基于薄层CT的三维影像组学在预测亚厘米磨玻璃样肺腺癌浸润程度的价值

3D radiomics analysis based on thin-slice CT images for preoperatively predicting the invasiveness of pulmonary adenocarcinoma appearing as sub-centimeter ground-glass nodules

目的:探讨基于术前薄层CT的三维影像组学预测亚厘米磨玻璃样肺腺癌浸润程度的临床应用价值。方法:回顾性分析华东医院2013年1月-2017年7月经病理证实的394例亚厘米肺腺癌患者(共446个结节)的术前肺部薄层CT和临床资料。选取2013年1月-2015年12月的253例患者的286个结节为验证集;2016年1月-2017年7月141例患者的160个结节为训练集。所有病例参照病理金标准分为浸润前病变和浸润性病变,且所有结节均逐层勾画ROI而得到其容积感兴趣区(VOI)。采用Matlab 2016b软件从每个结节的VOI中提取475个影像组学特征,利用最小绝对收缩和选择算子(LASSO)进行特征筛选,随后建立影像组学标签。采用单因素及多因素分析方法筛选出训练集中两组病变间差异有统计学意义的变量并建立回归模型,进一步在验证集中对此模型进行验证。采用ROC曲线评价模型对结节浸润性的预测效能。结果:经可重复性分析及LASSO降维,最终筛选出13个影像组学特征并建立影像组学标签。多因素分析结果显示影像组学标签和CT值是预测肺癌浸润程度的独立危险因子。在训练集中,回归方程、CT值和影像组学标签的ROC曲线下面积(AUC)分别为0.785(95%CI:0.730~0.840)、0.742(95%CI:0.681~0.802)和0.696(95%CI:0.630~0.760)。在验证集中,回归模型、CT值和影像组学标签的AUC分别为0.704(95%CI:0.618~0.790)、0.683(95%CI:0.595~0.772)和0.674(95%CI:0.588~0.761)。结论:基于薄层CT的三维影像组学特征联合临床资料建立的多因素logistic回归模型对预测亚厘米级磨玻璃结节样肺腺癌的浸润程度具有很好的临床应用价值及发展前景。

Objective:The purpose of this study was to investigate the value of 3D radiomics based on thin-slice CT images in preoperatively predicting the invasiveness of pulmonary adenocarcinoma appearing as sub-centimeter ground-glass nodules.Methods:A retrospective analysis of CT images and clinical data of 394 patients with 446 sub-centimeter GGNs from January 2013 to July 2017 was conducted in the study.A total of 253 patients(286 nodules)from January 2013 to December 2015 were included into a training set;and 141 patients(n=160)from January 2016 to July 2017 were selected as a validation set.Each case was divided into pre-invasive lesion group or invasive lesion group based on pathological results.And the area of interest(ROI)was delineated layer by layer in all the nodules and then volumetric ROI(VOI)was obtained for measurement.Matlab 2016b software was used to extract 475 radiomic features from each VOI,and then the least absolute shrinkage and selection operator(LASSO)was used to select features and construct a radiomic signature.Univariate and multi-variate analyses were used to train statistically significant variables between the two groups and to establish a model,and the model was validated in the validation set.The model performance was assessed by the area under the curve(AUC)of receiver operating characteristic(ROC).Results:After reproducibility analysis and LASSO logistic regression,13 robust radiomic features were selected to establish the radiomic signature.After multivariate analysis,radiomic signature and CT values were identified as independent risk factors for predicting the invasiveness of sub-centimeter GNNs.In the training set,the AUCs of the regression model,CT value and radiomics signature were 0.785(95%CI:0.730~0.840),0.742(95%CI:0.681~0.802)and 0.696(95%CI:0.630~0.706),respectively.In the validation set,the AUCs for the predictive power of the regression model,CT value,radiomics signature were 0.704(95%CI:0.618~0.790),0.683(95%CI:0.595~0.772)and 0.674(95%CI:0.588~0.761),respectively.Conclusion:The multivariate logistic regression model based on the 3D radiomic signature of thin-slice CT images and clinical data performs well in predicting the invasiveness of sub-centimeter GNNs.