Cardiac MRI Segmentation Based On Semi-supervised Learning

Accurate and efficient segmentation of cardiac MRI is a key step in analysis and diagnosis of clinical cardiac diseases,which can provide important basis for doctors to evaluate the condition of patients with cardiovascular diseases,formulate treatment plans and plan for later rehabilitation.Nowadays,the fully automatic cardiac segmentation model based on deep learning requires a large amount of manually labeled data as training samples to obtain ideal segmentation results,and the labeling them not only requires professional knowledge,but also time consuming.Furthermore,the current segmentation model seldom considers the contour quality of heart,which makes the segmentation results often unsatisfactory.In this paper,we propose an improved semi-supervised segmentation framework based on the Mean-Teachers model to solve the problem above.Additionally,a Fourier ellipse descriptor based contouraware algorithm is proposed to solve the inaccurate results of the cardiac segmentation model at the edge.The main works of the thesis include:(1)To solve the problem of time-consuming and laborious labeling of cardiac MRI image datasets,we propose an improved semi-supervised cardiac segmentation framework based on Mean-Teachers model which designed two branches of network based on U-Net to predict the segmentation results.Additionally,the regularization method of consistency loss is added to the two branches,using a small amount of labeled data and making full use of unlabeled data to improve accuracy of cardiac MRI segmentation.(2)To solve the problem that the accuracy of the cardiac segmentation model under the semi-supervised method needs to be further improved,we propose an improved semisupervised segmentation method named Dual Attention Uncertainty-aware Mean Teachers(DA-UAMT).This method uses the U-Net-based semi-supervised framework as the backbone and designs a shape attention module(SAM)and a channel attention module(CAM)for cardiac segmentation task combined with the idea of attention mechanism.We integrate these two modules into a semi-supervised model in parallel.Furthermore,to improve the accuracy of semi-supervised learning,an uncertainty-aware module based on Monte Carlo Dropout is proposed to select samples with higher confidence in prediction,which ensures that the student model learns from more reliable teacher model predictions.(3)To solve the problem of unsatisfactory segmentation quality at the edge of the current cardiac segmentation model,we add a contour constraint to the heart on the basis of the semisupervised framework,and propose a segmentation method called Contour-aware Mean Teachers(CA-MT),which uses Fourier ellipse descriptors to represent the extracted results and attention map contours step by step,by jointly optimizing the attention map contour loss function and class level contour loss function,and control the number of descriptors during training,so that the network gradually learns the representation method of contour coarse-tofine.The proposed method improves the representation accuracy of the model at the heart segmentation edge.The experiments on ACDC and LASC’13 dataset show that proposed DA-UAMT and CA-MT methods improved the data efficiency and achieved better segmentation performance,which improved by 1-2% in average compared to popular semi-supervised methods.The proposed methods solved the problem of difficulty in data labeling and segmentation accuracy,especially at the edge of heart,which have good practical value clinically.