【作者】 张杰慧；

【导师】 何中市；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2011， 博士

【摘要】 肺栓塞在西方国家被列为常见的三种心血管疾病之一,具有诊断率低、死亡率高的特点。在中国,肺栓塞长期以来被误认为是一种罕见病,被临床医生严重忽视,但近年来逐渐引起医学界的重视并广泛普及。肺栓塞在临床上无特异性,极易被忽视,未及时诊断和有效诊治会导致大部分的栓塞患者死亡。计算机断层扫描与血管造影技术的出现使现代影像学检查技术成为肺栓塞检测诊断的重要手段,尤其是CT肺动脉造影技术已经成为临床上检测肺栓塞的一种主要途径,配合使用计算机辅助检测技术(CAD)可以进一步提高检查栓塞的准确率和效率。肺栓塞CAD系统的研究经历了近十年,但是远未成熟,在理论和实际应用中都还有许多亟待解决的问题,在国内更是尚未引起研究者的重视。本文基于肺栓塞CAD系统的常规流程,提出了一套新颖的构建肺栓塞CAD系统的技术路线,该路线将纵膈内部的中心型肺栓塞也纳入重点检测对象,并致力于降低肺栓塞检测的假阳性。本文的研究按照CAD系统的技术路线展开,对各步骤进行了深入的研究,所完成的主要研究工作和创新点如下:①为得到左右肺分离的完整肺区域,提出了一种基于解剖学知识的3D肺区域自动分割算法。首先,分析CT图像各组织密度分布采用阈值法进行预分割。其次,针对支气管导致肺区域分割误差的问题,提出采用自顶向下的强规则区域生长法提取气道内腔以避免气道渗漏至肺实质,再膨胀至气道壁,然后提出区域面积分析法和投影法以判断与分离左右肺,并进行肺实质边界修补。实验验证了该算法对CT增强扫描和普通扫描数据的适用性,能有效排除两肺间的气管/支气管通道,在复杂图像的分割实验中也取得了良好的效果。②作为纵膈内肺动脉分割和中心型肺栓塞的工作基础,本文首次将纵膈区域的分割加入肺栓塞CAD系统,并提出了一种3D纵膈区域自动分割算法。纵膈是一个解剖区域,无明显的边界和形状特征,故提出缘点的概念用以描述2D纵膈区域的解剖学结构,并采用基于角点检测的方法提取缘点和基于快速行进的方法提取缘点间路径,同时在2D纵膈分割过程中引入上下文轮廓的位置信息以约束和保障3D纵膈表面的连续性。此外,用与金标准结果间的偏差距离累积分布来客观评估分割效果。实验结果显示,本文基于上下文轮廓的3D纵膈区域自动分割算法能得到完整且平滑的纵膈区域,分割具有良好的准确率和效率。③为了能够在肺动脉中更精确地检测栓塞,提出一种由纵膈到外围的肺动脉分割方案,将常规的肺血管分割与动静脉分离两个任务集为一体。针对跟踪过程中的渗漏问题和血管中断问题,提出了相应策略。提出基于区域生长和分段行进的血管树跟踪算法作为肺动脉跟踪的主线。针对渗漏问题,提出一种基于多尺度二阶导数的特征图过滤技术以排除或减少相邻对象间的大面积接触区域,并结合3D形态学血管分枝评估技术防止渗漏。对于血管中断部分则通过提取最小代价路径进行补偿。实验结果表明,本方案对纵膈内主肺动脉的分割具有很高敏感度和鲁棒性,能有效排除大部分肺部静脉,同时证实了方案中各技术的有效性。④针对CTA图像中肺栓塞检测的医学特性,在实现前人研究成果的基础上,提取候选栓塞的不同区域对象并另提出了与周围环境的灰度差异特征、边界灰度特征、部分形态特征以及与基于肺动脉树的特征,用以描述肺栓塞与相关环境及器官对象之间的联系。对各特征的检验概率分析表明本文提取的特征大部分具有更好的真实栓塞识别能力。⑤联合使用各类特征选择最优特征子集并优化分类器以改善肺栓塞计算机辅助检测系统的性能。首先提出一种基于蚁群算法的组合式特征选择算法,再选择分类器分别评估不同类别特征、特征选择子集、与人工筛选特征子集以筛选出分类性能最优的特征子集。在此基础上,进一步提出采用集成学习的方法改善分类器的性能。实验结果表明采用基于蚁群算法得到的特征子集,结合以随机森林为基分类器的集成学习算法能获得良好的分类性能。与其他肺栓塞CAD系统的性能比较说明本文所提出的CAD系统能够检测中心型、叶段、段以及亚段动脉中的栓塞,同时具有相对很低的假阳性率。更多还原

【Abstract】 Featured by low rates of diagnosis and high motality rates, pulmonary embolism (PE) is one of the three most common cardiovascular diseases in Western countries. In China, it has long been mistakenly regarded as one of the rare diseases, but is gathering attention rapidly in recent years. Due to its lack of clinical specificities, PE could be easily neglected in clinical diagnosis. The mortality rate is high without early diagnosis and efficient treatment. The invention of computed tomography and angiography has made modern imaging examination a significant method for PE diagnosis. Amongst, CT pulmonary arteriography has been used as one of the main methods of clinical diagnosis. The assistance by computer-aided detection(CAD) system could further enhance accuracy and efficiency of PE diagnosis.The research on PE CAD system incepted nearly ten years ago. However, it is still far from mature, with many issues remaining to be resolved in both theory and application. In China, the research is primitive at best. The current research, by reference to the normal procedures of PE CAD systems, proposes a new technical route for the designing of PE CAD systems. It takes into account central PE inside mediastinum as one of the key detection objects, and aims to reduce false positives in PE diagnosis. The tasks and innovations made at each step of the research are as follows:Firstly, an automatic 3D lung segmentation algorithm, based on anatomy knowledge, is proposed with an aim to segment the left and right lungs. At first, pre-segmentation is made by thresholding method, after an analysis of the density distribution of tissues in CT images. With regards segmentation errors caused by bronchus, the top-to-down region growing method is proposed to extract airway lumen and to avoid leakage from airway to parenchyma, to dilate to include airway wall; then the methods of regional area analysis and area projection are proposed to judge and separate left and right lungs, then to repair the edges of parenchyma. Experimental results show the applicability of the algorithm to CT plain scan/enhancement scan data, the effectiveness in excluding airways between the two lungs and in the segmentation of complicated images.Secondly, in order to provide a basis for segmentation of pulmonary artery inside mediastinum and diagnosis of central PE, the study, in an innovative fashion, adds the module of mediastinum segmentation to PE CAD system and proposes an automatic 3D mediastinum segmentation algorithm. As an anatomic region, the mediastinum does not have clear boundaries or shape. The study uses the idea of margin points in the description of the anatomical structure of 2D mediastinal area. The points are extracted by a method based on corner detection, and the paths between points by fast marching method. Meanwhile, the location information of context-contour is introduced in 2D segmentation process in order to limit and guarantee the continuity of the surface of 3D mediastinum. The segmentation result is evaluated by the cumulative distribution of deviation distance from the gold standard result. The experiment shows that the automatic segmentation algorithm based on context-contour mediastinum could achieve complete and smooth mediastinum, and the segmentation is accurate and efficient.Thirdly, with an aim to detect PE in the pulmonary artery more accurately, the study proposes a segmentation plan from mediastinal to peripheral pulmonary artery, thus combining the tasks of traditional segmentation of pulmonary vascular and separation of artery and vein. With regards the problem of leakage and unconnected vessels in the tracing process, the following proposals are made: a vascular tree tracking method based on regional growing and slice marching is used for artery tracking; a characteristic pattern filtration method based on multiple-dimensioned second-order derivative is applied in order to eliminate or reduce the contact area of neighboring objects and to prevent leakage by combination with 3D morphological vessel section assessment technique; the unconnected vessels are repaired by extracting minimum cost path. Experimental results show the high sensitivity and robustness for segmentation of main pulmonary artery inside mediastinum. It is also proved that the proposed method could exclude most pulmonary veins effectively.Forthly, with regards the medical characteristics of PE detection in CTPA images, the present research extracts different regions for candidate PE, based on existing research results, and proposes another eighteen features for describing the relationship between PE and related environment and organs, such as gray-level difference from surrounding environment, gray level of borders, partial morphological feature and feature based on pulmonary artery tree. The chi-square test shows that most of the proposed features have better detection capability for real PE.Fifthly, the performance of the CAD system for PE diagnosis is improved by selecting an optimal subset among the proposed features and optimizing the classifier. The research proposes a hybrid feature selection algorithm based on ant colony algorithm. Then selected classifiers are used to evaluate features of different types, feature selection subsets and manually classified feature subsets, in order to acquire the optimal feature subset for classification. The performance of classifier is then improved by the method of ensemble learning. Experimental results show the satisfactory performance of the combination of the feature subset deriving from ant colony algorithm and the ensemble learning algorithm with random forest as the base classifier. A comparison of the performance of CAD systems for PE detection reveals that the proposed CAD system could detect central, segmental, section and sub-segmental embolism, and could reduce false positives.更多还原