CT平扫图像纹理分析鉴别浸润性肺腺癌与非钙化结核球

目的 探讨基于CT平扫图像纹理分析鉴别诊断浸润性肺腺癌与非钙化结核球的可行性。方法 回顾性分析52例经病理证实的单发肺结节患者的平扫CT资料，其中31例浸润性肺腺癌，21例非钙化结核球。采用MaZda软件于2种病灶各提取300个纹理特征，之后以费希尔参数法（Fisher）、最小分类误差与最小平均相关系数法（POE+ACC）、相关信息测度法（MI）分别筛选出10个最佳纹理特征，并将其合并得到3种方法联合的最佳纹理特征组合（MPF）。采用线性判别分析（LDA）和非线性判别分析（NDA）对4组最佳纹理特征进行分类，LDA及NDA分别以K-近邻分类器（K-NN）及人工神经网络（ANN）进行分类。分析4组纹理特征鉴别2种病变的最小错误率，比较2组病变间30个最佳纹理特征的差异，并绘制其鉴别2种病变的ROC曲线，计算AUC，评价其诊断效能。结果 对于单组最佳纹理特征，NDA/ANN-Fisher法的错误率最低，为7.69%（4/52）；对于MPF，NDA/ANN-MPF法的错误率最低，为5.77%（3/52）；而NDA/ANN-Fisher法的错误率与NDA/ANN-MPF法差异无统计学意义（χ²=0.15，P>0.05）。2种病变间存在10个纹理特征差异有统计学意义，其中差异熵S（1，1）、差方差S（1，1）及梯度方差的诊断效能较好（AUC=0.71、0.71、0.70），3者间AUC差异无统计学意义（P均>0.05）。结论 基于CT平扫图像纹理分析可较好地区分浸润性肺腺癌和非钙化肺结核球，为鉴别诊断提供可靠的客观依据。

Identification of invasive lung adenocarcinoma and non-calcified lung tuberculoma on plain CT images based on texture analysis

Abstract] Objective To investigate the feasibility of differential diagnosis of invasive lung adenocarcinoma and non-calcified tuberculosis using CT plain image texture analysis. Methods A retrospective analysis of 52 cases of pathologically confirmed single pulmonary nodules, including invasive lung adenocarcinoma in 31 cases, non-calcified tuberculosis in 21 cases. 300 texture features were extracted from 52 nodules by MaZda software. The optimized texture parameters of texture analysis were selected with fisher coefficient (Fisher), probability of classification error and average correction coefficient minimization of both classification error probability and average correlation coefficients (POE+ACC) , mutual information coefficients (MI) as well as combination of the above three methods (MPF) , respectively. The texture characteristics were analyzed by using linear discriminant analysis (LDA) and nonlinear discriminant analysis (NDA) provided by B11 module in the Mazda software, the minimum error probability of differential diagnosis of invasive lung adenocarcinoma and non-calcified tuberculosis. The texture features for identification were examined by Mann-Whitney U, the ROC curves were established for the texture features with significant differences, the area under the curve AUC was calculated, and the texture diagnostic performance was evaluated. Results The NDA/ANN-Fihser method had a minimum error rate of 7.69% (4/52), and the lowest error rate of the three combined with NDA/ANN-MPF was 5.77 (3/52). Among the 30 texture features, there were 11 significant difference entropy-related texture features for the two lesions. The difference entropy S(1,1) difference variance S(1,1) and gradient variance have good diagnostic performance (AUC>0.7). Conclusion Based on CT plain scan texture analysis, it is valuable to identify invasive lung adenocarcinoma and non-calcified tuberculosis.