【作者】 陈侃；

【导师】 田联房；

【作者基本信息】 华南理工大学 ， 模式识别与智能系统， 2013， 博士

【摘要】 在当今世界许多国家中，肺癌是导致人类死亡的主要原因之一。临床证明，早期发现与诊断肺癌可以显著地提高患者的生存机率。在早期阶段，肺癌以肺结节的形式表现出来。肺结节的形状和大小变化是良恶性诊断的主要依据。精确地分割肺结节是获取这些信息的先决条件，因此肺结节精确分割是发现与诊断肺癌的关键技术。随着医学影像设备的发展，计算机辅助诊断(Computer-Aided Diagnosis, CAD)系统得到了迅速的发展。在现有肺结节检测CAD系统中，主要的研究对象是孤立性的肺结节，而缺乏对磨玻璃(Ground Glass Opacity, GGO)肺结节和血管粘连(Juxta-vascular, JV)肺结节的研究。研究表明，与孤立性肺结节相比，GGO肺结节很有可能表现为恶性结节。除此之外，大量的肺结节表现为JV肺结节。因此，本论文将GGO肺结节和JV肺结节作为研究对象。近年来，活动轮廓模型(Active Contour Model, ACM)已经广泛运用在肺图像分割领域,主要是存在以下两个原因：(1)活动轮廓模型很容易现实数字化计算；(2)活动轮廓模型可以获取平滑和封闭的分割结果。为了精确分割GGO肺结节和JV肺结节，本文提出了几种改进的活动轮廓模型，并对这些模型进行了深入的研究。本论文的主要研究内容和创新点描述如下：首先，提出了基于模糊速度的活动轮廓模型的肺结节分割方法。目前基于边界的活动轮廓模型的肺结节分割方法和基于区域的活动轮廓模型的肺结节分割方法，存在边界泄漏问题。对此，本文提出基于模糊速度函数的活动轮廓模型(Fuzzy Speed-basedActiveContour Model, FSACM)的肺结节分割方法。该活动轮廓模型能够精确分割GGO肺结节(具有模糊边界)和JV肺结节(肺结节与血管之间具有相似的亮度值)。首先，结合亮度特征和局部形状特征构造二维向量，并且根据该二维向量计算三个模糊速度函数。其次，将模糊速度函数引入到活动轮廓模型中，使得在肺结节的边界处，这些模糊速度函数等于零，轮廓曲线停止演变，从而完成肺结节的精确分割。其次，提出基于后验概率和小波能量的活动轮廓模型的磨玻璃型肺结节的分割方法。由于GGO肺结节具有模糊的边界、低对比度和亮度非均匀等特征，所以使用局部区域的活动轮廓模型很难分割该类型结节。为了精确分割GGO肺结节，提出了基于后验概率和小波能量的活动轮廓模型(Active Contour Model Based on Posterior Probabilityand Wavelet Energy, ACMPPWE)。首先，将小波能量引入到活动轮廓模型的中。因为小波能量增强肺结节与背景之间的对比度，并且在曲线演变的过程中考虑到局部区域信息，所以该活动轮廓模型能够精确分割低对比度和亮度非均匀的磨玻璃型肺结节。其次，基于后验概率的速度函数引入到活动轮廓模型的规则项中。在磨玻璃型的边界处，曲线进化的速度函数等于零，所以该活动轮廓模型能够分割具有模糊边界的磨玻璃型肺结节。最后，根据后验概率，可以获取轮廓曲线演变的初始轮廓。由于获取的初始轮廓位于真实分割目标边界的附近，所以轮廓曲线演变可以获得全局最小能量。再次，提出基于局部与全局隶属的活动轮廓模型的磨玻璃型肺结节的分割方法。为了精确分割GGO肺结节，提出了另外一种活动轮廓模型，即基于局部与全局隶属的活动轮廓模型(Active Contour Model Based on Local and Global Membership, ACMLG)。首先，采用全局隶属度计算边界停止函数。由于在磨玻璃型肺结节边界处，边界停止函数等于零，所以该活动轮廓模型能够精确分割具有模糊边界的磨玻璃型肺结节。其次，采用局部隶属度计算轮廓模型的数据项。由于局部隶属度可以增强肺结节与背景之间的对比度，并且考虑到局部区域信息，所以该活动轮廓模型可以精确分割低对比度和亮度非均匀的磨玻璃型肺结节。最后，根据全局隶属，可以获取轮廓曲线演变的初始轮廓。由于获取的初始轮廓位于真实分割目标边界的附近，所以轮廓曲线演变可以获得全局最小能量。最后，提出基于模糊速度和贝叶斯距离的活动轮廓模型的血管粘连型肺结节的分割方法。由于JV肺结节通常具有亮度非均匀性，并且肺结节与血管之间具有相似的亮度值，所以局部区域活动轮廓模型和基于形状的活动轮廓模型不能精确分割该类肺结节。针对JV肺结节的分割，提出基于模糊速度和贝叶斯距离的活动轮廓模型(Active ContourModel Based on Fuzzy Speed and Bhattachary Distance, ACMFSBD)。首先，基于模糊速度的边界停止函数引入到轮廓模型的规则项中。由于在JV肺结节与血管粘连处，边界停止函数等于零,所以该活动轮廓模型能够精确分割JV肺结节；其次，根据亮度和形状信息联合的贝叶斯距离计算活动轮廓模型的数据项。由于该贝叶斯距离可以增强JV肺结节与血管之间的对比度，并且考虑到局部区域信息，所以该活动轮廓模型可以精确分割亮度非均匀的JV肺结节。更多还原

【Abstract】 Lung cancer is the leading casue of death in many countries of the world. The clinicalevidence verifies that early detection and diagnosis of lung cancer can significantly improvethe survival rate of patients. Pulmonary nodule is the representation form of early stage oflung cancer. A main factor in determining a nodules malignancy status is the change in thenodule size and shape information. Pricise nodule segmentation is a prerequisite for obtainingthe change in the nodule and shape information, so pricise nodule segmentation is criticalstage. With the development of medical imaging equipment, the computer-aided diagnosis(CAD) system has been rapid development. In the CAD system of pulmonary noduledetection, the researches are mainly aimed to solid pulmonary nodules, but without moreresearches on the research of ground glass opacity (GGO) pulmonary nodules andjuxta-vascular (JV) pulmonary nodules. Studies have shown that GGO pulmonary noduleshave higher risks of being malignant than solid pulmonary nodules. Additionally, JVpulmonary nodules account for the largest typology of lung nodules. So GGO pulmonarynodules and JV pulmonary nodules are research objects in this dissertation. Recently, theactive contour models have been extensively applied to lung image segmentation. There aretwo reasons:(1) the active contour models can be easily formulated;(2) the active contourmodels can provide smooth and closed contour as segmentation results.In this dissertation, to segment GGO pulmonary nodules and JV pulmonary nodules,some improved active contour models are proposed and those models are studied intensively.The main contents and innovation are as follows.Firstly, the fuzzy speed function-based active contour model for segmentation ofpulmonary nodules is proposed. At present, the segmentation algorithm of pulmonary nodulesusing the edge-based active contour model and the region-based active contour model maycause boundary leakage. In order to avoid this phenomenon, a new segmentation algorithm ofpulmonary nodules using active contour model based on fuzzy speed function is proposed.This active contour model can accurately segment GGO pulmonary nodules, which havefuzzy boundary, and JV pulmonary nodules, which have the similar intensity as vessel. First,the two-dimensional vectors, which use gray feature and local shape feature, are constructedand three fuzzy speed functions are calculated using those two-dimensional vectors. Second,the three fuzzy speed functions are incorporated into the active contour model. At theboundary of pulmonary nodules, the three fuzzy speed functions equal to zero and theevolution of the contour curve stops, so that the accurate segmentation of pulmonary nodules in completed.Secondly, the segmentation of GGO pulmonary nodules using active contour modelbased on posterior probability and wavelet energy is proposed. Due to the fuzzy boundary,low contrast and intensity inhomogeneity of GGO pulmonary nodules, it’s really hard toaccurately segment them with the local region-based active contour models. To accuratlysegment GGO pulmonary nodules, a novel active contour model based on posteriorprobability and wavelet energy is proposed. First, the wavelet energy is incorporated into theactive contour model. The wavelet energy can enhance the dissimilarity between pulmonarynodule with background, and the local region information is considered, so the proposedactive contour model can segment low contrast GGO pulmonary nodules and those withintensity inhomogeneity. Second, the speed function based on posterior probability isincorporated into the active contour model. At the boundary of GGO pulmonary nodules, thespeed function of evolution equals to zero, so the active contour model can segment GGOpulmonary nodules with fuzzy boundary. Third, the initial contour is set near the boundary ofthe GGO pulmonary nodules using posterior probability, so the evolution curve may obtain aglobal minimum.Thirdly, the segmentation of GGO pulmonary nodules using active contour model basedon based on local and global membership is proposed. In order to segment GGO pulmonarynodules, another novel active contour model based on local and global membership isproposed. First, an edge stopping function based on the global membership is formulated. Theedge stopping function equals to zero at the boundary of GGO pulmonary nodules, so theproposed active contour model can accurately segment GGO pulmonary nodules with fuzzyboundary. Second, the data compoment of active contour model is calculated by using localmembership. The local membership can enhance the dissimilarity between pulmonary nodulewith background, and the local region information is considered, so the proposed activecontour model can segment low contrast GGO pulmonary nodules and those with intensityinhomogeneity. Third, the initial contour is set near the boundary of the GGO pulmonarynodules using global membership, so the evolution curve may obtain a global minimum.Finally, the juxta-vascular pulmonary nodules segmentation using active contour modelbased on fuzzy speed function and Bahattachary distance is proposed. Because of similarintensity in adjacent regions and intensity inhomogeneity, JV pulmonary nodules are notaccurately segmented by using the local region-based active contour models and theshape-based active contour models. For JV pulmonary nodules segmentation, a novel activecontour model based on fuzzy speed function and Bahattachary distance is proposed. First, the edge stopping function base on the fuzzy speed is incorporated into the regularizationcomponent of active contour model. The edge stopping function equals to zero at the placebetween JV pulmonary nodules and vessel, so the proposed active contour model canaccurately segment JV pulmonary nodules. Second, the data component of tha active contourmodel is formulated according to the shape-intensity joint Bhattacharya distance. TheBhattacharya distance can enhance the dissimilarity between JV pulmonary nodules andvessels, and the local region information is considered, so the proposed active contour modelcan segment JV pulmonary nodules with intensity inhomogeneity.更多还原