论文部分内容阅读
目的:探究n 18F-脱氧葡萄糖(FDG) PET/CT影像结合机器学习算法对肺腺癌(LUAD)表皮生长因子受体(EGFR)突变亚型的预测价值。n 方法:回顾性收集2016年4月至2020年5月于天津医科大学肿瘤医院行n 18F-FDG PET/CT检查的238例LUAD患者的影像及病理资料，其中EGFR突变型为126例(男54例，女72例，中位年龄62岁)；EGFR野生型112例(男68例，女44例，中位年龄61岁)。分别对PET、CT图上感兴趣体积(VOI)进行三维勾画，提取三维和二维影像组学特征。使用K-最近邻(KNN)、支持向量机(SVM)和Adaboost分别对CT、PET和PET/CT融合组学特征进行学习，并对EGFR突变亚型进行预测。使用受试者工作特征(ROC)曲线对预测性能进行评估。n 结果:126例EGFR突变型患者中3例为18号外显子突变，6例为20号外显子突变，42例为19号外显子突变，75例为21号外显子突变。前2个亚型患者因数量较少，分类器难以进行充分的训练而被排除。PET/CT平均融合特征模型预测EGFR突变亚型的效果[Adaboost：曲线下面积(AUC)＝0.87, 95% n CI：0.75～0.99]优于PET特征模型(Adaboost：AUC＝0.64, 95% n CI：0.46～0.83；n z＝2.04，n P<0.05)和CT特征模型(Adaboost：AUC＝0.64, 95%n CI：0.45～0.83; n z＝2.06，n P0.05)。n 结论:机器学习结合n 18F-FDG PET/CT融合影像组学特征在EGFR突变亚型预测中有一定的价值。n “,”Objective:To explore the predictive values for mutation subtypes of epidermal growth factor receptor (EGFR) in patients with lung adenocarcinoma (LUAD) based on machine learning and n 18F-fluorodeoxyglucose (FDG) PET/CT images.n Methods:18F-FDG PET/CT images and pathological data of 238 patients with LUAD (126 patients (54 males, 72 females, median age 62 years) with EGFR mutation; 112 patients (68 males, 44 females, median age 61 years) with wild-type EGFR)) were retrospectively collected at Tianjin Medical University Cancer Institute and Hospital between April 2016 and May 2020. Volumes of interest (VOI) of PET and CT images were delineated respectively and three-dimensional-based and two-dimensional-based radiomics features were extracted from VOIs. Three machine learning classifiers of K-nearest neighbor (KNN), support vector machine (SVM) and Adaboost were trained in training set with CT, PET and fusion PET/CT radiomics features respectively. Well trained classifiers were tested in test set. Each predictive model was evaluated by using the receiver operating characteristic (ROC) curve.n Results:A total of 126 patients were EGFR mutation including 3 patients with 18 exon mutation, 6 patients with 20 exon mutation, 42 patients with 19 exon mutation, and 75 patients with 21 exon mutation. Finally, patients with 18 exon mutation and 20 exon mutation were removed due to the scale was too small to be trained adequately by machine learning classifiers. Predictive performance of mean PET/CT feature-based model (Adaboost: area under curve (AUC)=0.87, 95% n CI: 0.75-0.99) in EGFR mutation subtypes was better than PET feature-based model (Adaboost: AUC=0.64, 95% n CI: 0.46-0.83; n z=2.04, n P<0.05) and CT feature-based model (Adaboost: AUC=0.64, 95%n CI: 0.45-0.83; n z=2.06, n P0.05).n Conclusion:Machine learning and n 18F-FDG PET/CT radiomics features can provide predictive value for EGFR mutation subtypes in patients with LUAD.n