Research On Key Techniques Of Left Ventricular Myocardium Segmentation In Cardiac MRI Images Based On Full Convolutional Network

With the rapid development of modern medical imaging technology,there are more and more studies on left ventricular segmentation of MR(Magnetic Resonance,MR)images of the heart.The estimation of various cardiac functional parameters plays an important role in the diagnosis and treatment of related diseases.Traditional segmentation of left ventricular regions often has the disadvantages of low segmentation accuracy or extremely low segmentation efficiency caused by unautomatic segmentation.With the recognition of deep learning,especially the application of convolution neural network in medical image segmentation,the accuracy of the left ventricular region segmentation of cardiac MR images has been greatly improved,and the automatic segmentation has been realized.In this paper,the key technologies of left ventricular segmentation in cardiac MRI images based on full convolutional network are researched,and in view of the deficiency of U-Net network segmentation results,several key technologies are proposed to improve the network.The main work and innovation of this paper are as follows:(1)A dual-channel left ventricular inner and outer membrane segmentation network based on codec structure for cardiac MR image is designed.The batch normalization technique is adopted,and an optimizer with adaptive learning rate and a dual-channel three-label loss function training network are proposed.According to the characteristics of the inner and outer membrane,different up-sampling methods are adopted to realize the left ventricular segmentation of the cardiac MR image.In the experimental part,the publicly available left ventricular segmentation data set is introduced,which is randomly divided into training set,test set and verification set,and used as the training and test data of this paper.The U-Net network achieves Dice coefficients of 0.9098 and 0.9292,Jaccard coefficients of 0.8536 and 0.8843,Sensitivity coefficients of 0.9201 and 0.9314 and Precision coefficients of 0.9032 and 0.9374 in the inner and outer films of the test set,respectively.(2)In order to solve the problem of small segmentation area of left ventric in cardiac MR images,a hybrid attention mechanism is proposed to suppress the influence of non-left ventricular regions on the segmentation results.According to the difference between intima and adventitia,different attention mechanisms are used to train dual-channel networks.Because the U-Net network is difficult to extract long-distance information,the intimal area is smaller and small targets are difficult to extract,so in the intimal network,the self-attention mechanism is introduced in the bottleneck layer of the U-Net network,and in order to obtain fast and accurate segmentation,the cross-attention mechanism is introduced at the jump junction.In the outer membrane network,the gated attention mechanism is introduced at the jump joint of the U-Net network to achieve accurate segmentation.In the test set,the inner and outer membranes of the respective networks obtained Dice coefficients of0.9252 and 0.9400,Jaccard coefficients of 0.8619 and 0.8916,Sensitivity coefficients of 0.9230 and0.9466 and Precision coefficients of 0.9292 and 0.9402,respectively.(3)Aiming at the problems of U-Net network,such as limited receptive field,loss of image details,poor segmentation of irregular left ventricular shape image,complex model,large number of parameters and slow network convergence,a left ventricular segmentation technology based on dualchannel parallel dilated convolution network is proposed.In order to expand the receptive field of the U-Net network to the left ventricle,five parallel cavity convolution with different void ratios are introduced into the intimal and adventitia networks.The dilated convolution does not increase the amount of computation and does not reduce the size of the feature graph.In order to reduce the model complexity and reduce the loss of image details,the network depth is set to 3.This reduces the number of model parameters,and the network convergence speed is increased at the same time.This achieves more accurate left ventricular segmentation,especially for images with irregular left ventricular inner and outer membrane shape.The experimental results show that the Dice coefficient,Jaccard coefficient,Sensitivity coefficient and Precision coefficient of the improved network are all improved.