东北大学学报(自然科学版) ›› 2009, Vol. 30 ›› Issue (9): 1249-1252.DOI: -

• 论著 • 上一篇    下一篇

基于K-L变换和支持向量机的三维肺结节识别

刘阳;赵大哲;刘积仁;   

  1. 东北大学医学影像计算教育部重点实验室;东北大学信息科学与工程学院;
  • 收稿日期:2013-06-22 修回日期:2013-06-22 出版日期:2009-09-15 发布日期:2013-06-22
  • 通讯作者: Liu, Y.
  • 作者简介:-
  • 基金资助:
    国家自然科学基金资助项目(60671050)

Recognition of 3-D lung nodules based on K-L transform and support vector machine

Liu, Yang (1); Zhao, Da-Zhe (1); Liu, Ji-Ren (2)   

  1. (1) Key Laboratory of Medical Image Computing, Ministry of Education, Northeastern University, Shenyang 110004, China; (2) School of Information Science and Engineering, Northeastern University, Shenyang 110004, China
  • Received:2013-06-22 Revised:2013-06-22 Online:2009-09-15 Published:2013-06-22
  • Contact: Liu, Y.
  • About author:-
  • Supported by:
    -

摘要: 针对传统二维结节感兴趣区的特征提取忽略了肺结节在三维空间的细部特征问题,提出了一种基于K-L变换和支持向量机(SVM)的肺结节识别新方法.首先对三维肺结节的几何特征和密度特征进行分析,计算并提取三维特征形成原始特征空间,然后使用K-L变换方法进行原始空间变换,去除特征间相关性,最后采用支持向量机分类方法来进行肺结节识别,并引入ROC曲线对算法性能进行评价.实验针对36组具有临床标注"金标准"的肺部HRCT数据进行,结果表明该方法的识别准确率可以达到94.33%,ROC曲线的Az值为0.94.

关键词: K-L变换, 支持向量机, 三维肺结节, 特征选择, ROC曲线

Abstract: Based on the K-L transform and support vector machine, a new recognition algorithm for the 3-D lung nodule was presented to solve the problem that the details of lung nodule were often ignored in the 3-D space when extracting the characteristics of lung nodule in the interested area by the conventional 2-D method. The geometry and intensity features of lung nodule were analyzed and the 3-D features were calculated and extracted to form an original feature space. Then, the K-L transform method was used to exclude the correlativity between features, and the support vector machine (SVM) was introduced to categorize and recognize the potential lung nodules with the receiver operating characteristic (ROC) curve introduced to evaluated the performance of the algorithm proposed. A test involving 36 sets of clinical lung HRCT data was carried out, where the lung nodules were marked with 'golden standard'. The results indicated that the validity of the algorithm or the accuracy of recognition is up to 94.33% and the value of Az (area under the ROC curve) is 0.94.

中图分类号: