Research On Cardiac MRI Analysis Method Based On Deep Convolutional Network And Image Registration

The heart is the core organ of the human body.Evaluating the shape of the heart(such as the position of the left ventricle,the volume of key tissues of the heart)and the operation of the myocardium allows doctors to quickly and intuitively obtain valuable information to determine the health of the heart.For this reason,clinicians or anatomists need to manually mark the heart images one by one to obtain the various tissues of the heart.This process is tedious and time-consuming.Computer-assisted cardiac image analysis,as a branch of medical image analysis,can free doctors from manual marking,and has become a research hotspot at home and abroad.In this context,based on the deep convolution network,medical image registration,dictionary learning and other technologies,this thesis has carried out the research of cardiac image analysis methods,such as left ventricular detection in cardiac MRI,segmentation of key cardiac tissues(left ventricular myocardium,left atrium,left ventricle,right atrium,right ventricle,aorta,pulmonary artery),and multi-level visualization of myocardial motion.The main research content and innovation results of this thesis are as follows:In the aspect of left ventricle detection,a CNN left ventricle detection method based on discriminative dictionary learning and sequence tracking is proposed,aiming at the large changes of left ventricle shape and size in cardiac MRI,and some of them are small size regions.The proposed method uses discriminative dictionary learning combined with SLIC algorithm and sequence tracking(for the same individual)to form two adaptive anchor point generation methods for the left ventricular region.The generated adaptive anchors are added to RPN(Region Proposal Network)and Proposal Layer,and realize left ventricular detection through regression and classification.In the aspect of heart key tissue segmentation,a two-stage CNN segmentation method based on dual attention mechanism and metric classification is proposed to solve the problems of complex distribution and close arrangement of tissues and organs,low differentiation of some organs,large size span and unclear edge contour in cardiac MRI.In the coarse segmentation stage,the proposed method combines the multi-level features of the Log-Gabor filter and the channel attention mechanism to enhance the original image,and introduces the dual attention network mechanism to generate the coarse segmentation result.In the fine segmentation stage,the metric classification network is used to optimize the segmentation edge.In the aspect of multi-level visualization analysis of myocardial motion,a multi-level visualization analysis method of myocardial motion based on image registration is proposed to solve the problems of spatial misalignment of cardiac MRI image sequence,uneven change of myocardial motion size,large difference of regional motion and low efficiency of myocardial AHA17 segmentation labeling.The proposed method simulates the complex motion process of the myocardium through multiple transformation methods such as rotation,zoom,and distortion,and uses the information measurement criterion to evaluate the image similarity to obtain the myocardial motion deformation field.The proposed method is used to construct the myocardial AHA17 segmentation label,and the myocardial movement displacement field is visualized and analyzed at multiple levels and multiple perspectives.The method in this thesis is tested on cardiac MRI data sets to verify the rationality and effectiveness of the method.Compared with related methods,it shows better performance and can assist doctors in the clinical judgment of heart health.