基于静脉期增强CT影像组学的机器学习模型术前预测局部进展期胃癌脉管侵犯

Preoperative prediction of vessel invasion in locally advanced gastric cancer based on venous phase enhanced CT radiomics and machine learning

目的评估基于静脉期增强CT影像组学特征的机器学习模型术前预测局部进展期胃癌脉管侵犯(VI)的价值。方法回顾性分析2011年7月至2020年12月郑州大学第一附属医院经病理证实的296例局部进展期胃癌患者,VI阳性213例、阴性83例,采用分层抽样方法按7∶3的比例将数据分为训练集(207例)和测试集(89例)。记录患者临床特征,采用多因素logistic回归筛选胃癌VI的独立危险因素。利用Pyradiomics软件提取肿瘤静脉期CT影像组学特征,采用最小绝对收缩和选择算法(LASSO)进行特征筛选,得到最优特征子集,建立影像组学标签。使用极端梯度提升(XGBoost)、逻辑回归(logistic)、朴素贝叶斯(GNB)和支持向量机(SVM)4种机器学习算法,对影像组学标签和筛选出的临床独立危险因素构建预测模型。采用受试者操作特征曲线评估模型预测胃癌VI的效能。结果分化程度(OR=13.651,95%CI 7.265~25.650,P=0.003)、Lauren分型(OR=1.349,95%CI 1.011~1.799,P=0.042)和CA199(OR=1.796,95%CI 1.406~2.186,P=0.044)是预测局部进展期胃癌VI的独立危险因素。基于静脉期增强CT图像提取了864个影像组学特征,经LASSO筛选出18个最优特征构建组学标签。训练集中,XGBoost、logistic、GNB和SVM模型预测胃癌VI的曲线下面积(AUC)分别为0.914(95%CI 0.875~0.953)、0.897(95%CI 0.853~0.940)、0.880(95%CI 0.832~0.928)和0.814(95%CI 0.755~0.873),测试集中分别是0.870(95%CI 0.769~0.971)、0.877(95%CI 0.788~0.964)、0.859(95%CI 0.755~0.961)和0.773(95%CI 0.647~0.898)。logistic模型在测试集中AUC最大且稳定性高。结论基于静脉期增强CT影像组学特征的机器学习模型术前预测局部进展期胃癌VI均具有较高的效能,其中logistic模型的诊断效能最佳。

Objective To evaluate the value of preoperative prediction of vessel invasion(VI)of locally advanced gastric cancer by machine learning model based on the venous phase enhanced CT radiomics features.Methods A retrospective analysis of 296 patients with locally advanced gastric cancer confirmed by pathology in the First Affiliated Hospital of Zhengzhou University from July 2011 to December 2020 was performed.The patients were divided into VI positive group(n=213)and VI negative group(n=83)based on pathological results.The data were divided into training set(n=207)and test set(n=89)according to the ratio of 7∶3 with stratification sampling.The clinical characteristics of patients were recorded,and the independent risk factors of gastric cancer VI were screened by multivariate logistic regression.Pyradiomics software was used to extract radiomic features from the venous phase enhanced CT images,and the minimum absolute shrinkage and selection algorithm(LASSO)was used to screen the features,obtain the optimal feature subset,and establish the radiomics signature.Four machine learning algorithms,including extreme gradient boosting(XGBoost),logistic,naive Bayes(GNB),and support vector machine(SVM)models,were used to build prediction models for the radiomics signature and the screened clinical independent risk factors.The efficacy of the model in predicting gastric cancer VI was evaluated by the receiver operating characteristic curve.Results The degree of differentiation(OR=13.651,95%CI 7.265-25.650,P=0.003),Lauren′s classification(OR=1.349,95%CI 1.011-1.799,P=0.042)and CA199(OR=1.796,95%CI 1.406-2.186,P=0.044)were independent risk factors for predicting the VI of locally advanced gastric cancer.Based on the venous phase enhanced CT images,864 quantitative features were extracted,and 18 best constructed radiomics signature were selected by LASSO.In the training set,the area under the curve(AUC)of XGBoost,logistic,GNB and SVM models for predicting gastric cancer VI were 0.914(95%CI 0.875-0.953),0.897(95%CI 0.853-0.940),0.880(95%CI 0.832-0.928)and 0.814(95%CI 0.755-0.873),respectively,and in the test set were 0.870(95%CI 0.769-0.971),0.877(95%CI 0.788-0.964),0.859(95%CI 0.755-0.961)and 0.773(95%CI 0.647-0.898).The logistic model had the largest AUC in the test set.Conclusions The machine learning model based on the venous phase enhanced CT radiomics features has high efficacy in predicting the VI of locally advanced gastric cancer before the operation,and the logistic model demonstrates the best diagnostic efficacy.