当前位置: 首页>博士论文>资源详情
基于心脏CT的左心耳腔体分割与腔内血栓自动诊断
中文摘要

随着全球人口日趋老龄化,房颤攀升为心脑血管疾病中高发病种。左心耳是房颤患者身体上心源性血栓主要形成部位,血栓脱落随血液游离到脑部,闭塞脑血管将导致脑卒中。脑卒中具有发病率高、致残率高、死亡率高、复发率高及经济负担高等特点,严重危害人类的生活质量和生命健康。为了定量诊治房颤、有效评估心源性血栓形成和脱落的风险,本文开展了基于心脏CT的左心耳腔体分割与腔内血栓自动诊断的研究。成果主要包括以下四个方面: 左心耳单时相分割。针对左心耳复杂多变的形态与其相关病变存在的强相关性,提出基于“全卷积神经网络+改进的三维条件随机场”的方法,这种方法先利用全卷积神经网络生成左心耳初始分割结果,而后充分考虑体素点间的位置和灰度信息,利用改进的三维条件随机场进行精细分割。该工作为颈部建模提供了基础模型,为多时相分割提供了初期前景种子点,为腔内物质检测提供了跟踪体素点的选取空间。 左心耳颈部建模。针对左心耳封堵手术风险高的问题,提出两种基于优化的自适应算法,实现术前左心耳开口自动检测与颈部自动建模,为封堵器类型、样式、规格的选配提供参考;通过建立标准坐标系和颈部各点封堵支撑张力的测算,预先确定术中封堵器的植入部位、植入姿态,以此制定合理的手术规划来辅助封堵手术。 左心耳多时相分割及房颤自动诊断。针对单时相分割结果不能全面反映运动伸缩的左心耳的真实形态,本文提出在全心动周期内多时相分割左心耳的方法,即附带时间连续性的迭代方式来自动选取前景、背景种子点的三维图割。针对动态心电诊断房颤未能反映房颤发作时左心耳的功能变化,提出利用腔体体积变化比测算左心耳的功能指标,以此构建支持向量机分类器和多因素逻辑回归分析的算法,实现房颤定量诊断和腔内血栓形成的风险评估。 左心耳腔内血栓的自动检测与诊断。针对左心耳腔内血栓检测与诊断,提出了基于时频特征分析腔内体素的三维运动轨迹的方法,构建离散时间傅里叶变换的血栓识别模型,进行小波变换的血栓信号的奇异性检测。最终检测结果包括:血液的形态;不同类型血栓的位置、形态、大小和质地;血栓的龄期;血栓脱落的风险及血栓对腔体的平均负荷。 关键词:房颤;心脏CT;左心耳分割;左心耳颈部建模;血栓检测与识别

英文摘要

As the world’s population ages, atrial fibrillation(AF) has risen to a high incidence in cardiovascular and cerebrovascular diseases. Left atrial appendage (LAA) is the main site for the formation of cardiogenic thrombosis in patients with AF. The thrombi fall off and drift along blood to the brain. Cerebral vascular occlusion caused by thrombi may lead to stroke. Stroke has the characteristics of high incidence, high disability rate, high mortality rate, high recurrence rate and high economic burden, which seriously endangers the quality of life and health of human beings. In order to evaluate the risk of cardiac thrombi formation and exfoliation effectively, the research of LAA cavity segmentation and intracavitary substance automatic detection based on cardiac CT are focused in this paper. It mainly includes the following four aspects: Single-phase segmentation of the LAA. In view of the strong correlation between the complex and changeable morphology of the LAA and its associated lesions, the “FCN+ improved CRF”method was introduced to eliminate the interference of adjacent tissues and organs and to precisely segment single-phase LAA. Based on the segmentation results, the basic model is provided for its neck modeling; the initial foreground and background seed points are provided for its multi-phase segmentation: and the selection spaces of all tracked voxels are provided for its intracavitary substance detection. Neck modeling of LAA. Aiming at the problem of high risk of LAA occlusion, an adaptive algorithm based on optimization is proposed to realize the automatic neck modeling of LAA before operation, which can help to select the type, style and specification of occluder. By establishing the standard coordinate system and measurement of the support tension of the closure at each point of the neck, the implanted position and posture of the occluder are determined in advance. So as to develop a reasonable surgical plan, assisted closure operation. Multi-phase segmentation of the LAA and automatic diagnosis of AF. The LAA is a constantly deforming tissue, single-phase segmentation results are not enough to reflect the true shape of the LAA, hence an 3D Graph-cut approach with time and space continuity is proposed to automatically select the foreground and background seed points, which is used to segment multi-phase LAAs in a entire cardiac cycle. This batch processing improves segmentation efficiency and performance. The multi-phase segmentation results provides a stable imaging basis for subsequent LAA function analysis and disease diagnosis. Aiming at the diagnosis of AF for conventional dynamic electrocardiogram can not directly reflect the functional changes of the LAA during AF, the seven key functional indices of LAA are estimated by volume change ratios based on the results of multi-phase segmentation result. Furthermore, the quantitative diagnosis of AF and the risk assessment of LAA thrombosis are achieved by constructing the SVM classifier and multivariate logistic regression analysis algorithm. Automatic detection and diagnosis of substances in the LAA. In view of the fact that LAA thrombosis is the direct cause of stroke, it is of great clinical significance to detect and analyze the substances in the LAA. A approach based on time-frequency features to analyze the 3D motion trajectories of intracavitary voxels is proposed. This approach includes constructing a thrombus recognition model based on a discrete-time Fourier transform and performing singularity detection of thrombi signals based on wavelet transformation. The final results included: normal blood morphology; abnormal morphology of mild, moderate and severe SEC; location, shape, size and texture of different types of thrombus; age of thrombus: initial jelling, calcified, organic? the risk of thrombi falling off; and the average load of the thrombi in the cavity. The proposed approach based on voxel level can be extended to the detection and diagnosis of other moving tissues or organs. Key words: Atrial fibrillation; cardiac CT; Left atrial appendage segmentation; neck modeling of left atrial appendage; detection and diagnosis of thrombosis

作者相关
主题相关
看过该书的人还在看哪些书