摘要
肺部CT图像中疑似病灶感兴趣区域(ROI)的准确分割是肺部计算机辅助检测/诊断(CAD)的重要环节。本文提出结合Hessian矩阵滤波的均值漂移聚类肺部疑似病灶ROI区域分割算法。对原图像进行多尺度Hessian矩阵圆形滤波,图像中圆形的疑似结节病灶区域得到滤波增强、直线形的气管/血管区域得到抑制,将Hessian矩阵滤波后的形状特征、灰度、空间位置3种信息引入特征空间,将均值漂移聚类的核函数分解为3种特征信息所分别对应的核函数乘积形式,最后采用自适应计算带宽的方法确定每个待分割疑似区域的带宽进行均值漂移聚类分割。对来自LIDC等127个包含不同类型肺结节的病例进行实验,实验结果表明引入Hessian矩阵圆形滤波信息的均值漂移聚类能够分割出与血管或气管相连或者交叉的结节区域,去除ROI中包含的非结节区域,能有效分割出基于灰度信息难以分割的毛玻璃型(GGO)结节;对于3种类型的结节区域:血管相连结节(VPN)、毛玻璃型结节(GG0)、孤立性结节(SPN)分割平均准确率分别为92.80%、86.13%、95.08%。
The accurate segmentation of the region of interest of the suspected lesions in thoracic CT images is an important issue for thoracic computer-aided detection/diagnosis （CAD）. In this paper, a ROC region segmentation algorithm for the suspected lung lesions is proposed based on Hessian matrix circle filtering combined with mean-shift clustering. The original image is filtered with multi-scale Hessian matrix circle filtering, the round suspected nodular lesions in the image are enhanced with filtering, the linear shaped trachea / vascular regions are suppressed, and the regions of Interest （RO1） are segmented initially by a simple threshold. Then, three types of feature information including the shape feature after Hessian matrix filtering, gray feature, spatial position are introduced to the feature space ; the kernel function of mean shift clustering is decomposed into the product form of three kinds of kernel functions corresponding to the three types of feature information. Finally, the adaptive bandwidth calculation method is adopted to determine the bandwidths of various suspected regions to be segmented, and the mean-shift clustering segmentation is conducted. Experiments on 127 cases from LIDC and etc. including different types of pulmonary nodules were conducted. The experiment results show that the mean-shift clustering with introducing the Hessian matrix circle filtering information can segment the nodule regions that are connected to or intersected with the blood vessels or trachea, remove the non - nodular regions contained in the regions of interest （ROIs） properly, and can also effectively segment the ground glass object （GGO） nodular regions that are difficult to be detected only with gray information. Using the proposed algorithm, the average segmentation accuracies for the three types of nodular regions： vascular pulmonary nodules （VPN）, ground glass object （GGO） and solitary pulmonary nodules （SPN） are 92.80%, 86.13% and 95.08%, respectively.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2016年第1期192-199,共8页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(61370152)
中央高校基本科研业务费专项资金(N130204003)项目资助
