摘要
目的代价敏感分类策略可以最小化误分类代价并有效提高识别率。针对计算机辅助肺穿刺样本的病理图像分析,提出一种新的代价敏感图像层的肺癌诊断系统(cost-sensitive image-level lung cancer diagnosis system,CILCDS)。方法基于潜在狄利克雷分配模型(latent dirichlet allocation,LDA)、代价敏感支持向量机(cost-sensitive support vector machine,CSSVM)以及多类支持向量机(multi-class support vector machines,mcSVM)等最新计算机技术,对肺穿刺病理细胞学涂片进行识别诊断。实验数据集是271例肺穿刺样本图像,其中240例肺癌样本有术后组织病理诊断结果对照。结果通过与基于轴平行矩形(axis-parallel rectangles,APR)、K近邻(citation K-nearest neighbors,Citation-kNN)、集成多样性密度(ensemble ofdiversity density,EM-DD)、多分类多示例自适应增强法(multi-class multi-instance adaptive boosting,MCMI-AdaBoost)等方法进行对比,CILCDS在癌与正常细胞的判断以及癌细胞的分类识别诊断过程中能够取得更低的错误分类代价以及较好的组织病理结果符合率。结论 CILCDS对肺癌细胞涂片诊断率高,并能减少既往肺癌细胞病理诊断过程中假阳性结果。
Objective The cost-sensitive classification strategy can effectively improve the rate of identification at the lowest cost of misdiagnosis. The aim of this study is to propose a novel cost-sensitive image-level lung cancer diagnosis system （CILCDS） for computer-aided pathological analysis of needle biopsy specimens. Methods CILCDS mainly incorporated the model of latent Dirichlet allocation （LDA）, a cost-sensitive support vector machine （CSSVM） and a multi-class support vector machine （mcSVM）. Using CILCDS, we conducted differential diagnoses on 271 needle biopsy specimens of the lungs, 240 of them with postoperative his- topathological results for comparison. Results We achieved obviously lower cost of misdiagnosis and higher precision of normal/cancerous identification based on the results of comparisons made with multi-instance learning algorithms such as axis-parallel rectangles,citation K-nearest neighbors and ensemble of diversity density, as well as our previous multi-class multi-instance adaptive boosting （MCMI- AdaBoost）. Conclusion CILCDS affords a high accuracy and reduces false positive results in the cytopathological diagnosis of lung cancer.
出处
《医学研究生学报》
CAS
北大核心
2012年第6期567-570,共4页
Journal of Medical Postgraduates
基金
国家自然科学基金(61035003)
关键词
代价敏感图像层肺癌诊断系统
肺肿瘤/诊断/病理学
图像处理
计算机辅助
Cost-sensitive image-level lung cancer diagno-sis system
Lung neoplasms/diagnosis/pathol-ogy
Image processing
Computer-assisted
