Entropy-based feature extraction and decision tree induction for breast cancer diagnosis with standardized thermograph images
Entropy-based feature extraction and decision tree induction for breast cancer diagnosis with standardized thermograph images
复制标题
DOI：
10.1016/j.cmpb.2010.04.014
复制
发表时间：
2010-12-01
影响因子：
6.1
通讯作者：
Yang, Chi-Shih
中科院分区：
文献类型：
--
作者：
Lee, Ming-Yih;Yang, Chi-Shih
在这项研究中，提出了一种基于计算机辅助熵的特征提取和决策树诱导方案，用于利用热像图诊断乳腺癌。首先，采用Beier-Neely场变形和线性仿射变换分别对全身和局部区域进行几何标准化；统计分析同一解剖位置像素群的灰度值，进行异常区域分类。形态学关闭和打开操作用于识别统一的异常区域。提取3种类型的25个特征参数（10个几何、7个拓扑和8个热）进行参数因子分析。通过决策树对阳性和阴性异常区域进行重新分类，归纳出基于病例的诊断规则。最后，利用解剖器官匹配方法，识别出具有阳性异常区的相应器官。为了验证所提出的基于病例的诊断方案的有效性，对71例和131例患有和未患乳腺癌的女性患者进行了分析。实验结果表明，共检测到1750个异常区域（703个阳性，1047个阴性），并将822个分支分解到决策空间中。14支有4个以上阳性异常区。这些阳性异常区域占比不到10%（61/703 = 8.6%）的关键诊断路径，可以有效地对以上14个分支中一半以上的癌症患者（42/71 = 59.2%）进行分类。2010爱思唯尔爱尔兰有限公司版权所有。
In this study, a computer-assisted entropy-based feature extraction and decision tree induction protocol for breast cancer diagnosis using thermograph images was proposed. First, Beier-Neely field morphing and linear affine transformation were applied in geometric standardization for whole body and partial region respectively. Gray levels of pixel population at the same anatomical position were statistically analyzed for abnormal region classification. Morphological closing and opening operations were used to identify unified abnormal regions. Three types of 25 feature parameters (i.e. 10 geometric, 7 topological and 8 thermal) were extracted for parametric factor analysis. Positive and negative abnormal regions were further reclassified by decision trees to induce the case-based diagnostic rules. Finally, anatomical organ matching was utilized to identify the corresponding organ with the positive abnormal regions. To verify the validity of the proposed cased-based diagnostic protocol, 71 and 131 female patients with and without breast cancer were analyzed. Experimental results indicated that 1750 abnormal regions (703 positive and 1047 negative) were detected and 822 branches were broken down into the decision space. Fourteen branches were found to have more than 4 positive abnormal regions. These critical diagnostic paths with less than 10% of positive abnormal regions (61/703 = 8.6%) can effectively classify more than half of the cancer patients (42/71 = 59.2%) in the abovementioned 14 branches. (c) 2010 Elsevier Ireland Ltd. All rights reserved.