Research On Brain Atlas Decoding Algorithm Based On Graph Convolutional Neural Network

As the human nerve center,the brain controls various activities of the human body.Therefore,it is particularly important to understand the structure and operation of the brain.Functional Magnetic Resonance Imaging(fMRI)has been widely used in brain research due to its non-radioactive and non-invasive advantages.Brain decoding is to predict the external stimulus information from the collected brain response activities.In the external stimulus information,visual information is one of the important sources.Decoding the fMRI image based on visual stimulation is helpful to understand the mechanism of brain visual nerve and diagnosis and treatment of related diseases.Brain decoding methods are mainly divided into three categories:early traditional methods,methods based on machine learning,and methods based on deep learning.Early traditional methods require manual construction of mathematical models,but fMRI data contains a large amount of information for three-dimensional slice images,and manual construction of mathematical models is very difficult and the matching degree is not high.Machine learning methods are usually shallow models and require manual feature extraction,which can easily lead to information loss.Commonly used deep learning methods based on convolutional neural networks often ignore the functional correlation feature between different regions of the brain.Therefore,in view of the above problems and the characteristics of fMRI data,this thesis proposes a brain map decoding algorithm based on graph convolutional neural network.The main work is as follows:(1)According to the brain’s visual nerve partition and the response to external visual stimuli,11 regions of interest(ROI)are selected,and then these 11 regions are segmented from the original fMRI image and input to the network for processing,which can reduce information and it also avoids noise interference in non-visual areas of the brain.(2)A deep three-dimensional convolutional neural network is proposed for feature extraction.Because there are 11 ROIs in total,11 three-dimensional convolutional neural networks are used in parallel,which can effectively extract deep-level features,and the simultaneous input of fMRI images at multiple time points ensures that the time-varying information contained in the data is not lost.(3)The graph convolutional neural network is used to extract the functional correlation features between different regions.Considering the problem of over smoothing when the number of layers of graph convolution neural network is too many,so this thesis adopts the residual connection method to make the graph convolutional neural network have a deeper network structure and help to merge different levels of features,so as to improve the accuracy of the network.The algorithm in this thesis was tested on an open data set,and the overall accuracy rate reached 94.37%.Compared with 3D-CNN(3 dimensional convolutional neural networks),C3d-LSTM(3DCNN-LSTM),ResNet(residual neural networks),cGCN(connectivity-based graph convolutional network),the algorithm in this thesis has achieved the best performance.