基于^(18)F-FDG PET/CT影像组学特征和临床因素预测肺腺癌EGFR突变状态  

Prediction of the Mutation Status of EGFR in Lung Adenocarcinoma via ~(18)F-FDG PET/CT Radiomics Features and Clinical Factors

作　　者：胡峙珩 赵艳萍[1] 巴雅[2] 宗平 夏露花[1] 张泽宇 吴旻[1] 张帆 HU Zhiheng;ZHAO Yanping;BA Ya;ZONG Ping;XIA Luhua;ZHANG Zeyu;WU Min;ZHANG Fan(Department of Nuclear Medicine,Affiliated Tumor Hospital of Xinjiang Medical University,Urumqi 830011,China;不详)

机构地区：[1]新疆医科大学附属肿瘤医院核医学科,新疆乌鲁木齐830011 [2]新疆医科大学第一附属医院核医学科,新疆乌鲁木齐830011 [3]新疆农业大学计算机与信息工程学院,新疆乌鲁木齐830052

出　　处：《中国医学影像学杂志》2023年第8期844-851,共8页Chinese Journal of Medical Imaging

摘　　要：目的 建立基于^(18)F-FDG PET/CT影像组学特征和临床风险因素的肺腺癌表皮生长因子受体(EGFR)突变预测模型,并验证其准确性。资料与方法 回顾性纳入2018年7月—2022年1月在新疆医科大学附属肿瘤医院行^(18)F-FDG PET/CT,并经病理证实且有EGFR基因检测结果的肺腺癌患者155例。按照7∶3随机划分为训练集和测试集。对所有EGFR突变临床风险因素进行单因素及多因素Logistic回归,建立预测肺腺癌EGFR突变的临床风险因素模型。所有图像经过预处理后,分别在PET、CT图像上勾画感兴趣区,提取影像组学特征。经特征降维后,基于最佳影像组学特征建立支持向量机、随机森林分类器(RF)和逻辑回归模型,并预测肺腺癌EGFR突变。采用受试者工作特征曲线评估3种模型的诊断效能。然后选取最佳模型结合临床风险因素构建复合模型,并绘制列线图。应用决策分析曲线评估列线图的临床效用。结果 155例肺腺癌中,EGFR野生型69例,突变型86例。多因素Logistic回归显示吸烟为预测肺腺癌EGFR突变的临床风险预测因子。影像组学分析,经特征提取、特征筛选、降维后保留了15个CT影像组学特征和3个PET影像组学特征。3种不同机器学习(支持向量机、RF、逻辑回归)方法在影像组学特征中预测EGFR突变状态训练集曲线下面积分别为0.769、0.857、0.783,测试集分别为0.777、0.775、0.764。经Delong检验,RF模型显示出更好的预测性能。通过融合临床风险因素与影像组学特征,构建复合模型,训练集曲线下面积为0.882;测试集曲线下面积为0.822。决策曲线显示,复合模型获益较大,具有较好的净获益。结论 纳入的吸烟史和RF模型预测分数构建的复合模型对预测肺腺癌的EGFR突变有较大的应用价值。Purpose To establish and validate a prediction model of EGFR mutation status in patients with lung adenocarcinoma based on ~(18)F-FDG PET/CT radiomic features and clinical risk factors,and verify its accuracy.Materials and Methods A total of 155 patients with lung adenocarcinoma who underwent ^(18)F-FDG PET/CT and were pathologically confirmed with EGFR gene test from July 2018 to January 2022 in the Affiliated Tumor Hospital of Xinjiang Medical University were retrospectively included.All patients were randomly divided into training set and test set in the ratio of 7∶3.Univariate and multivariate Logistic regression was performed on all EGFR mutation clinical risk factors to establish a clinical risk factor model for predicting EGFR mutations in patients with lung adenocarcinoma.After preprocessing,all images were subjected to region of interest outlining on PET and CT images,respectively,to extract radiomics features.After feature dimensionality reduction,support vector machine,random forest classifier(RF) and Logistic regression models were built based on the best radiomics features to predict EGFR mutations in patients with lung adenocarcinoma.The diagnostic efficacy of the three models was evaluated via receiver operating characteristic curves.The best model was further selected to construct a composite model via combining clinical risk factors,and a column-line diagram was drawn.Decision curve analysis curves were applied to assess the clinical utility of the column charts.Results Of 155 cases with lung adenocarcinoma,there were 69 patients with EGFR wild-type and 86 patients with mutant.Smoking was exclusively shown to be the clinical risk predictor for predicting EGFR mutations in patients with lung adenocarcinoma via multifactorial Logistic regression.The radiomics analysis showed that 15 CT radiomics features and 3 PET radiomics features were retained after feature extraction,feature screening,and dimensionality reduction.Three different machine learning methods(including support vector machine,RF,and

关 键 词：肺腺癌 正电子发射断层显像术 氟脱氧葡萄糖F18 表皮生长因子 预测 

分 类 号：R734.2[医药卫生—肿瘤] R445[医药卫生—临床医学] R730.44