Research On Segmentation Method Of Cardiac MRI Image Based On Deep Learning

Cardiovascular disease has become the disease with the highest mortality rate in my country and the number one killer that threatens human life and health.Quantitative evaluation of the results of the segmentation of cardiac MRI images is the best method for the diagnosis of cardiovascular diseases.In clinical medical diagnosis,the cardiac MRI images are often manually segmented by imaging experts,which is very time-consuming and prone to manual errors.Therefore,a fast and accurate automatic segmentation method is needed.The main research content of this paper is to realize the automatic segmentation of cardiac MRI images based on deep learning.This paper first designs an improved Dense U-Net network based on U-Net and Dense Net.In the process of network downsampling,the improved Inception module is added to expand the receptive field of the network and reduce the number of parameters of the network,adding residual connections at the same time improves the accuracy of network training.During the upsampling process of the network,the feature fusion is performed by superposition,which effectively reduces the network parameters and GPU memory usage.The improved skip projection connection is also used to realize downsampling and upsampling to transfer feature information across channels.In addition,this paper also proposes a dual loss function combining the weighted Dice loss function and the weighted cross-entropy loss function,which effectively alleviates the problem of class imbalance in the image and improves the segmentation result.In this paper,a cooperative hospital dataset is established based on the cooperative hospital data,and multiple experiments are conducted simultaneously with the ACDC dataset to further demonstrate the reliability of the network and improved method proposed in this paper through different datasets.Among them,there are 150 cases in the ACDC data set,and 80 cases in the cooperative hospital data set.The average Dice coefficients of the corresponding segmentation results of the left ventricle,right ventricle,and myocardium of the ACDC dataset are 0.940,0.903,and 0.891,respectively.The average Dice coefficients of the corresponding segmentation results of the left ventricle,right ventricle,and myocardium of the cooperative hospital dataset are 0.941,0.890,and 0.888,respectively.The experimental results show that the improved network proposed in this paper has good generalization ability and robustness,and can effectively improve the segmentation performance and accuracy.In order to further improve the segmentation results,this paper proposes an improved segmentation method based on the Dense U-Net network.The network is improved by adding ROI detection preprocessing based on Canny edge detection and Hough transform and post-processing based on maximum connected domain analysis.The improved segmentation method can remove the mis-segmentation in the basal and apical slices,effectively improve the Dice coefficient,and greatly reduce the Hausdorff distance.The average Hausdorff distance of the ACDC dataset was reduced from 9.811 mm to 5.658 mm,and the average Hausdorff distance of the cooperative hospital dataset was reduced from 9.539 mm to 5.691 mm.