心脏疾病的高发病率和病患死亡率严重威胁着人类的健康，如何运用精准的检测手段自动对医学影像进行处理，准确及时地发现心脏功能异常具有重要意义。然而，由于心肌与周围组织反差小、影像数据分辨率低、人体运动伪影以及病变结构异常等因素，致使对心脏疾病的自动诊断是一项具有挑战性的工作。针对该问题，本文基于医学影像数据对计算机辅助诊断心脏疾病的相关工作进行了深入研究。主要包括：基于心脏MR影像（Cine Magnetic Resonance Imaging，Cine MRI）数据，通过改进的全卷积深度神经网络，对心脏不同结构进行分割，进而通过特征抽取，实现心脏疾病的自动诊断；针对CT血管造影图像（Computed Tomography Angiography，CTA），在对心脏结构、冠脉进行分割之后，基于冠脉的轮廓形变信息，对冠脉狭窄进行检测与量化评估。具体内容如下： 1）基于全卷积深度神经网络分割与诊断框架，对心脏心室和心肌进行了定量及定性分析，实现了对心脏的分割和心脏疾病的自动诊断。具体流程包括：4D心脏区域分割，心脏特征提取，分类器训练。在分割后的心脏数据中，提取和计算心脏结构中的多种特征，之后进行分层抽样得到训练集。基于5倍交叉验证分数估计，对集合分类器中的组成分类器进行选择。第一阶段的分类器根据多数投票原则来进行心脏疾病判别，结合第二阶段的专家分类器，对心脏Cine MR图像进行分类，将心脏图像根据特征信息自动分为五类，包括：扩张型心肌病、肥厚型心肌病、既往心肌梗死、右心室异常及正常，从而实现了对心脏疾病的自动辅助诊断。 2）构建并优化了冠脉狭窄的判定Pipeline，包括：血管分支预处理，血管行进方向提取及三维重建和狭窄区域确定与病变等级评估。为对冠状动脉狭窄进行更准确的检测与量化，本文提出了一个新的基于截面形变几何信息的冠脉狭窄双重判定方法。首先，在三维冠脉CTA中，用骨架化方法抽取血管中心线，记录中心线上体素点的位置信息。其次，通过维度变换重建经过各体素点的血管横截面。然后，基于二维血管截面轮廓的形变几何信息对冠脉狭窄进行双重判断，并自动标注出可疑狭窄位置及相应的狭窄等级。最后，结合近邻截面的狭窄信息，对冠脉狭窄进行最终的检测与量化。所提算法充分考虑了中心线位置、狭窄位置及血管几何形状等信息，实现了血管狭窄的精准鲁棒判定。为验证算法有效性，将本文算法实验结果与基于血管截面面积的检测方法和基于血管半径的检测方法进行比对。结果表明，所提算法能够更准确直观地对冠状动脉狭窄进行检测和量化，并具有良好的鲁棒性。

The high incidence of heart disease and the mortality of patients are a serious threat to human health. How to use precise detection methods to automatically process medical images and accurately and timely discover cardiac dysfunction is of great significance. However, due to factors such as small contrast between the myocardium and surrounding tissues, low resolution of image data, artifacts of motion and abnormal structure of the lesions, automatic diagnosis of heart disease is a challenging task. In response to this problem, this paper has conducted in-depth research on computer-aided diagnosis of heart disease based on medical image data. It mainly includes: based on Cine Magnetic Resonance Imaging (Cine MRI) data, through the improved fully convolutional neural network, the different structures of the heart are segmented, and then the heart disease is automatically diagnosed by feature extraction; for CT angiography images (Computed Tomography Angiography (CTA), after segmentation of the heart structure and coronary artery, based on the contour deformation information of the coronary artery, the coronary stenosis was detected and quantified. The details are as follows: In this article, the related work of computer-aided diagnosis of heart disease based on medical imaging data is deeply studied. After segmentation of the coronary arteries and cardiac structures, in MR images, an automated assisted diagnosis of heart disease was completed based on an improved fully convolutional network. And the health status of the coronary arteries was evaluated based on the contour deformation information of the coronary arteries in CT images. The research results have important theoretical and clinical value for the clinical diagnosis and treatment of heart disease. The details are as follows: 1) Based on the fully convolutional network segmentation and diagnostic framework, the heart ventricle and myocardium were quantitatively and qualitatively analyzed, and the segmentation of the heart and the automatic diagnosis of heart disease were realized. The specific process includes: 4D heart region segmentation, cardiac feature extraction, and classifier training. In the segmented cardiac data, various features in the cardiac structure are extracted and calculated, and then stratified sampling is performed to obtain a training set. The component classifiers in the set classifier are selected based on a 5-fold cross-validation score estimate. The first stage of the classifier performs heart disease discrimination according to the majority voting principle, and combines the second stage expert classifier to classify the cardiac Cine MR images, and automatically classifies the heart images into five categories according to the characteristic information, including: dilated cardiomyopathy. Hypertrophic cardiomyopathy, previous myocardial infarction, right ventricular abnormalities and normal. Thereby an automatic auxiliary diagnosis of heart disease is achieved. 2) Pipeline was established and optimized for coronary stenosis, including: vascular branch pretreatment, vascular travel direction extraction, three-dimensional reconstruction and stenosis determination with lesion grade assessment. In order to more accurately detect and quantify coronary artery stenosis, this paper proposes a new dual-determination algorithm based on cross-sectional deformable geometry of blood vessels. First, the centerline of the coronary artery is extracted and position information of voxels belonging to the centerline is recorded. Then, vessel cross-sections covering the voxel points along the centerline are reconstructed by dimensional transformation. Coronary stenosis can be detected by dual-determination using the two-dimensional cross-sectional profile of the blood vessel combined with the deformable geometry information. Possible areas of stenosis and their corresponding stenosis level can be automatically computed and marked. Furthermore, the final detection and quantification of each stenosis is performed in conjunction with proximal section information. The proposed algorithm accurately and robustly quantify coronary artery stenosis, while fully considering the position of the centerline, the position of the stenosis, and the geometry of the vessels. The proposed cross-sectional deformable geometry-based method is compared with cross-sectional area- and radius-based methods, showing that the proposed algorithm can detect and quantify coronary artery stenosis more accurately and intuitively, as well as provide satisfactory robustness.