capture the highly non-linear structure in the brain network only by using the superficial model of subgraph mining.It is of great significance that how to use artificial intelligence technology in non-linear data mining and analysis based on brain network to help doctors understand,recognize and prevent various brain diseases which are currently difficult to cure.Based on the above challenges and opportunities,the main research work of this paper is as follows:First,aiming at the shortcomings of that the existing convolution patterns can not effectively capture the local patterns of nodes in graph data as well as how to apply deep learning to the mining and classification of graph topological data.In this paper,a graph local pattern suitable for convolution neural network is derived by using the direct connection between the graph node and the neighboring nodes.Then,a row convolution filter is defined based on the graph local pattern.In order to ensure that the graph topology information is not lost during the network layer stacking process,the corresponding structure preserve pooling operation is defined.It solves the serious over-fitting phenomenon caused by the loss of too much graph structure information when the row convolution neural network is stacked.Finally,the brain disease diagnosis classification experiments on real brain network datasets show that the proposed method achieves the best results in all representative methods.And the extended experiment also shows that the method has the characteristic semantic expression function,which can effectively mine the ROI related to brain diseases in the brain network.Second,most of the training samples in real life are covered by undetectable noise,especially the assumption that medical image data has different central distribution noise.In order to make better use of limited and valuable medical data,this paper proposes a deep integration learning framework based on sample noise distribution.The framework assumes that the noise of the sample has random strength and direction.Under this assumption,this paper analyzes the generalization error bounds of single learners.The analysis shows that in the actual learning task,the influence of medical image noise on the single learner is not enough to destroy the learning performance of the learner which means the generalization error is less than 0.5.At the same time,single learners learned based on multiple sample center datasets have some differences.Therefore,integration through generalization errors and difference metrics can effectively improve the generalization ability of ensemble classifiers.Experiments show that our ensemble learning framework can further improve the classification accuracy of brain disease diagnosis.