【作者】 李丽；

【导师】 邱天爽；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2010， 硕士

【摘要】 肺结节是肺部最常见的病变之一,肺结节的早期检测和诊断对于肺癌的早期诊治十分重要。近年来,随着多层螺旋CT(MSCT)、高分辨CT(HRCT)及低剂量胸部CT(LDCT)的应用,计算机辅助诊断(CAD)系统的必要性和重要性也日益显现,因此涌现了许多利用图像分割等进行病变识别的方法,本课题正是基于这样的现实意义,以CT图像肺结节提取为处理目标进行算法的研究,成功可靠的算法可以明显提高诊断医生的工作效率,为更多的患者服务。肺结节提取通常分为四个环节：肺实质分割、感兴趣区域的提取、特征提取、分类识别。肺实质分割过程本文尝试了交互式及非交互式两种方法,互为补充,能够准确得到不同图像源的肺实质图像,并具有一定的鲁棒性。感兴趣区域提取方面本文进行了三种方法的仿真,分别为模糊区域生长法、结合CI特征的均值平移法、快速行进法,其中本文创新性地将CI特征应用于均值平移法中,获得了良好的疑似结节提取效果。在特征提取方面,除了较为常用的灰度特征、形状特征之外,本文也验证了分形维对于识别肺结节的有效性,创新性地将其加入到特征提取中,最后利用BP神经网络对疑似结节进行分类识别。除了二维提取之外,本文的最大创新之处在于利用三维血管重建实现了肺结节的提取。由于肺部血管切片形状也为类圆形,与肺结节相似,往往造成较高的结节识别假阳性率。本文提出的方法打破常规思维,在高分辨率CT图像基础上,从肺部血管三维重建入手,间接去掉血管组织对结节提取的干扰。首先利用数学形态学及凸包算法获得二维完整肺实质,再利用区域增长法提取肺部软组织,间接得到疑似结节图像,然后利用三维Hessian矩阵特征值的几何意义,构造三维血管结构的增强因子,得到完整的肺部血管图像,将其与疑似结节图像进行对比,重合区域即可除去,最大限度得剔除了血管的干扰,最后再利用疑似区域的几何特征剔除残余的肺部杂质,与二维提取相比,最终获得了较低的假阳性率,提取准确率也有了大幅增加。更多还原

【Abstract】 Lung nodule is one of the most common pathological changes, thus early detection and diagnosis of lung nodule is very important for the medical treatment of lung cancer. In recent years, as the application of Multi-slice spiral CT, High-resolution CT and Low-dose chest CT, Computer-aided diagnosis (CAD) system will be more essential and more important. Thus a lot of methods emerge, which apply the image segmentation idea on detecting the pathological changes. Based on the realistic, the paper has done arithmetic research aiming to abstracting lung nodules from CT lung images.Lung nodule abstraction usually contains four steps:lung region segmentation, area-of-interest (ROI) abstraction, feature extraction and classification and recognition. Firstly, in lung region segmentation process, both interactive and non-interactive methods have been applied, which can be complimentary to each other. Both of them can accomplish the lung region image with robustness to a certain extent from different images. Secondly, in ROI abstraction, the paper uses three methods:fuzzy regional growing, mean shift with convergence index (CI) feature and fast marching method. Combining the mean shift and CI feature is a creation, which also acquires nice result. Thirdly, in the feature extraction, except the normal gray scale feature, shape feature and so on, the effect of fractal dimension in lung nodule detection is evaluated before adding it into the feature group. Finally, in classification and recognition, BP (Back Propagation) neural network is used.Besides the 2D abstraction, the most creative point of this paper is the successful 3D lung nodule abstraction by taking advantage of vessel reconstruction. Due to the circle shape of the lung vessel which is similar to lung nodules, the false positive rate is always quite high. The paper breaks the routine with the 3D reconstruction of vessels in lung as the beginning, then avoid the interference indirectly. First, we get the 2D complete lung regions. Second, the lung soft tissue is acquired by region growing method, so we can get the possible nodules. Third, use the geometrical meaning of 3D Hessian matrix’s Eigen values to finish vessel 3D enhancement and reconstruction. Then we compare the vessels with the possible nodule image, remove the overcast regions. Finally, we remove the small lung regions with the geometry features. The paper gets low false positive and high accuracy.更多还原