Myocardium Segmentation From Delayed Enhancement Cardiovascular Magnetic Resonance Imaging Of Myocardial Infarction Patients

For myocardial infarction(MI)patients,delayed enhancement(DE)cardiovascular magnetic resonance imaging(MRI)is a well validated technique for the visualization of MI.The myocardium viability assessment with DE MRI can play important roles in diagnosis,treatment management,therapeutic strategy evaluation and prognosis,where myocardium segmentation is a prerequisite for quantitative assessment.However,developing automatic myocardium segmentation methods can be arduous,primarily due to the characteristics of DE images,i.e.the heterogeneous intensity of myocardium,the overlapping intensity range of myocardium and surrounding tissues,and the complex enhancement patterns.The goal of this research is to develop fully automatic approach for accurate myocardium segmentation from DE MRI of MI patients.The major work and contribution of this thesis include:1)A segmentation framework based on multi-component Gaussian mixture(MCGM)model and coupled level set(CLS)is constructed.In this method,the heterogeneous intensity of myocardium is modeled explicitly with MCGM and converted into a relatively uniform probability distribution,based on which the myocardium segmentation is performed.The CLS is introduced to impose a constraint of relatively uniform myocardium thickness,so as to deal with the overlapping intensity distribution between myocardium and surrounding tissues and guide the segmentation of myocardium.For initialization of the segmentation framework,the atlas-based approach is adopted to utilize the anatomy information in the prebuilt atlas.2)A segmentation framework based on guided random walks and sparse shape model is constructed.In this method,the guided random walks algorithm is used to deal with the heterogeneity within the segmented region in a natural way.The shape model based on sparse representation is built,using the data in a public database as the shape templates,to introduce stronger shape constraint and achieve accurate and robust segmentation results.3)A dual-sequence myocardial segmentation framework with registration error compensation is constructed.In this method,the complementary information in the DE MRI and the T2 weighted MRI is combined for unified segmentation.Compared with intensity characteristics,we consider shape information to be a superior approach for complementary information sharing between the two sequences.Considering the myocardium shape discrepancy between the two sequences due to non-perfect registration,an error term is added to explicitly model this difference.To reduce the computation time and achieve better initialization in apical slices,the conditional generative adversarial network(cGAN)is introduced for initialization.In this thesis,33 sets of MRI data were collected from patients with MI,and the segmentation results of the three methods from DE MRI were evaluated.The myocardial Dice similarity coefficients(DSC)of 72.22±6.49%,74.31±7.98% and 78.13±6.22% were obtained respectively,which were similar to the inter-observer differences.These results demonstrate that the proposed methods have potential in clinical application.