Research And Implementation Of Semantic Segmentation Of Echocardiographic Images

Cardiac ultrasound imaging is a valuable imaging tool in clinical practice,while its semi-automatic or manual annotation task is time-consuming and operator-dependent,which adversely affects the accuracy and efficiency of clinical diagnosis.In recent years,researchers have begun to use deep learning method to carry out automatic segmentation of ultrasonic images.However,a variety of inherent limitations increase the difficulty of segmentation,such as: low signal-to-noise ratio and speckle of echocardiographic images hinder the robustness of the segmentation method,low image contrast between the blood pool and the myocardium makes it difficult to determine the contour of the ventricle,and limited number of echocardiographic images.Furthermore,the existing convolutional neural network fails to make full use of the extracted information for the semantic segmentation tasks and ignores the semantic relevance between pixel pairs,so that the segmentation accuracy of the parts of interest needs to be improved.Based on the above reasons,this thesis explores the application of deep learning algorithm in multi-objective semantic segmentation of echocardiographic images.To quickly and accurately achieve the task of automatic segmentation of the interested parts in the ultrasound cardiac images,so as to reduce the burden of cardiologists.The main work contents are as follows:(1)In this thesis,a pseudo-label generation network for echocardiographic images is established.Based on consistency regularization of the important theory of pseudo tags,reliable pseudo tags are selected.Then,from the perspectives of pixel and whole image processing,the generated results of pseudo tags are optimized to generate higher quality pseudo tags.(2)In this thesis,a two-branch multi-objective semantic segmentation network is proposed to improve the segmentation accuracy of the parts of interest in echocardiographic images from the perspective of semantic relevance and spatial structure.Feature maps containing rich deep semantic information are obtained in the branch of semantic information extraction path.The semantic correlation between pixel pairs in the local areas is acquired and strengthened,so that it can be used as supplementary information to make residual connection with the primitive underlying feature maps,so as to enrich the semantic information contained in the feature maps.In the branch of spatial information extraction path,the innovative end-to-end spatialchannel attention module is used to make the features of interested regions more prominent and those of disinterested regions suppressed,so that the model can better capture the parts to be segmented.(3)In this thesis,a multi-objective semantic segmentation system for echocardiographic images was established.The target population of this system is doctors in clinical scenarios,and the system is simple and easy to operate.It can predict the results in real time under the condition of ensuring the segmentation accuracy,so as to reduce the workload of cardiologists and help doctors to better carry out research in related medical fields.