≤20 mm孤立性肺结节良恶性预测模型的建立与验证

Establishment and verification of prediction model for benign or malignant of≤20 mm solitary pulmonary nodules

目的基于人工智能(AI)建立并验证≤20 mm孤立性肺结节(SPN)良恶性预测模型。方法收集2018年11月至2020年5月在厦门大学附属中山医院接受手术切除并获得明确病理诊断的≤20 mm SPN患者279例(338个SPN),回顾性分析其临床特征(年龄、性别、吸烟史、恶性肿瘤史及家族恶性肿瘤史)、影像特征(最大径、最小径、实性占比、体积、分叶征、毛刺征、空泡征、空洞征、胸膜凹陷征)、及影像组学特征(最大CT值、最小CT值、平均CT值、中位数CT值、CT值标准差、偏度、峰值、能量、熵)。采用完全随机法将SPN按8∶2比例分为训练集(271个)和验证集(67个)。训练集数据中,首先使用最小收缩和选择算子(LASSO)回归方法对临床特征、影像特征及影像组学进行筛选,再进行多因素logistic回归分析筛选出≤20 mm SPN良恶性相关的独立危险因素,并实现列线图预测模型构建。最后将测试集数据传入该模型进行验证,绘制ROC曲线和校准曲线,评估模型预测价值。结果训练集中271个≤20 mm SPN,其中良性81个、恶性190个。经LASSO回归及多因素logistics回归分析筛选得出年龄、性别、最大径、空泡征、实性占比5个因素为预测最大径≤20 mm SPN良恶性的独立预测因子,构建预测模型为:P=e^(x)/(1+e^(x)),x=-2.583+0.027×年龄+1.519×性别+0.127×结节最大径-2.132×实性占比+1.720×空泡征。该模型预测≤20 mm的SPN为恶性的ROC曲线下面积为0.850,灵敏度为73.7%,特异度为82.7%,准确度为82.3%。验证集67个SPN,其中良性22个、恶性45个,预测模型的AUC为0.882,灵敏度为82.2%,特异度为81.8%,准确度为85.1%。训练集和验证集预测模型的校准曲线与理想曲线重合度良好(训练集:P=0.688,验证集:P=0.618)。结论基于AI建立的≤20 mm SPN的良恶性预测模型可获得预测概率并具有良好的诊断效能。

Objective To establish and verify the prediction model of benign or malignant of solitary pulmonary nodules(SPNs≤20 mm)based on artificial intelligence.Methods Totally 338 SPNs(≤20 mm)from 279 patients,confirmed by operation and pathology,were selected in Zhongshan Hospital Xiamen University from November 2018 to May 2020.Clinical data(age,gender,smoking history,individual and family history of malignancy),image features(maximum diameter,minimum diameter,solid proportion,volume,lobulation sign,burr sign,vacuole sign,cavity sign,pleural indentation sign,and radiomic features(maximum CT value,minimum CT value,average CT value,median CT value,CT value standard deviation,skewness,peak,energy,entropy)were analyzed retrospectively.All the data of patients were randomly divided into training set(271 SPNs)and test set(67 SPNs).In the training set,the clinical,image features and radiomic features were first selected by the least absolute shrinkage and selection operator(LASSO)regression,then the independent risk factors of SPN(≤20 mm)were screened out by multi-variate logistic regression analysis,and the nomogram prediction models were constructed.Finally,the data of test set were used to verify the prediction model by the ROC curve and calibration curve(CC).Results In the training set of 271 SPNs,81 SPNs were benign and 190 malignant.After analysis of LASSO regression and multi-factor logistics regression,the independent predictors of benign or malignant SPN were age,gender,largest diameter,vacuole sign and solid proportion.The prediction model was P=e^(x)/(1+e^(x)),x=-2.583+0.027×age+1.519×gender+0.127×maximum diameter-2.132×solid proportion+1.720×vacuole sign.The results of the model showed that the area under curve(AUC)of ROC was 0.850,and the sensitivity was 73.7%,specificity was 82.7%and accuracy was 82.3%.In the test set of 67 SPNs,22 SPNs were benign and 45 malignant.The results showed that the AUC of ROC was 0.882,and the sensitivity was 82.2%,specificity was 81.8%and accuracy was 85.1%.The calibration nomogram of prediction model showed that CC from training set or test set well coincided with its individual ideal curve(Ptraining=0.688,Ptest=0.618).Conclusion Prediction model of benign or malignant SPN≤20 mm is established based on AI;it can obtain the prediction probability and has good diagnostic efficiency.