4DCMR Registration And Segmentation Based On Multi-Task Joint

Cardiac Magnetic Resonance(CMR)image registration and segmentation can effectively assist clinicians in the diagnosis and prediction of cardiovascular diseases.The purpose of timeseries image registration is to find the spatial alignment of sequence images and target images;the purpose of time-series image segmentation is to extract the spatial structure of the target full sequence frame from the region where the corresponding organ or tissue is located.With the development and application of deep learning in the field of medical images,the accuracy of registration and segmentation have been greatly improved.However,due to the heart movement,frame and inter-frame structure information does not correspond completely,which puts forward higher requirements for CMR registration and segmentation.In order to solve the above problems,the main work of this thesis includes:(1)This thesis designs a temporal CMR registration network(Dense Sampling Temporal Pyramid Network,DS-TPN)based on multi-step fusion.The network relies on the pyramid cascade structure to integrate the spatial-temporal characteristics and increase the spatialtemporal sensing domain of the network;the adjacent dense sampling interpolation loss is designed at the top of the pyramid to further constrain the reliability of deformation in the shortstep registration process.This thesis has done experiments on two temporal CMR datasets,this thesis verifies the accuracy and reliability of DS-TPN in time series CMR registration.Compared with other methods,the best accuracy and smoothest deformation field are obtained.(2)This thesis designs a temporal CMR segmentation network(Spatial-temporal Attention Cyclic Network,SACN)based on multi-task joint learning.Through the network structure of spatial-temporal motion attention,the network uses the DS-TPN optimized deformation field as temporal feature for spatial-temporal fusion;at the same time,the continuous motion loss is designed to obtain smoother and excellent continuity constraints for intermediate frames;The updated multi-task joint network structure makes the outputs on both sides of the network get the effect of mutual optimization and gradually converge to the optimal interval.This thesis also conducts quantitative and qualitative experiments on two temporal CMR datasets,and finally obtains the most accurate time-series full-sequence frame segmentation results compared to other methods.(3)This thesis develops a PyQt5 desktop application software Medical-4DCMR system based on the above-mentioned time series CMR registration and segmentation deep learning algorithm,which serves the daily work needs of medical image processing practitioners.The design of the system follows the MVC model;functionally customized time-series image visualization,built-in FTP transfer file,one-click operation of deep learning registration,segmentation process and can complete image saving and index calculation,and finally achieve a low-coupling,Lightweight desktop application software with high ease of use.