Cardiac MRI Images Based On Semi-supervised Learning Segmentation Method Research

Heart disease is one of the top causes of death in the world,and cardiac MRI is a noninvasive imaging technique that does no harm to the human body.Compared with CT,it does not need ionizing radiation,and it can directly see the ejection fraction,stroke output,heart structure and other important contents of the heart.Therefore,cardiac MRI image is currently regarded as the gold standard of quantitative cardiac analysis.However,accurate segmentation of these images is a challenging task due to the characteristics of heart images such as artifacts,large organ gap,low resolution of short axis and difficulty in acquiring annotated data.Therefore,the current segmentation methods mainly rely on manual segmentation.In view of the problems encountered in cardiac MRI image segmentation,this thesis designs two effective semi-supervised segmentation algorithms for cardiac MRI images.The main work is as follows:(1)A Dual-stream Semi-supervised Network(DSS-Net),which is supervised by boundary sensing network and image disturbance consistency network,is proposed.The boundary awareness network enhances the perception ability of the boundary information through the combination of sign distance function and inverse transformation function.By using several auxiliary decoder structures to generate a variety of pixel-level disturbance graphs,the generalization ability of the image disturbance consistency network is improved.An uncertainty sensing module composed of correction functions and regularization terms is designed to help the image disturbance consistency network screen out safe and reliable pseudo-tags.The experimental results show that DSS-Net achieves better segmentation effect than supervised learning network and mainstream semi-supervised learning network on LA data set and ACDC data set with a large amount of unlabeled data.(2)Based on Mean-Teacher certainty,a Dual Teacher Un-guidance Model(DTU)is proposed.The following improvements are made on the basis of Mean-Teacher model.First,the data enhancement concept combining strong/weak enhancement is applied to the student-teacher model,so that the student model can learn from the teacher model more comprehensively and enhance the generalization ability of the model.Secondly,the output entropy of pseudo-annotation is minimized to enhance the stability of pseudo-annotation generation.Finally,the method of screening pseudo-annotations by fixed threshold is abandoned,and an adaptive uncertainty perception method is proposed to reasonably screen pseudo-annotations according to different data enhancement methods.Experimental results on LA and ACDC data sets show that the proposed method can reasonably use unlabeled data and achieve high performance with few labeled data.