Research On Heart Segmentation Based On Semi-supervised Deep Learning

Image semantic segmentation is a fundamental recognition task in computer vision.With the development of artificial intelligence,deep learning-based methods are increasingly being applied to this task with good results.Since the success of deep learning methods in image segmentation tasks usually relies on a large amount of training data with pixel-level manual annotation,it is extremely costly compared to other vision tasks such as image classification and target detection.Especially for 3D medical images,even though large amounts of images are easily available,professional annotation of image data is an extremely time-consuming and expensive process,which makes the semi-supervised nature of segmentation methods an important problem worth investigating,i.e.,training segmentation models with good performance using a very small amount of real annotated data as well as additional unannotated data.Cardiovascular disease,a disease with high morbidity and mortality due to abnormal heart function or structural defects,has been a major concern worldwide.With the advantages of no radiation damage to the human body,high quality imaging of tissues,and freedom to choose the profile,MRI has become an important aid for physicians to assess the structure and function of the heart,guide diagnosis,and generate treatment decisions.Effective semantic segmentation of magnetic resonance imaging(MRI)of the heart is an important prerequisite for computeraided medical diagnosis and an important technical basis for tissue structure visualization.Therefore,the main research work of this paper is to achieve the effective segmentation of cardiac MRI images including atria,ventricles and ventricular wall muscles by using a very small amount of labeled data as well as a large amount of rich unlabeled data to achieve the performance improvement of segmentation models through semi-supervised learning methods,which are mainly as follows:(1)A dual-task cross-consistency network-based atrial segmentation algorithm is proposed for the segmentation task of 3D MRI data.The network model in this method is extended with V-Net as the base architecture,and the data perturbation by data enhancement and random noise and the model perturbation operation by adding an auxiliary decoder enhance the overall antiperturbation capability of the model while the feature extraction capability is enhanced.In addition,an implicit dynamic geometric prior knowledge is introduced into the model through the dual training of segmentation and regression tasks,thus making better use of the valid information in the large amount of unlabeled data.The method also utilizes the idea of consistent regularization to optimize the segmentation network by using cross-consistent regularization of the prediction results between the dual decoders and between the dual tasks,which further enhances the segmentation effect and generalization ability of the model.(2)To enhance the utilization of global features of images,two different paradigms of CNN(Convolutional Neural Networks,CNN)and Transformer are efficiently fused for semisupervised learning,and a dual-paradigm cross-guided architecture with uncertainty perception is proposed for more accurate segmentation of ventricular and ventricular wall muscles.The training architecture enables each input image to be trained and finally predicted simultaneously by a CNN-based student segmentation network S and two teacher TA and teacher TB networks constructed by CNN and Transformer,respectively.The teacher TA based on Mean Teacher architecture MT(MT)instructs the student S by means of consistent regular constraints and combines data and model perturbations to enhance the generalization ability of the model;the teacher TB,on the other hand,cross-supervises the student S with each other by means of pseudo-label learning to achieve the instructional effect,so that the advantageous attributes of the two learning paradigms are effectively integrated in the architecture and complement each other.In addition,while the dual teachers each instruct student S,they further achieve mutual communication between teachers and continuously screen the prediction results for uncertainty assessment during the training process,so that the segmentation effect of student model S has stronger accuracy and robustness.(3)In order to alleviate the problem of insufficient capture of 3D cardiac MRI spatial information by 2D network models and based on the purpose of more effective utilization of valid information in a large amount of unlabeled data,a dual-teacher co-teaching cardiac segmentation strategy incorporating multiple views and multiple tasks is proposed.This semisupervised training strategy uses a novel approach to integrate the multi-view collaborative training strategy into the dual-teacher instruction architecture designed in the previous method,and proposes a semi-supervised heart segmentation method based on a 2.5D training approach,which alleviates the problem of 2D networks ignoring 3D image spatial information to a certain extent.In addition,dynamic geometric prior information is also effectively introduced in this segmentation strategy,and the training of different structures and tasks among networks is used to deliver richer and more accurate knowledge information to the student target network,allowing the effective information in the large amount of unlabeled data to be utilized to a greater extent.The method is able to hold good segmentation results with much reduced time and space complexity for both thin-layer and thick-layer cardiac MRI data,and achieves further improvements for both single-category and multi-category segmentation tasks.