能谱CT及碘基图影像组学特征鉴别肺炎性及恶性病变

Radiomics analysis of iodine overlay maps from spectral CT in differentiating inflammatory and malignant pulmonary lesions

目的:探讨并比较能谱CT定量参数及碘基图影像组学特征对肺炎性及恶性病变的鉴别诊断价值。方法:回顾性分析123例(炎性组38例,恶性组85例)行动脉期能谱CT增强扫描的肺结节或肿块患者;比较病灶感兴趣区(region of interest,ROI)在70 keV及40 keV单能量图像下的CT值、能谱曲线斜率(λ)、碘浓度(iodine concentration,IC)、有效原子序数值(effectiveZ,Zeff)、标准化碘浓度(normalized iodine concentration,NIC)、标准化有序原子序数值(normalized effective-Z,NZeff)及最大径(D),进行多因素分析并绘制受试者工作特征(receiver operating characteristic,ROC)曲线。手动分割ROI,提取影像组学特征,以7∶3拆分训练集和测试集,通过皮尔森相关系数法、递归特征消除筛选特征,5折法交叉验证,使用线性判别式分类器建立预测模型,并绘制模型的ROC曲线。结果:炎性组及恶性组在40 keV下的CT值、λ100-70 keV、λ70-40 keV、λ100-40 keV、IC、NIC、Zeff、NZeff以及D的差异具有统计学意义(P <0.05),多因素分析提示NIC为独立影响因素(P <0.001),NIC的曲线下面积(area under curve,AUC)为0.728。影像组学共筛选出10个特征建立预测模型,其AUC为0.843,二者的差异具有统计学意义(P=0.034)。结论:能谱CT定量参数及碘基图影像组学特征对于肺内炎性及恶性病变的鉴别诊断具有很大价值,碘基图组学预测模型取得更优的诊断效能。

Objective:To investigate the value of quantitative analysis from spectral CT imaging and radiomics analysis from iodine overlay maps in differentiating inflammatory and malignant pulmonary lesions. Methods:A total of 123 patients with solitary pulmonary nodules underwent contrast enhanced spectral CT scan in arterial phase were divided into two groups(38 cases in inflammatory group and 85 cases in malignant group). The gemstone spectral imaging(GSI)viewer was used for image display and data analysis. The parameters,including CT value on 70 keV and 40 keV monochromatic images,slope of spectrum energy curve(λ),iodine concentration(IC),effective-Z(Zeff)value,normalized iodine concentration(NIC),normalized effective-Z(NZeff)value and the maximum diameter of lesion(D)were measured. After the multivariate logistic regression analysis,the individual diagnostic performance of independent factor was compared by receiver operating characteristic(ROC)curve. Radiomics features were extracted from manually segmented ROIs. Patients were randomized devided into training or test set in a ratio of 7∶3. Pearson’s correlation and recursive feature elimination(RFE)were used to select features. A radiomics-based predictive model using linear discriminant analysis was developed and calibrated with fivefold cross-validation. ROC curve was generated to assess the diagnostic performance. Results:CT value on 40 keV monochromatic image,λ100-70 keV,λ70-40 keV,λ100-40 keV,IC,NIC,Zeff,NZeff and D showed significant differences between inflammatory and malignant lesions(P < 0.05). NIC was an independent factor based on the multivariate logistic regression analysis(P < 0.001). The area under the curve(AUC)value of NIC was 0.728. Ten radiomic features were selected for predicting malignant lesions. The AUC value of model(0.843)was significantly higher than NIC(P=0.034). Conclusion:Spectral CT with GSI mode helps to identify inflammatory and malignant pulmonary lesions. The radiomics-based predictive model provides a more promising tool for differentiating.