基于CT纹理特征的非小细胞肺癌EGFR基因突变预测模型的构建

Establishment of EGFR mutation prediction model for non-small cell lung cancer based on CT texture features

目的探讨CT纹理分析对非小细胞肺癌(NSCLC)表皮生长因子受体(EGFR)基因突变的预测价值。方法收集312例经术后病理证实为NSCLC患者的临床资料,术前3个月内均进行了CT扫描,并进行EGFR基因突变测试。提取NSCLC的CT影像形态学征像(密度、大小、分叶、形态),构建诊断EGFR基因突变的主观预测模型。采用Fire Voxel软件提取NSCLC的CT纹理特征,并构建CT纹理特征预测模型。通过ROC曲线评估纹理特征模型、CT主观征像模型预测EGFR基因突变效能。结果多因素Logistic回归分析结果显示,分期、吸烟史、位置、密度及支气管空气征是NSCLC患者EGFR突变主观独立预测因子;而峰度、能量、自相关、熵值是EGFR突变的CT纹理预测参数。训练组中,CT主观征像模型ROC曲线下的面积为0.78,敏感度为72.3%,特异度为84.7%;纹理特征模型ROC曲线下的面积为0.85,敏感度为80.3%,特异度为87.1%;联合预测模型ROC曲线下的面积为0.89,敏感度为84.3%,特异度为92.1%。验证组中,联合预测模型同样显示最佳预测效能,ROC曲线下面积为0.88,其次为纹理特征模型,主观征像模型效能最低。结论多种CT纹理特征与NSCLC的EGFR基因突变状态有关,结合影像学特征,可以较好地预测NSCLC的EGFR基因突变状态。

Objective To explore the predictive value of CT texture analysis in epidermal growth factor receptor(EGFR)mutation of non-small cell lung cancer(NSCLC).Methods A total of 312 patients with NSCLC confirmed by postoperative pathology were collected.CT scans were performed within 3 months before operations,and EGFR gene mutations were performed simultaneously.The morphological signs(density,size,lobulation and morphology)of CT images of NSCLC wereextractedto construct a subjective prediction model for diagnosis of EGFR gene mutations.FireVoxel software was usedto extract the CT texture features of NSCLC and construct a CT texture feature prediction model.The ROC curve was used to evaluate the efficiency of the texture feature model and CT subjective sign model for predicting EGFR gene mutation.Results Multivariate Logistic regression identify the stage,smoking history,location,density and air bronchus as subjective independent predictors of EGFR mutation in theNSCLC patients.Kurtosis,energy,autocorrelation and entropy are CT texture prediction parameters for EGFR mutation.The area under the ROC curve of the subjective CT model in the training is 0.782,the sensitivity is 72.3%,and the specificity is 84.7%;while the area under the ROC curve of the texture feature model is 0.851,the sensitivity is 80.3%,and the specificity is 87.1%;the area under the ROC curve of the joint prediction model is 0.891,the sensitivity is 84.3%,and the specificity is 92.1%.In the testing,the joint prediction model also shows the best prediction performance with area under the ROC curve of 0.883,the texture feature model was the second,and the subjective feature model was the lowest.Conclusion A variety of CT texture features are related to the EGFR mutation status of NSCLC,and combined with imaging features,which can better predict EGFR mutation status of NSCLC.