Research On Medical Image Classification Based On Graph Convolution Neural Network

With the rapid development of information technology,graph convolutional neural networks for medical image classification research have broad development prospects.Building a medical image classification model through a graph convolutional neural network can effectively assist clinicians in speeding up diagnosis efficiency,improving diagnosis accuracy,and reducing the burden on patients.However,there are some deficiencies in medical image classification methods based on graph convolutional neural networks,such as: When a graph convolutional neural network is deeply stacked,the representation of all nodes converges to the same value;The memory increment problem caused by no jump connections in existing deep graph neural networks cannot be solved;The existing composition methods have high randomness and poor interpretability;Inadequate utilization of multi-modal medical image data in the study of brain disease diagnosis.In order to solve the above problems,we propose three methods for medical image classification based on graph convolutional neural networks:(1)A medical image classification method based on a deep graph convolutional neural network.We use the weighted undirected complete graph sparse pruning algorithm for graph structure transformation and propose a nodal self-convolution algorithm based on the fusion of graph structure and channel features to construct a classification model using a dense connectivity mechanism to construct a deep graph convolutional neural network for lung CT image classification.The experimental results show that the deep graph convolutional neural networks perform better than the shallow graph convolutional neural networks,and the classification accuracy is improved by 8%.(2)A graph convolutional neural network medical image classification method based on transfer learning and feature reconstruction.First,we propose a new feature reconstruction algorithm based on image pixel neighbourhood structure.Compared with the traditional feature reconstruction algorithm,the obtained graph structure is more stable,which improves the robustness of the model.Second,based on the transfer learning strategy,a variety of top-layer removal schemes are proposed to extract the pixellevel features of the medical image with the excellent feature extraction capability of deep convolutional neural networks to reduce the training cost and improve the model performance.The final experimental results show that the deep graph convolutional neural network using the feature reconstruction algorithm and transfer learning performs well,with an average classification accuracy of 96.4%,1% higher than the state-of-theart methods.(3)A Multi-modal medical image classification method based on graph attention.We propose a feature fusion method based on graph attention,a 3D feature extraction module,and a 3D feature reconstruction module.We design a medical image classification architecture with multi-channel and multi-modal feature fusion,which can classify medical images according to the fusion features and use the branch structure for additional training.The experimental results show that the multi-modal medical image classification model proposed in this paper is better than other multi-modal classification models,and the classification sensitivity is improved by 6%.This paper has 31 figures,30 tables,and 79 references.