基于术前CT影像组学特征预测肺腺癌患者EGFR突变状态的研究

Study on Predicting EGFR Mutations Status in Patients with Lung Adenocarcinoma Based on Preoperative CT Radiomics Features

目的探讨术前CT影像组学特征在预测肺腺癌患者EGFR突变中的价值。方法回顾性搜集经病理证实为肺腺癌且检测EGFR突变状态的患者共286例,其中,EGFR突变126例,野生型160例。按7∶3的比例随机分为训练集和验证集。从每个患者的感兴趣区内分别提取1468个组学特征,采用Wilcoxon检验、最小绝对收缩和选择算子(LASSO)回归和Logistic回归筛选影像组学特征。在训练集中采用Logistic回归的方法构建预测模型,并在验证集中评估其性能。通过ROC曲线评价模型的预测性能,并计算曲线下面积(AUC)、敏感度、特异度和准确性。DeLong检验用于比较各模型AUC之间的差异。结果两个临床因素(性别和吸烟史)与EGFR突变独立相关(P<0.05),而年龄、CEA和肿瘤位置在两组之间差异无统计学意义(P>0.05)。9个影像组学特征在两组之间有显著差异(P<0.05)。训练集中,临床模型、组学模型和综合模型的AUC分别为0.692、0.762和0.814,而在验证集中分别为0.712、0.779和0.827。训练集中,综合模型与临床模型、组学模型AUC的差异具有统计学意义(P<0.05),而其余模型AUC的差异无统计学意义(P>0.05)。验证集中,综合模型和临床模型AUC的差异具有统计学意义(P<0.05),而其余模型AUC的差异均无统计学以意义(P>0.05)。结论基于术前CT影像组学特征构建的模型可以用于预测EGFR突变状态,其中,综合模型具有较高的预测性能。

Objective To explore the value of predicting EGFR mutations in patients with lung adenocarcinoma based on preoperative CT radiomics features.Methods A total of 286 patients with lung adenocarcinoma and EGFR mutation status detected in the pathology department were retrospectively collected,of which 126 were ECFR mutations and 160 were wild-type.According to the ratio of 7:3,it is randomly divided into training and validation sets.1468 radiomics features were extracted from the region of interest(ROI)of each patient.Wilcoxon test,least absolute contraction and selection operator(LASSO)regression and Logistic regression were used to screen radiomics features.In the training set,the Logistic regression was used to construct the prediction model,and its performance was evaluated in the validation set.The prediction performance of the model was evaluated by ROC curve,and the area under the curve(AUC),sensitivity,specificity and accuracy were calculated.Delong test was used to compare the difference of AUC between the models.Results Two clinical factors(sex and smoking history)were independently correlated with EGFR mutation(P<0.05),but there was no significant difference in age,CEA and tumor location between the two groups(P>0.05).A total of 9 radiomics features were significantly different between the two groups(P<0.05).The AUCs of clinical,radiomics and comprehensive models were 0.692,0.762 and 0.814 in training set and 0.712,0.779 and 0.827 in validation set,respectively.In the training set,the difference in AUC between comprehensive model,clinical model and radiomics model was statistically significant(P<0.05),while the difference in AUC of the other models was not statistically significant(P>0.05).In the validation set,the difference in AUC between comprehensive model and clinical model was statistically significant(P<0.05),while the differences in the AUC of the other models were not statistically significant(P>0.05).Conclusion The model constructed based on the preoperative CT radiomics features can be used to predict EGFR mutations status.Among them,the comprehensive model has high predictive performance.