基于CT影像组学特征的机器学习和SHAP方法预测肝细胞癌微血管浸润状态  

The prediction of microvascular invasion status in hepatocellular carcinoma by machine learning,and SHAP method based on CT imaging radiomics features

作　　者：陈鹏 郑屹峰[1] 张成孟 沈伟强[1] 夏冬 沈健[1] 付治中 周玮[1] Chen Peng

机构地区：[1]浙江省湖州市中心医院,313000 [2]四川成都电子科技大学生命科学与技术学院,610000

出　　处：《浙江临床医学》2025年第3期329-332,336,共5页Zhejiang Clinical Medical Journal

基　　金：湖州市公益性应用研究项目(2020GY06)。

摘　　要：目的探讨基于CT影像组学特征构建机器学习模型预测肝细胞癌(HCC)微血管浸润(MVI)状态。方法回顾性分析经组织病理学明确证实为HCC且已评估MVI状态的108例患者,依据MVI状态将其划分为MVI(+)组41例及MVI(-)组67例。所有患者均于术前2周内接受腹部CT平扫与双期增强扫描。通过手动方式勾画全肿瘤区域,并提取影像组学特征。在训练集中,应用最大相关最小冗余(mRMR)以及最小绝对收缩和选择算子(LASSO)方法对影像组学特征降维处理。降维后的影像组学特征,分别采用随机森林(RF)、逻辑回归(LR)、支持向量机(SVM)、K近邻(KNN)、梯度提升决策树(lightGBM)、自适应增强机(AdaBoost)这6种模型对MVI进行预测。采用受试者工作特征曲线(ROC)的曲线下面积(AUC)和Delong检验对6种机器学习模型的预测准确度进行评估。从上述6种机器学习模型中选出预测准确度最优的机器学习模型,采用SHAP方法分析各影像组学特征与MV状态的相关性。计算每例患者SHAP值,并对机器学习模型的特征重要性进行可视化。结果经过降维和筛选流程后,最终有7个影像组学特征(其中包括5个平扫特征、1个动脉期特征以及1个静脉期特征)被应用于模型构建。在6种分类器模型中,RF模型表现最为优异,其训练组AUC为0.988,验证组AUC为0.901。在SHAP方法中,通过混淆矩阵可知,RF模型对验证集准确预测27例,其中阳性患者准确预测4例,阴性患者准确预测23例。在计算每例患者SHAP值后发现,SHAP值越大,预测结果越倾向于MVI(+);反之,预测结果则更偏向于MVI(-)。结论基于CT影像组学特征所构建的机器学习模型能够有效预测HCC病灶的MVI状态,借助SHAP方法解释了机器学习模型的决策过程,对于临床指导制定合理的个性化治疗方案具有重要的辅助价值。Objective To explore the accurate prediction of the microvascular invasion(MVI)status of hepatocellular carcinoma(HCC)by constructing a machine learning model based on the radiomics features of CT.Methods A retrospective analysis was conducted on 108 patients with HCC confirmed by histopathology and evaluated for MVI status.They were divided into the MVI(+)group(41 cases)and the MVI(-)group(67 cases)according to the MVI status.All patients underwent abdominal CT plain scan and two-phase enhanced scan within two weeks prior to surgery.The entire tumor area was manually mapped,and the radiomics features were extracted.In the training set,the max-relevance and min-redundancy(mRMR)and the least absolute shrinkage and selection operator(LASSO)methods were used to reduce the dimensionality of the radiomics features.For the radiomics features after dimensionality reduction,six classifiers,namely random forest(RF),logistic regression(LR),support vector machine(SVM),K nearest neighbor(KNN),light gradient boosting machine(lightGBM),and adaptive boosting(AdaBoost),were utilized to build machine learning models.The six machine learning models could precisely predict the microvascular invasion(MVI)of hepatocellular carcinoma(HCC).The predictive accuracy of the six machine learning models was evaluated using the area under the curve(AUC)of the receiver operating curve(ROC)and Delong tests.The machine learning model with the best prediction accuracy was selected from the above six machine learning models and the SHAP method was used for analysis to clarify the specific correlation between the radiomics features and the MVI status of the patients.SHAP values were calculated for each patient and the feature importance of the machine learning model was visualized.Results After the reduction and screening process,a total of 7 radiomics features(including 5 plain scan features,1 arterial phase feature,and 1 venous phase feature)were ultimately utilized for model construction.Among the six classifier models,the RF model performed the

关 键 词：肝细胞癌 微血管浸润 体层摄影术 X线计算机 影像组学 机器学习 

分 类 号：R735.7[医药卫生—肿瘤]