Automatic segmentation of lung nodules with growing neural gas and support vector machine
Automatic segmentation of lung nodules with growing neural gas and support vector machine
复制标题
DOI：
10.1016/j.compbiomed.2012.09.003
复制
发表时间：
2012-11-01
影响因子：
7.7
通讯作者：
Gattass, Marcelo
中科院分区：
文献类型：
--
作者：
Barros Netto, Stelmo Magalhaes;Silva, Aristofanes Correa;Gattass, Marcelo
肺癌是所有其他类型癌症中发病率和死亡率最高的癌症之一。不幸的是，这种疾病往往诊断较晚，影响治疗结果。为了帮助专家在断层图像中搜索和识别肺结节，许多研究中心已经开发了计算机辅助检测系统(CAD系统)来自动化程序。这项工作旨在开发一种自动检测肺结节的方法。所提出的方法包括获取肺的计算机断层图像，通过提取胸部、提取肺和重建实质的原始形状的技术来缩小感兴趣的体积。之后，生长神经气体(GNG)被用来更多地约束比肺实质密度更高的结构(结节、血管、支气管等)。下一阶段是将类似肺结节的结构与其他结构分离，如血管和支气管。最后，通过形状和纹理测量，结合支持向量机，将结构分为结节结构和非结节结构。该方法确保以86%的敏感性和91%的特异性发现合理大小的结节。这导致了在29次考试的48个结节样本中进行10次训练和测试的平均准确率为91%。在分析的29项考试中，每项考试的假阳性率为0.138。(C)2012爱思唯尔有限公司。保留所有权利。
Lung cancer is distinguished by presenting one of the highest incidences and one of the highest rates of mortality among all other types of cancer. Unfortunately, this disease is often diagnosed late, affecting the treatment outcome. In order to help specialists in the search and identification of lung nodules in tomographic images, many research centers have developed computer-aided detection systems (CAD systems) to automate procedures. This work seeks to develop a methodology for automatic detection of lung nodules. The proposed method consists of the acquisition of computerized tomography images of the lung, the reduction of the volume of interest through techniques for the extraction of the thorax, extraction of the lung, and reconstruction of the original shape of the parenchyma. After that, growing neural gas (GNG) is applied to constrain even more the structures that are denser than the pulmonary parenchyma (nodules, blood vessels, bronchi, etc.). The next stage is the separation of the structures resembling lung nodules from other structures, such as vessels and bronchi. Finally, the structures are classified as either nodule or non-nodule, through shape and texture measurements together with support vector machine. The methodology ensures that nodules of reasonable size be found with 86% sensitivity and 91% specificity. This results in a mean accuracy of 91% for 10 experiments of training and testing in a sample of 48 nodules occurring in 29 exams. The rate of false positives per exam was of 0.138, for the 29 exams analyzed. (C) 2012 Elsevier Ltd. All rights reserved.