影像组学联合机器学习预测胃癌脉管癌栓及神经侵犯的可行性研究  

Feasibility of radiomics combined with machine learning in predicting lymphovascular and perineural invasion of gastric cancer

作　　者：艾双全 杨苗 袁子龙[1] 贺瑶瑶 聂婷婷 刘玉林[1] AI Shuangquan;YANG Miao;YUAN Zilong;HE Yaoyao;NIE Tingting;LIU Yulin(Deparment of Radiology,Hubei Cancer Hospital,Wuhan 430079,China;School of Biomedical Engineering,South-Central Minzu University,Wuhan 430000,China)

机构地区：[1]湖北省肿瘤医院放射科,湖北武汉430079 [2]中南民族大学生物医学工程学院,湖北武汉430000

出　　处：《实用放射学杂志》2024年第5期746-751,共6页Journal of Practical Radiology

摘　　要：目的探讨基于CT扫描所获得的影像组学特征联合不同机器学习方法预测胃癌脉管癌栓及神经侵犯的可行性。方法回顾性选取经手术病理证实的胃癌脉管癌栓患者142例,其中阳性96例,阴性46例;神经侵犯患者137例,其中阳性76例,阴性61例。采用3D-Slicer包勾画,使用Pyradiomics包提取影像组学特征,按照8︰2的比例将数据分为训练集和测试集。采用组内相关系数(ICC)、Pearson相关性分析以及最小绝对收缩和选择算子(LASSO)算法进行特征筛选。再利用支持向量机(SVM)、K最邻近(KNN)、决策树(DT)、随机森林(RF)、极端树(ET)、极端梯度提升树(XGBoost),LightGBM7个分类器对脉管癌栓及神经侵犯分别进行建模比较。采用受试者工作特征(ROC)曲线以及曲线下面积(AUC)评估模型的预测效能。结果SVM、KNN、DT、RF、ET、XGBoost,LightGBM7个分类器在脉管癌栓组训练集中对应的AUC分别为0.926、0.753、1.000、0.999、1.000、1.000、0.917,测试集AUC分别为0.894、0.692、0.456、0.678、0.753、0.650、0.650;神经侵犯组训练集中对应的AUC分别为0.864、0.794、1.000、1.000、1.000、1.000、0.866,测试集AUC分别为0.861、0.706、0.700、0.672、0.731、0.667、0.678。结论基于静脉期CT影像组学特征联合机器学习方法可以在术前对胃癌脉管癌栓和神经侵犯进行预测;在诸多分类器中,SVM对胃癌脉管癌栓及神经侵犯的预测效果最好。Objective To explore the feasibility of radiomics features combined with different machine learning methods based on CT scans to predict lymphovascular and perineural invasion in patient with gastric cancer.Methods A total of 142 patients with gastric cancer lymphovascular confirmed by operative pathological examination were retrospectively selected.Among all patients,there were 96 positive cases and 46 negative cases.Among 137 patients with perineural invasion,there were 76 positive cases and 61 negative cases.The 3D-Slicer package was used for delineation,and the Pyradiomics package was used to extract radiomics features.All data were randomly divided into training set and test set in an 8:2 ratio.Intraclass correlation coefficient(ICC),Pearson correlation analysis,least absolute shrinkage and selection operator(LASSO)algorithm were used for feature selection.Support vector machine(SVM),K-nearest neighbor(KNN),decision tree(DT),random forest(RF),extreme tree(ET),extreme gradient boosting(XGBoost),and LightGBM were used to compare the models of lymphovascular and perineural invasion,respectively.Receiver operating characteristic(ROC)curve and area under the curve(AUC)were used to evaluate the predictive performance of these models.Results The lymphovascular group AUC of SVM,KNN,DT,RF,ET,XGBoost,and LightGBM in the training set were 0.926,0.753,1.000,0.999,1.000,1.000,and 0.917,and the AUC in the test set were 0.894,0.692,0.456,0.678,0.753,0.650,and 0.650,respectively.The perineural invasion group AUC of SVM,KNN,DT,RF,ET,XGBoost,and LightGBM in the training set were 0.864,0.794,1.000,1.000,1.000,1.000,and 0.866,and the AUC in the test set were 0.861,0.706,0.700,0.672,0.731,0.667,and 0.678,respectively.Conclusion Based on venous phase CT radiomics features combined with machine learning methods,it is feasible to predict lymphovascular and perineural invasion of gastric cancer preoperatively.Among the variousmachine learning methods,SVM shows the best predictive performance for lymphovascular and perineural invasio

关 键 词：胃癌 机器学习 脉管癌栓 神经侵犯 影像组学 计算机体层成像 

分 类 号：R735.2[医药卫生—肿瘤] TP181[医药卫生—临床医学] R445[自动化与计算机技术—控制理论与控制工程] R814.42[自动化与计算机技术—控制科学与工程]