【作者】 吴龙海；

【导师】 周荷琴；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 肺癌是导致癌症病人死亡率高的主要原因,对它进行早期诊断和早期治疗是提高患者生存率的主要手段。肺癌的早期症状一般以肺结节的形态出现,因此正确检测出肺结节,及早了解肺部病变情况,进而及时治疗,对挽救肺癌病人的生命具有重要意义。近几年,我国各大医院中多层螺旋CT的装机数量显著上升,它们能提供肺部高分辨率CT(High Resolution Computerized Tomography,HRCT)图像,可用来更好地评价肺组织中肺结节的界面以及结节的内部结构,为肺部疾病的正确诊断提供了强有力的工具。但是海量的CT数据在提供更加详细、更加准确的诊断信息的同时,也给读片医生增加了很重的工作负担,为了提高医生的诊断效率和减轻他们的劳动强度,计算机辅助诊断(Computer-AidedDiagnosis,CAD)系统应运而生。要研究肺部病变的计算机辅助诊断,帮助医生诊断早期肺癌,首先必须面对和解决的问题就是胸部CT图像中肺结节的计算机自动检测。要实现肺结节的自动检测,需要涉及到图像分割中的一系列处理和分析方法。本文以HRCT数据为研究对象,以肺部病变的计算机自动检测为目标,对该领域的国内外研究现状作了广泛的调研,通过结合人体组织的解剖知识,对肺部病变检测及其相关的医学图像处理方法进行了深入的研究,完成了如下有新意的工作:(1)提出一种肺部区域自动分割算法。肺部分割结果的好坏将直接影响到后续处理的效率和效果。该算法针对CT图像中肺组织的灰度值与人体内部其他组织的灰度值有明显差异的特点,利用迭代计算求取最优分割阈值的方法,减小了阈值的选取对分割效果的影响。研究了相邻两个CT层面中气管/支气管位置的相互关系,利用前层图像中气管/支气管的位置自动地确定下层图像中气管/支气管区域生长的种子点,提出了气管/支气管的自动区域生长方法以去除气管/支气管区域对肺部边界提取的干扰。使用基于8-邻域搜索的边界跟踪算法来去除CT图像中的背景干扰和获取肺部区域的边界,避开了多次形态学操作,节省了处理时间,并且根据躯干和肺部边界的光滑性的特点,对8-邻域搜索策略进行改进来提高边界跟踪的速度。该算法可以快速、准确地完成CT图像中肺部区域的自动分割。(2)早期肺癌一般以肺结节的形式出现,对CT图像中的肺结节进行增强可以提高肺结节检测的准确度。本论文在假设肺结节是球形、血管是圆柱形的基础上,提出一种基于相关矩阵的肺结节增强算法。首先根据Hessian矩阵特征值的正负,从一些灰度值比较大的像素点中筛选出要增强的点,然后计算需要增强的点的相关矩阵,利用该点相关矩阵的特征值的相互关系设计了肺结节增强滤波器。算法利用的是图像的一阶偏微分信息,能够在对肺结节进行有效增强的同时减小传统滤波器对噪声的敏感度。(3)综合考虑二维检测速度较快和三维检测精度较高的特点,提出一种基于肺结节三维空间结构特征的肺结节检测算法。该算法首先在二维层片上使用灰度收敛指数滤波来产生候选肺结节,然后计算候选肺结节的三维特征来去除候选肺结节中的假阳性肺结节。收敛指数滤波能够快速地找到CT层片上的圆形和椭圆形区域,产生候选肺结节;充分利用结节的空间结构信息来去除假阳性肺结节,提高了检测精度。该算法在执行过程中不需要人机交互,具有较高的灵敏度和低假阳性。(4)针对肺结节检测过程中因受血管影响而呈现的假阳性肺结节问题,提出一种基于血管剔除的肺结节检测算法。把血管的连续横断面看作具有二维高斯密度分布的管状体,设计了基于管状模型的血管检测滤波器,用它对候选肺结节做进一步筛选来降低假阳性率。所设计的血管检测滤波器还可以用于其他需要血管检测的场合。更多还原

【Abstract】 Lung Cancer is the main reason of high mortality caused by cancer. Early diagnosis and treat of lung cancer is the main means to improve the survival rate of patients. Lung nodule is the symbol of most early stage of lung cancer. Detecting lung nodules correctly, understanding and treating lung diseases in early stage is of great significance to save the lives of lung cancer patients. In recent years, the number of Multi-slice Spiral Computerized Tomography (MSCT) installed in China’s major hospital has increased significantly. MSCT can provide High Resolution Computerized Tomography (HRCT) images, which could be used to better describe surface and internal structure of pulmonary nodules. MSCT is a powerful tool for accurate diagnosis of lung diseases. However, while massive CT data provides more detailed and more accurate diagnostic information for radiologists, it brings heavy burden of work to radiologists. In order to improve the efficiency of medical diagnosis and to reduce radiologists’ labor intensity, Computer-Aided Diagnosis (CAD) system came into being.To study Computer-Aided Diagnosis of lung diseases, assist radiologists to diagnose early lung cancer, the problem first of all we must face and solve is automatic detection of pulmonary nodules in thoracic CT images on the computer. In order to realize automatic detection of pulmonary nodules, a series of processing and analysis methods in image segmentation must be studied. With a direction of research on Computer-Aided Diagnosis of lung disease, and with the goal of automatic detection of pulmonary diseases on the computer, we have made an extensive survey on the research status at home and aboard in this domain. By using HRCT data as research materials, and by combining knowledge of human tissue anatomy, we have made an in-depth study on pulmonary disease detection and related processing methods of medical images. The contributions of this dissertation are as follows:(1) An automatic segmentation algorithm for lung region abstraction from CT images is proposed. The quality of lung segmentation results will affect the efficiency and effectiveness of follow-up processing. Aim at the characteristic of gray value of lung tissue and others tissues in human body in CT images, optimal threshold is obtained by an iterative process of calculation, which can reduce the impact of threshold selection on segmentation results. The relationship between the locations of tracheal/bronchia in two adjacent CT slices is studied, the location of tracheal/bronchia in anterior slice is used to produce a seed point for automatic region growth of tracheal/bronchia in posterior slice. A border tracking algorithm based on 8-neighborhood searching method is adopted to eliminate background and to abstract the boundary of lung, which avoids many morphological operations, so processing time is saved. According to the smoothness of boundaries of human body and lung region, 8-neighborhood searching method is improved utilizing previous direction to increase the searching efficiency. The proposed algorithm is quite efficient and accurate for automated lung segmentation in CT images.(2) Lung nodule is the symbol of most early stage of lung cancer, enhancement of pulmonary nodules in CT images can improve, the precision of pulmonary nodule detection. Based on the assumption that nodule is spherical and vessel is cylindrical, an enhancement algorithm of pulmonary nodule is proposed. Points that need to be enhanced are selected from those pixels whose gray-value is relatively high, according to that whether the three eigenvalues of each point are all negative, and then correlation matrix is calculated. The relationship between eigenvalues of each point is used to design an enhancement filter for pulmonary nodules. By using the first order partial differential information of images, the pulmonary nodules are enhanced effectively and the sensitivity to noise is reduced.(3) By comprehensively considering the fact that two-dimensional detection is relatively faster and three-dimensional detection is more precise, a pulmonary nodule detection algorithm based on three-dimensional spatial structure of nodule is proposed. Nodule candidates are extracted by a two-dimensional Convergence Index (CI) filter firstly, and then three-dimensional features of each candidate are calculated to eliminate false-positive nodules from candidates. Rounded and elliptic regions can be found quickly by CI filter. This false-positive eliminating method is able to take full advantage of three-dimensional spatial structure information of nodules to improve detection precision. In the process of implementing the algorithm does not require manual intervention, and with high sensitivity and low false positive.(4) According to the disturbing of cross-sections of vessels in the process of nodule detection, a pulmonary nodule detection algorithm based on vessel eliminating is proposed. In the algorithm, a tubular vascular model is supposed, with Gaussian intensity distribution on cross-sections. A vessel detection filter based on tubular model is designed to eliminate false-positive nodules from candidates. This vessel detection filter may also be used in other occasions, where vessel detection is needed.更多还原