【作者】 王倩；

【导师】 宋恩民； 金人超；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2009， 博士

【摘要】 CT图像中肺部疾病的计算机辅助诊断系统已成为全球的研究热点之一，它可作为放射科医师的“第二双眼”，减少放射科医师的工作量及单独阅片的疏漏；还能通过图像分割、特征量化等技术为放射科医生提供精确的定量分析，在一定程度上提高放射科医师诊断结果的可重复性和对于图像、疾病解释的一致性。以肺癌（Lung cancer）、慢性阻塞性肺病（Chronic obstructive pulmonary disease，COPD）的CT图像为主要研究对象，针对肺结节、支气管狭窄等病灶各自的特点，对支气管病变及肺结节研究了不同的计算机辅助诊断方法。针对支气管病变的计算机辅助诊断的研究包括支气管自动检测及支气管形态特征量化。（1）支气管断面自动检测要解决的关键问题是如何检测到更多的适合于气管形态参数量化的支气管断面。在Zheng提出的方法基础上，提出了一种基于圆形度的多方向支气管断面自动检测方法；为检验该方法的有效性，对9个全肺CT图像进行了实验，结果表明该方法能够检测到可用于形态参数量化的支气管断面可达2007个／Scan，其中假支气管断面数目为0．67个／Section，过度粘连的气管断面数为0．12个／Section。所检测到的三级以下的支气管断面比例，COPD的CT图像要高于正常CT图像，这与医生对COPD的诊断结果具有一致性。（2）支气管形态参数量化的难点是支气管面积很小，PVE效应（Partial volume effect，PVE）对其分割影响较大，因此如何准确地分割支气管壁是要解决的关键问题。采用Zheng提出的方法，在最大-最小极值法分割支气管外壁的基础上，利用模糊技术精确量化支气管壁及管腔面积等形态参数；实验结果表明，对于所检测到的同位于第3级支气管的断面中，COPD患者的支气管壁面积量化值大于同年龄正常人的，这也表明该方法的形态参数量化结果与医生阅片结果具有一致性。针对肺结节计算机辅助诊断的研究主要包括肺结节的自动检测及肺结节特征量化。（1）肺结节的自动检测要解决的关键问题是在足够高的灵敏度条件下，尽量降低其假阳性率，其中毛玻璃结节是检测的难点。基于Li提出的多尺度选择性滤波方法对结节进行初始检测，然后在所得的候选结节分割基础上提取有效直径、紧凑度等20个特征，最后利用Fisher线性分类器对优化后的特征进行分类，去除分类结果中的假阳性结节。为评估结节检测的性能，实验利用LIDC（Lung Imaging DatabaseConsortium，LIDC）所提供的肺CT图像，采用leave-one-scan-out方法对结节检测结果进行了ROC曲线分析。评估结果表明，全部结节的检测水平为当TPF为87％时，假阳性率为0．01FPs／Section；当TPF为96％时，假阳性率为0．03FPs／Section。评估结果还表明GGO¹（Ground glass opacity，GGO）结节能够较好地被检出。（2）肺结节特征量化是对肺结节良、恶性判别的一个重要基础，其难点之一是肺结节的准确分割。针对该难题提出了一种基于动态规划及多方向融合的肺结节分割方法。实验结果表明该方法能够较精确地分割LIDC所提供的23个肺结节，其平均overlap、Seg_TP及Seg_FP分别为66％，75％，15％；在结节准确分割的基础上，精确量化毛刺征、胸膜凹陷征、分叶征等具有临床诊断意义的结节特征。由于其影像表现复杂且各不相同，因此对于不同的结节特征需要研究特定的量化算法，这是另一难点所在。针对毛刺征提出了一种基于边界法线-梯度正交性的毛刺量化方法，并与已有的毛刺量化方法进行了比较评估，分析表明该方法的量化结果与真实的毛刺水平具有更高的一致性；实验结果还表明分叶征、胸膜凹陷征等其它特征的量化分级大部分具有87％以上的正确率。更多还原

【Abstract】 The computer-aided diagnosis of lung disease in CT images has become one of theworld’s research hotspots, which can be used as "secondary opinion" of radiologists toreduce the reading time and improve the correction of diagnosis. Computerizedsegmentation, feature quantification provided the precision analysis, which can improvereproducibility and the consistency of the diagnosis results.In this study, we developed different computer-aided diagnosis methods for airwaypathology and lung nodules in CT images with lung cancer and chronic obstructivepulmonary disease （COPD） based on their respective features.At first, we discuss the method of computer-aided diagnosis of COPD with thebronchial wall thickening in CT images. The key problem to be solved is how to detect themore suitable bronchial sections for quantification. In our study, based on ref, wepresent an automatic detection of bronchial sections based on roundness in multidirections. To validate this approach, we tested on 9 whole CT scans （including 7 normalCT scans and 2 COPD CT scans）, the results show that, this approach can detect 2007approximate perpendicular bronchial sections per scan, with 0.67 false bronchial sectionsper section and 0.12 excessive attachment bronchial sections per section. The radios ofdetected bronchial sections under 3^rd stage of the bronchial tree in COPD scans are higherthan normal ones, which is coincident with the diagnosis made by radiologists. Thequantification of bronchial morphological parameters is very difficult because of the verysmall area and thin wall of bronchia, the partial volume effect influences the accuracy ofthe segmentation, which is the key problem to be solved in this study. We adopted themethod used max-min of CT value to segment the boundary of the bronchial wall, andquantify the parameters, such as the bronchial wall, the bronchial lumen area, and the ratioof the bronchial lumen area to the total bronchial section area. The results show that, forthe bronchial sections of the same stage in the tree. The bronchial wall areas of COPD arelarger than normal ones in the CT of patients with the same age, which is also coincidentwith the diagnosis made by radiologists.In another work of our researches, we study the method of computer-aided diagnosisof lung cancer with lung nodules in CT images. The main two contents include automaticdetection and feature quantification of lung nodules. Automatic detection of the lungnodules is the most popular issue of the research about the CAD of lung diseases. The curial problem to be solved is how to improve the sensitivity and reduce the false positivefaction of nodules at the same time. During the processing, the GGO nodules are moredifficult to be detection. In our work, we developed a multi-scale selective filter based onref to initially detect the candidate nodules, and then, we extract 20 features, such aseffective diameter, compactness et al, from the candidate nodules. We optimized thefeature set by feature selection, and used Fisher linear classifier to identify the true andfalse positive nodules. We used leave-one-scan-out cross validation on LIDC （LungImaging Database Consortium, LIDC） dataset to evaluate this detection, and draw theROC of all nodules, multi-scale nodules, multi-pattern nodules, multi-shape nodules andmulti-attached-degree nodules. The results show that when TPF is 87% of the detection ofall nodules, the FPF is 0.01FPs/Section, and when TPF is 96%, the FPF is0.03FPs/Section.The results also show that the GGO nodules can be detect well by thismethod. The feature quantification of nodules is very important to its benign andmalignant discrimination. One of its difficulties is an accurate segmentation. We present anodule segmentation approach used dynamic-programming and multi-direction fusiontechniques. The segmentation result on 23 nodules provided by LIDC indicate that theaverage overlap, Seg_TP and Seg_FP are 66%, 75%, 15% respectively. Based on thesegmentation results, we quantify the nodule features contribute to the diagnosis, such asspiculation signs, pleural indentation, and lobulation sign et al. Specific quantificationalgorithm need to be study for each sign because of different performances, which isanother difficulty of this study. We present a novel quantification method of spiculationsbased on boundary-normal gradient direction and make an evaluation by ROC analysis.The results show that the quantification of spiculations by this method is with highconsistency with the real spiculation level. Moreover, in our experiments, we compare thequantification of other features, such as pleural indentation, lobulation sign et al, with thequalitative classification provided by radiologists. The results indicate that mostquantification of the features can obtain consistency higher than 87%.更多还原