Research On Heart Ischemia Deficiency Degree Algorithm Based On Semantic Segmentation Network

Coronary heart disease is the most common form of cardiovascular disease.As one of the non-invasive imaging techniques,myocardial perfusion imaging(MPI)provides a basis for the diagnosis of myocardial ischemia in coronary heart disease.With the patient as the unit,a MPI includes several slices of the myocardium in three axial directions and a bull-eye from an overlooking angle.Professional imaging physicians can obtain the diagnosis result of the degree of myocardial ischemia in patients by analyzing the axial diagram and bull-eye together.However,a large number of studies have been conducted on the prediction and diagnosis of myocardial ischemia in coronary heart disease based on axial section,but few studies have been carried out on bull-eye,ignoring its own information value,and few studies have accurately distinguished the specific degree of myocardial ischemia.Aiming at the bull-eye,this thesis improved the network structure design and fusion of traditional features based on U-Net,to achieve accurate segmentation of the degree of myocardial ischemia loss,so as to locate the necrotic myocardium and provide technical support for subsequent quantitative calculation of its proportion in the whole heart.Firstly,based on U-Net,a branch structure consisting of a multilayer transposed convolution up-sampling splicing module and a four-channel weighted channel attention module was proposed,and the output results of the branch structure were fused with the output results of the main U-Net to achieve accurate segmentation of the deficient parts.The experimental results show that the multilayer transpose convolution up-sampling splicing module can effectively reduce the interference of the severe sparsity degree which is similar to the deficient degree.The four-channel weighted channel attention module can further improve the ability to distinguish between two similarity degrees and the ability to learn target edge details,and retain richer edge details features.Through comparative experimental analysis on self-built datasets,the model presented in this thesis is superior to other models optimized based on U-Net,and Jaccard is 5.0% higher than U-Net.Based on the above research,in order to further improve the segmentation accuracy of the deficient parts,similar to mask generation algorithm and Canny edge detection algorithm are introduced,and use the convolution neural network to extract similar to mask images,edge images,and the characteristics of RGB images,through three fusion location and integrated approach to complete four characteristics of integration,choose the optimal combination model.In order to achieve better deep fusion.Experimental results show that the fusion of similar to mask generation features can effectively reduce the misjudgment and omission caused by the small sample size and insufficient neural network learning features,fusion edge detection algorithm to extract features can further improve the ability of the model to extract the details of the target edge and distinguish the similarity degree.the optimal combination model Jaccard further increased by 2.3%.