To develop a radiomics signature based on CT image features to estimate the expression level of Ki-67 in non-small cell lung cancer (NSCLC). Methods: A total of 108 NSCLC patients, who underwent non-enhanced and contrast-enhanced CT scan in our hospital from January 2014 to November 2017, were retrospectively analyzed. They were confirmed by histopathological examination and undergone Ki-67 expression level test within 2 weeks after CT examination. The non-enhanced and contrast-enhanced CT three-dimensional structural images of the lesions were manually delineated by MaZda software, and the texture features of the region of interest were extracted. Combination of feature selection and classification methods were used to build radiomics signatures, and the classification were assessed using misclassification rates. The MaZda software provides texture feature selection methods including mutual information (MI), Fisher coefficients (Fisher), classification error probability combined with average correlation coefficients (POE+ACC), and Fisher+POE+ACC+MI (FPM), and texture feature analysis including raw data analysis (RDA), principal component analysis (PCA), linear classification analysis (LDA) and nonlinear classification analysis (NDA). Results: Among the 108 patients, 50 cases were at high levels of Ki-67 expression and 58 cases were at low levels of Ki-67 expression, respectively. The differences of gender, age and pathological type between the two groups were statistically significant (P<0.05). The radiomics signature built by FPM feature selection combined with NDA feature analysis based on non-enhanced CT images achieved the best performance for predicting the level of Ki-67 with a misclassification rate of 14.81%. However, radiomics signature based on contrast-enhanced CT images did not reduce the misclassification rate. Conclusion: The radiomics signature based on conventional CT image texture features is helpful to predict the expression of Ki-67 in NSCLC lesions, which can provide a non-invasive technique for assessing the invasiveness and prognosis for NSCLC.
目的：建立基于CT图像特征的影像组学标签，预测非小细胞肺癌(non-small cell lung cancer，NSCLC)的Ki-67表达水平。方法：回顾性收集2014年1月至2017年11月行肺部CT平扫及增强扫描并在检查后2周内经病理证实、行Ki-67表达水平检测的108例NSCLC患者。在MaZda软件中分别手动勾画出病灶CT平扫及增强的三维立体结构影像，提取感兴趣区的纹理特征参数。经过多种特征选择方法[交互信息(mutual information，MI)、Fisher系数(fisher coefficient，Fisher)、分类错误概率联合平均相关系数(classification error probability combined with average correlation coefficients，POE+ACC)以及联合方法(Fisher+POE+ACC+MI，FPM)]和分析方法[原始数据分析(raw data analysis，RDA)、主成分分析(principal component analysis，PCA)、线性分类分析(linear discriminant analysis，LDA)和非线性分类分析(nonlinear discriminant analysis，NDA)]组合建立影像组学标签，以误判率评价其诊断效果。结果：108例患者中Ki-67高表达组50例，Ki-67低表达组58例。Ki-67高表达与低表达两组间性别、年龄和病理类型差异均有统计学意义(P<0.05)。基于CT平扫图像纹理特征的FPM特征选择和NDA特征分析方法建立的影像组学标签预测Ki-67表达水平效果最佳，错判率为14.81%。基于CT增强图像纹理特征建立的影像组学标签错判率并未低于基于CT平扫图像纹理特征建立的影像组学标签错判率。结论：基于常规CT图像纹理特征建立的影像组学标签有助于预测NSCLC病灶Ki-67的表达水平，可为评估肿瘤侵袭性和预后情况提供一种无创性技术手段。.