CT增强影像组学模型鉴别肾脏嗜酸细胞腺瘤与肾脏嫌色细胞癌的价值  

The value of contrast-enhanced CT radiomics model in differentiating renal oncocytoma from chromophobe renal cell carcinoma

作　　者：李可 师毅冰[1] 梁弦弦 赵恒亮 郭迪 LI Ke;SHI Yibing;LIANG Xianxian;ZHAO Hengliang;GUO Di(Department of Radiology,Xuzhou Central Hospital,Xuzhou,Jiangsu Province 221009,China;Department of Orthopedics,Xuzhou Central Hospital,Xuzhou,Jiangsu Province 221009,China)

机构地区：[1]徐州市中心医院影像科,江苏徐州221009 [2]徐州市中心医院骨科,江苏徐州221009

出　　处：《实用放射学杂志》2025年第3期452-456,共5页Journal of Practical Radiology

摘　　要：目的 探讨基于CT增强的影像组学机器学习模型鉴别肾脏嗜酸细胞腺瘤(RO)和肾脏嫌色细胞癌(chRCC)的价值.方法 回顾性分析经病理证实且临床及影像资料完整的RO及chRCC患者65例,按照7︰3的比例随机分为训练集(45例)和测试集(20例).使用3D Slicer软件在患者术前CT图像上勾画出肿瘤边界,后用Radiomics插件提取影像组学特征,使用单因素分析法、递归特征消除法(RFE)、最小绝对收缩和选择算子(LASSO)算法筛选最佳的影像组学特征.在训练集上构建3种机器学习模型,并用网格搜索法选择最佳超参数组合.采用受试者工作特征(ROC)曲线、校准曲线、决策曲线评估训练集、测试集上各机器学习模型的效能.结果 随机森林模型、逻辑回归模型、支持向量机模型均可以较好鉴别RO与chRCC.在训练集中,随机森林模型、支持向量机模型的曲线下面积(AUC)分别为0.950[95%置信区间(CI)0.901~0.998]、0.955(95%CI 0.908~1.000),高于逻辑回归模型的AUC 0.882(95%CI 0.806~0.956),DeLong检验有统计学差异(P<0.05);在测试集中,随机森林模型、逻辑回归模型、支持向量机模型AUC分别为0.876(95%CI 0.758~0.993)、0.883(95%CI 0.768~0.997)、0.883(95%CI 0.768~0.997),DeLong检验各模型AUC无明显统计学差异(P>0.05).决策曲线表明3种模型均有明显临床净获益.结论 基于CT增强的影像组学机器学习模型可以有效鉴别RO和chRCC.Objective To investigate the value of machine learning models based on contrast-enhanced CT radiomics in differentiating renal oncocytoma(RO)from chromophobe renal cell carcinoma(chRCC).Methods A total of 65 patients with RO and chRCC confirmed by pathology with complete clinical and imaging data were analyzed retrospectively.The patients were randomly divided into training set(n=45)and test set(n=20)according to a ratio of 7︰3.The tumor boundaries were delineated on the preoperative CT images using 3D Slicer software,and radiomics features were extracted using the Radiomics plugin.Univariate analysis,recursive feature elimination(RFE),least absolute shrinkage and selection operator(LASSO)algorithms were used to select the best radiomics features.Three machine learning models were constructed on the training set and the grid search method was used to select the best combination of hyperparameters.The receiver operating characteristic(ROC)curve,calibration curve and decision curve were used to evaluate the performance of each machine learning model on the training set and test set.Results Random forest model,logistic regression model and support vector machine model can better identify RO and chRCC.In the training set,the area under the curve(AUC)of random forest model and support vector machine model were 0.950[95%confidence interval(CI)0.901-0.998]and 0.955(95%CI 0.908-1.000),respectively,which were higher than the AUC of logistic regression model 0.882(95%CI 0.806-0.956).Statistical differences were found by DeLong test(P<0.05);In the test set,the AUC of random forest model,logistic regression model and support vector machine model were 0.876(95%CI 0.758-0.993),0.883(95%CI 0.768-0.997)and 0.883(95%CI 0.768-0.997),respectively.There was no significant statistical difference in the AUC of each model by DeLong test(P>0.05).The decision curve showed that all three models had significant net clinical benefits.Conclusion The machine learning model based on contrast-enhanced CT radiomics can effectively distinguish

关 键 词：影像组学 机器学习 肾脏嫌色细胞癌 肾脏嗜酸细胞腺瘤 

分 类 号：R445[医药卫生—影像医学与核医学] TP181[医药卫生—诊断学] R737.11[医药卫生—临床医学]