摘要
目的探讨基于钆塞酸二钠增强MRI肝胆期影像组学特征的不同机器学习模型术前预测肝细胞癌(HCC)微血管侵犯(MVI)的价值。方法回顾性分析2015年6月至2020年6月在苏州大学附属第一医院经病理证实的132例HCC患者的资料,MVI阳性72例、阴性60例。按照7∶3的比例以随机种子法分为训练集和验证集。利用PyRadiomics软件提取肝胆期图像影像组学特征,采用最小绝对收缩和选择算子(LASSO)回归5折交叉验证法对训练集临床和影像组学特征进行筛选,得到最优特征子集,然后用6种机器学习方法(决策树、极端梯度提升、随机森林、支持向量机、广义线性模型和神经网络)构建预测模型,采用ROC曲线评估模型的预测能力,采用DeLong检验比较6种机器学习算法曲线下面积(AUC)的差异。结果经LASSO回归筛选后获得14个特征组成最优特征子集,包括2个临床特征(肿瘤最大径和甲胎蛋白)和12个影像组学特征。训练集中基于最优特征子集构建的决策树、极端梯度提升、随机森林、支持向量机、广义线性模型和神经网络模型预测HCC MVI的AUC值分别为0.969、1.000、1.000、0.991、0.966和1.000,验证集的AUC值分别为0.781、0.890、0.920、0.806、0.684和0.703。验证集中,极端梯度提升与广义线性模型、神经网络的AUC的差异有统计学意义(Z=2.857、3.220,P=0.004、0.001),随机森林与支持向量机、广义线性模型和神经网络AUC的差异有统计学意义(Z=2.371、3.190、3.967,P=0.018、0.001、<0.001),支持向量机与广义线性模型AUC的差异有统计学意义(Z=2.621,P=0.009),其余机器学习模型间AUC的差异均无统计学意义(P>0.05)。结论基于钆塞酸二钠增强MRI肝胆期图像的影像组学特征构建的机器学习模型可用于术前预测HCC MVI,其中,极端梯度提升和随机森林模型具有较高的预测效能。
Objective To explore the value of different machine learning models based on Gd-EOB-DTPA enhanced MRI hepatobiliary phase radiomics features in preoperative prediction of microvascular invasion(MVI)of hepatocellular carcinoma(HCC).Methods The data of 132 patients with HCC confirmed by pathology in the First Affiliated Hospital of Soochow University from January 2015 to May 2020 were retrospectively analyzed,including 72 cases of positive MVI and 60 cases of negative MVI.According to the proportion of 7∶3,the cases were randomly divided into training set and validation set.The radiomics features of hepatobiliary phase images for HCC were extracted by PyRadiomics software.The clinical and radiomics features of the training set were screened by the least absolute shrinkage and selection operator(LASSO)regression with 5 fold cross-validation,and then the optimal feature subset was obtained.Six machine learning algorithms,including decision tree,extreme gradient boosting,random forest,support vector machine(SVM),generalized linear model(GLM)and neural network,were used to build the prediction models,and the ROC curves were used to evaluate the prediction ability of the models.DeLong test was used to compare the differences of area under the curve(AUC)for 6 machine learning algorithms.Results Totally 14 features selected by LASSO regression were obtained to form the optimal feature subset,including 2 clinical features(maximum tumor diameter and alpha-fetoprotein)and 12 radiomics features.The AUCs of decision tree,extreme gradient boosting,random forest,SVM,GLM and neural network based on the optimal feature subset were 0.969,1.000,1.000,0.991,0.966,1.000 in the training set and 0.781,0.890,0.920,0.806,0.684,0.703 in the validation set,respectively.There were significant differences in the AUCs between extreme gradient boosting and GLM or neural network(Z=2.857,3.220,P=0.004,0.001).The differences in AUCs between random forest and SVM,GLM,or neural network were significant(Z=2.371,3.190,3.967,P=0.018,0.001,<0.001).The
作者
郁义星
王希明
胡春洪
范艳芬
胡梦洁
诗涔
朱默
张妤
胡粟
Yu Yixing;Wang Ximing;Hu Chunhong;Fan Yanfen;Hu Mengjie;Shi Cen;Zhu Mo;Zhang Yu;Hu Su(Department of Radiology,the First Affiliated Hospital of Soochow University,Institute of Imaging Medicine,Soochow University,Suzhou 215006,China)
出处
《中华放射学杂志》
CAS
CSCD
北大核心
2021年第8期853-858,共6页
Chinese Journal of Radiology
基金
国家自然科学基金(81801692)
国家重点研发计划(2017YFC0114300)
苏州市科技计划项目(SYS2020125,SS2019057)
苏州市民生科技示范工程(SS201808)。
关键词
癌
肝细胞
磁共振成像
影像组学
机器学习
微血管侵犯
Carcinoma,hepatocellular
Magnetic resonance imaging
Radiomics
Machine learning
Microvascular invasion
