Research On Some Key Problems During Left Ventricle Segmentation From Cardiac CTA Images In Small Data Set

Cardiovascular diseases(CVDs)remain the leading cause of death and disability globally.For years,a great effort has been dedicated to the prevention,diagnosis,treatment and research of CVDs.The hardware and software developments have been helping to this health effort with the increasing use of double source cardiac computed tomography angiography(CTA),magnetic resonance imaging(MRI)and other imaging equipment.It is essential to detect the important structures of a left ventricle myocardium from CTA scans in a clinical-decision support system dedicated to improving the early diagnosis of critical CVD diseases.For example,accurate myocardium location will be very helpful for subsequent processing such as cardiac image registration and tissue segmentation,also for understanding cardiac anatomy how to adapt to disease.Computer-aided automatic detection and segmentation methods provide great potential to solve this problem instead of tedious,time-consuming,and poorly reproducible manual delineation.However,this has been a challenging task due to the complex structure of cardiac anatomy,and low image quality such as presence of noise,low contrast and intensity non-uniformity.In this dissertation,based on the analysis on the background and current mainstreams of cardiac image processing works,the automatic cardiac segmentation methods involving left ventricle object detection,left ventricle landmark localization and myocardium segmentation are studied according to the characteristics in the small cardiac CTA sample set.Specifically,the research work and the contributions of this dissertation are mainly presented as follows:(1)In the research work of left ventricular detection,a hybrid method is proposed to detect myocardium by combining region proposal and deep feature classification and regression.The model firstly generates candidate regions using new structural similarity-enhanced supervoxel over-segmentation plus hierarchical clustering.Then it adopts a deep stacked sparse autoencoder(SSAE)network to learn the discriminative deep feature to represent the regions.Finally,the features are fed to train a novel nonlinear within-class neighborhood preserved soft margin support vector classifier(C-SVC)and multiple-output support vector regressor(-SVR)for refining the location of myocardium.To improve the stability and generalization,the model also takes hard negative sample mining strategy to fine-tune the SSAE and the classifier.The proposed model with impacts of different components were extensively evaluated and compared to related methods.Experimental results verified the effectiveness of proposed integrated components,and demonstrated that it was robust in myocardium localization and outperformed the state-of-the-art methods in terms of typical metrics.For instance,the proposed model achieved 0.903,94.1%,90.4%,and 0.93 respectively according to F1,Tpr,Ppv and AUC(Area Under Curve).This study would be beneficial in some cardiac image processing such as region-of-interest cropping and left ventricle volume measurement.(2)In the research work of left ventricular landmark localization,a learning-based landmark localization and identification model is proposed to localize three key landmarks based on salient region sampling,deep feature learning and structural forest regression and classification.The model includes two modules.In the important sample construction model,the salient edge region is built according to human visual characteristics,and the contextual patches are extracted in these regions and fed into the deep SSAE neural network and outputs a compact but more discriminative representation.In the learning module,these deep features are further fed to train the structural random forest regression and classification.A three-layer important weighting is also computed to both increase the accuracy of each decision tree and the diversity of the whole forest.After refinement process,the final output is used to predict the landmark locations.The proposed method is performed on cardiac data set and experimental results validate the reasonability of module settings in the proposed model.Further experiments show that the proposed model achieves mean localization error as 3.87 mm,mean identification accuracy as 96.15%,mean outlier error as 3.85%,which are more competitive in comparison to the state-of-the-art methods.This framework of combining SSAE deep learning and structured random forest is applicable for locating other anatomical landmarks in similar applications.(3)In the research work of myocardium tissue segmentation,a new geometric active contour(GAC)model that integrates high-level shape prior and low-level local intensity statistics is proposed.First,the cardiac-type specific shape priors are learned by structured graph-regularized principle component analysis,so that allows to be faithful to the shape of the desired myocardium.Second,local intensity distribution with inhomogeneity is modeled by cross-entropy energy functional and segmentation-oriented image decomposition.Third,the proposed model is solved as variational level set formulation and a distance regularized energy functional is also introduced for the stability of numerical computation.The model was evaluated on a set of cardiac CTA images with comparison to related shape prior and local region-based methods and multi-atlas joint label fusion methods,and experimental results show it achieves competitive accuracies of segmenting myocardial epicardium and endocardium parts.It also presents advantage of computational burden in comparison to some popular segmentation such as multi-atlas joint label fusion methods.Specifically,for segmenting epicardium,it achieves 0.894,0.823,1.18 mm and 12.8min;while for segmenting endocardium,it achieves 0.957,0.912,0.91 mm and 11.8min in terms of Dice ratio,Jaccard index,mean absolute distance and computation time.Since the framework of the method is general and adaptive,it could be potentially extended to segment similar objects in CT images.