Study Of 3D Echocardiographic Segmentation Of Left Ventricle By Fusing Anatomical Structure Priori

Cardiovascular disease is still the disease with the highest mortality in China,in which heart disease accounts for an important proportion.The structural changes of the left ventricle are closely related to the occurrence of cardiovascular diseases.Throughout the anatomy of the heart,the left ventricle delivers oxygenated blood to the entire body.Therefore,left ventricular function is an important measure of heart disease.The measurement of left ventricular function is achieved by segmenting the left ventricular part in different imaging modes,and then calculate the relevant functional indexes.At present,three-dimensional echocardiography can provide more context information,provide more anatomical structure information for clinical experts,and become a popular mode of left ventricular echocardiography.In order to avoid the problem of time-consuming and subjective differences in manual annotation of echocardiography,computer-aided echocardiography segmentation has become a research hotspot in view of its efficiency and consistency.Therefore,this paper is devoted to the study of computer-aided automatic and semi-automatic segmentation of left ventricular three-dimensional ultrasound.Firstly,in order to solve the initialization problem of deep learning model based on small number of samples,and fully utilize the spatial prior information of 3D echocardiography itself,a self-supervised sorting model based on spatio-temporal anatomical prior knowledge is proposed.Based on the sorting model,a self-supervised pre-training mechanism based on the sorting model is proposed and applied to specific ventricular segmentation tasks to improve segmentation performance.The pre-training mechanism based on sorting task proposed in this paper has good generalization ability and can be applied to other sorting modes and more extensive application fields.At the same time,it lays the foundation for the following methods of left ventricular segmentation.Secondly,a residual network segmentation method is proposed.In the aspect of network structure,multi-scale residual module is used to build fully convolutional network(FCN).In the deformation model method,the deformation surface model with anatomical structure constraint is proposed.As a post-processing model of multi-scale residual FCN,the deformable surface model with anatomical structure constraint overcomes the problems in the defect of ventricular base part,left ventricular wall,papillary muscle segmentation,and improves the accuracy of segmentation.Compared with the state-of-the-artsegmentation method,this method has obvious advantages in some indexes.In terms of methodology,the idea of fusing traditional segmentation method,deep learning method and shape priori is promising,which improves the interpretability of the model to some extent.Thirdly,based on the geometric prior integrity of map segmentation method,a multiscale information consistency constraint segmentation method based on deep learning technology and atlas is proposed.The proposed segmentation method integrates the atlas into the deep learning framework to form the deep atlas network,and proposes multi-scale information consistency constraints(including label consistency constraints,volume consistency constraints and adversial consistency constraints).The proposed segmentation method uses a light-weight network to achieve high segmentation accuracy and efficiency.Moreover,the proposed segmentation method can be trained based on the limited training set and achieve better training effect than the state-of-the-art method.The proposed multiscale information consistency constraint improves the accuracy of segmentation from different levels,and finally completes the task of left ventricular segmentation in the complex anatomical environment.It has good research potential in the field of 3D ultrasonic image segmentation based on deep learning.Finally,a 3D echocardiography segmentation method for the whole cardiac cycle is proposed,which uses a single labeled image to achieve the task of left ventricular segmentation in a unsupervised environment.The proposed method uses a registration framework based on deep learning to conduct the task of left ventricular segmentation.The proposed segmentation method uses the spatial deformation network to generate the deformation field between different gray-scale bodies,and uses the gray-scale body similarity measurement and smooth constraint to optimize the network.Based on the single labeled image,the registration method is used to transfer the label between adjacent volume,and then the segmentation task is completed.And this optimization process does not rely on any annotation.Based on the fact that the deformation field is shared between two grayscale objects,a semi-automatic left ventricular segmentation can be realized easily.This segmentation method has good application potential in the field of semi-automatic left ventricular segmentation.