Research On Glioma Grading Method Based On Graph Convolutional Neural Network

Glioma is the most common central nervous system tumor with significant morbidity and mortality.Generally,gliomas are categorized as low-grade gliomas and high-grade gliomas,and the treatment and prognosis of patients with different grades of gliomas vary.Accurate grading of glioma patients before treatment is helpful for clinical treatment decisions and prognosis evaluation.Magnetic resonance(MR)imaging is the most commonly used imaging examination method for glioma,which can generate multiparameter image data.Radiologists usually make corresponding diagnosis after comprehensive analysis of MR images with different parameters.The prediction model based on convolutional neural network can assist radiologists in grading gliomas more accurately and faster by mining deep-level features in MR images.However,convolutional neural network is not able to make full use of the complementary information between multiparameter MR images,and its awareness ability between data is limited,so the auxiliary diagnostic value of multiparameter MR images cannot be fully utilized.In view of the above problems,this thesis carries out the following research work based on graph convolutional neural network for multiparameter MR image-assisted grading of glioma:(1)Aiming at the problem that two-dimensional convolutional neural network cannot simultaneously fuse the slice context information in MR image and the complementary information between different parameters MR images.This thesis proposed a multiparameter-MRI information fusion model based on graph convolutional neural network,named MMIF-GCN.According to the characteristics of multiparameter MR images,the information is designed into a graph.The node information in the graph is defined as the features extracted by the convolutional neural network from image slices at the same position with different parameters,and the edge reflects whether the position between the slices is near.The information interaction and fusion between adjacent nodes are carried out by graph convolution operation.Multiparameter MR image analysis was performed on the public dataset BraTS2020 and the internal dataset GliomaHPPH2018 using a variety of common convolutional neural networks as feature extraction models.The experimental results show that the MMIF-GCN model can effectively fuse the slice context information of image and the complementary information between multiparameter MR images,and improve the classification performance of the model.(2)Aiming at the problem of ignoring the potential relationship information between multiparameter MR images of similar patients in three-dimensional convolutional neural network.This thesis proposed a multiparameter-MRI similarityaware model based on graph convolutional neural network,named MMSA-GCN,The node information in the graph is defined as the features extracted from each parameter MR image of each patient.Three receptive fields containing disease state information were designed to consider the disease state information of patients in the data,and cosine similarity and disease state information were used to calculate the information similarity between nodes.Edges are constructed according to the similarity of information between nodes and the relationship between multiparameter MR images.The similarity awareness and multiparameter MR image information fusion between nodes are carried out by graph convolution operation.Experiments are performed on the public dataset BraTS2020 and the internal dataset GliomaHPPH2018.In the MMSA-GCN model,the area under the receiver operating characteristic curve values are 0.970 and 0.988,and the accuracies are 94.7%and 93.6%,respectively.Experimental results show that MMSA-GCN can accurately capture the similarity information between images and the complementary information between multiparameter MR images,which is superior to the traditional multiparameter fusion method.