Research On Cardiac Magnetic Resonance Image Segmentation Based On Deep Learning

According to the world health organization,Cardiovascular Diseases(CVDs)have become one of the major threats to human health.The advance diagnosis and treatment of cardiovascular diseases has become a hot topic in the world.Cardiac MR images are widely used in the early diagnosis and treatment of cardiovascular diseases because of the advantages of non-invasive and high contrast.Clinicians can obtain important cardiac function parameters by cardiac MR image segmentation.In real-world medicine,however,ventricular segmentation is usually done manually by experienced doctors.Manual segmentation is not only time-consuming and laborious,but also highly subjective,so it is of great significance to study the efficient and accurate automatic segmentation method of cardiac MR images.In this thesis,based on the cardiac short-axis nuclear magnetic resonance image,two improved methods based on deep learning are proposed after analyzing the limitations of the traditional algorithm for cardiac MR image segmentation.The main work is as follows::1?In the traditional image segmentation algorithm,the threshold segmentation algorithm,region growth algorithm and k-means clustering algorithm are used in turn to complete the segmentation task of cardiac MR image.According to the experimental results,the limitations of the traditional algorithm in cardiac MR image segmentation are analyzed.The traditional segmentation method is prone to over-segmentation and under-segmentation in cardiac MR image segmentation,and the accuracy of segmentation results is relatively low..2?Based on the deep learning U-Net network.This thesis introduces a improved framework for the automated segmentation of the epi-and endo-cardial walls of the left ventricle from the cardiac MR images using a fully convolutional neural network similar to the U-net.In order to overcome the problem of class unbalance,an improved loss function is proposed to replace the traditional cross entropy loss function.The traditional loss function is easy to lead to the learning bias of the model.The improved loss function improves the overall accuracy of the model and reduces the learning bias caused by cross entropy.Compared with the traditional method,the improved method has better segmentation precision.3?In the task of cardiac MR image segmentation,this thesis adopts the Dense U-Net as the basic network structure.In order to reduce of the parameter calculation and memory footprint without affecting the quality of the segmentation output.,the skip connections from down-sampling path to up-sampling path used element-wise addition operation instead of concatenation operation and a projection operation was done to match the channel dimensions.By introducing shortcut connection of residual network in the upsampling path,Dense U-Net could better use of information in shallow networks.Additionally,a modified Inception structure was introduced at the initial layer of the network to replace the 3×3 convolution operation and increase the receptive field of the model.Finally,an improved network was obtained.Through experimental comparative analysis,it was found that the improved network improved the segmentation accuracy of cardiac MR image,and achieved good results on Dice and APD indexes.