Research On Cardiac Image Segmentation Method Based On Deep Unsupervised

There are many different imaging methods for cardiac images in modern clinics,and they usually involve many different types of equipment or different imaging protocols for the same equipment.With the help of various imaging technologies,it can provide clinicians with more abundant and comprehensive information for diagnosis.Among these imaging,Computed Tomography(CT)and Magnetic Resonance Imaging(MRI)of the cardiac provide clinicians with a deeper understanding of cardiac physiology and pathology.Nowadays,image segmentation methods based on deep learning have played an indispensable role in clinical practical applications,and have achieved better segmentation results.However,most of these deep learning segmentation methods only pay attention to a specific modality data,and are insufficient in simultaneously segmenting different modality data.In response to this situation,unsupervised domain adaptation algorithms are widely used in medical image segmentation to assist subsequent clinical doctors in the analysis.This dissertation mainly focuses on the research on the automatic segmentation algorithm of cardiac images of different modalities,and proposes a deep unsupervised cardiac image segmentation method.The relevant research contents are as follows:(1)A cross-modality cardiac image segmentation method based on a deep unsupervised domain adaptation framework.First,for unannotated target modality data,this dissertation uses image synthesis algorithm to synthesize target domain data under the condition of the source domain to form a target domain training set.Immediately after that,input the real target domain and the synthesized target domain and use the annotations of the source domain data as labels to train the target domain segmentation network.The network consists of a shared encoder and a pixel classifier,and the shared encoder extracts the corresponding domain features.In order to ensure that the shared encoder can extract higher-quality domain invariant features,the features obtained by the shared encoder are used to reconstruct and synthesize the source domain data respectively.In addition,the synthesized source domain data can form a source domain training set,and the real source domain and the synthesized source domain are input into the source domain segmentation model to obtain prediction results.At the same time,a cross-domain consistency loss is introduced between the prediction result and the real target domain prediction result to ensure the consistency of the target domain segmentation result to a certain extent.Finally,in order to capture the detailed information of different domains in the segmentation process,this dissertation combines high-level information and low-level information in the segmentation network,thereby further improving the segmentation performance of the target domain.Although the source domain segmentation module can provide effective supervision information for the target domain segmentation,the differences between domains still exist.In order to further reduce the differences between the different domains,this dissertation introduces the maximum mean difference constraint(maximum mean discrepancy,MMD)in the segmentation process.In this dissertation,the above method is applied to solve the problem of multi-sequence cardiac image segmentation,using annotated sequence images to assist the segmentation of unannotated sequence images.(2)Cardiac image segmentation method based on self-supervised learning.In order to solve the problem of segmentation of MR and CT images with more significant distribution differences,we introduce self-supervised learning and deep supervision mechanisms on the basis of the above methods,which can realize unannotated CT(MR)with the help of annotated MR(CT)segmentation.First,the image synthesis and reconstruction module is similar to the above method.Secondly,deep self-supervised learning is introduced in the source domain segmentation module,thereby providing pseudo-labels with rich information for target domain segmentation.At the same time,deep cross-domain self-supervised learning is introduced between the source domain segmentation module and the target domain segmentation module to further mine the original information of the unannotated target domain and assist the target domain segmentation.Finally,improve the above-mentioned segmentation network,introduce a multi-level feature aggregation segmentation network,and spread the high-level information to the low-level step by step to obtain more detailed target domain information.In order to enhance the domain invariance and comprehensiveness of the features used for segmentation,this dissertation applies deep supervision mechanism and adversarial learning ideas in both the source domain and target domain segmentation modules.The effectiveness of this method is verified by experiments on the 2017 Multi-Modality Whole Heart Segmentation(MMWHS)challenge dataset.