基于临床及CT影像组学特征预测肺腺癌结节

Clinical and CT radiomic features for predicting pulmonary nodules in adenocarcinoma

目的:探讨基于临床及影像组学特征构建机器学习模型对预测肺腺癌结节的准确性。方法:回顾性收集186例病理类型明确的肺结节患者,按病理类型分为腺体前驱病变组与腺癌组,按照7∶3比例将其分为训练集和测试集。采用3D Slicer软件对病灶容积感兴趣区(ROI)进行逐层手动勾画,通过Python软件提取影像组学特征。提取临床特征,包括人口统计学特征、临床表现、肿瘤标志物及CT影像学语义特征。选用单因素分析、最小绝对收缩和选择算子(LASSO)和逐步logistic回归分析进行特征筛选。在训练集中分别构建基于影像组学特征(模型1)及临床与影像组学特征相结合(模型2)的随机森林(LR)肺腺癌结节预测模型。通过ROC曲线及计算曲线下面积(AUC)对模型进行验证。结果:训练集130例,测试集56例。提取影像组学特征和临床特征数量分别为688个和25个。经特征筛选,共保留11个影像组学特征。临床特征中年龄、结节成分、结节最大径在训练集中组间差异显著(P<0.05)。训练集中模型1和模型2的AUC分别为0.991和0.960;测试集中模型1和模型2的AUC分别为0.913和0.884,准确率分别为0.875和0.839,精确度分别为0.872和0.824,召回率分别为0.976和1.0,F1分数为0.921和0.903。结论:基于临床及CT影像组学特征构建的RF模型能够准确预测肺腺癌结节。

Purpose:To investigate the accuracy of machine learning models based on clinical and radiomic features in predicting pulmonary nodules in adenocarcinoma.Methods:A total of 186 patients with definite pathological types of pulmonary nodules were collected retrospectively.Patients were divided into precursor glandular lesions group and adenocarcinoma group according to pathological types.All patients were divided into training set and test set according to the ratio of 7:3.Region of interest(ROI)of volumetric lesions were delineated manually layer by layer by 3D-Slicer software.Radiomic features were extracted by Python software.Clinical features were collected,including demographic features,clinical symptoms,tumor markers and CT semantic features.Univariate analysis,least absolute shrinkage and selection operator(LASSO)and stepwise logistic regression analysis were used to select features.Random forest(RF)models for predicting pulmonary nodules in adenocarcinoma were constructed based on only radiomic features(model 1)and the combination of clinical features and radiomic features(model 2).Predictive performance of the models was evaluated by drawing ROC curve and calculating the area under ROC curve(AUC).Results:There were 130 cases in training set and 56 cases in test set.A total of 688 radiomic features and 25 clinical features were extracted.11 radiomic features were selected after features selection.There were statistically significant differences in clinical features of age,components of nodule and maximal diameter of nodule between the two groups in training set(P<0.05).The AUC of model 1 and model 2in training set was 0.991 and 0.960,respectively.The results of model 1 and model 2 in test set were as follows:AUC was 0.913 and 0.884,accuracy was 0.875 and 0.839,precision was 0.872 and 0.824,recall was 0.976 and 1.0,F1score was 0.921 and 0.903,respectively.Conclusion:RF models based on clinical and radiomic features can predict pulmonary nodules in adenocarcinoma accurately.