摘要
目的:探讨基于机器学习的超声影像组学模型在术前预测乳腺癌患者程序性死亡配体1(PD-L1)表达水平的应用价值。方法:收集2018年4月至2024年3月在丽水市中心医院经病理学证实的在术前接受常规超声检查的177例乳腺癌患者,按7:3随机分为训练集(124例)和测试集(53例),提取超声图像上病灶的影像组学特征。组内相关系数用于去除一致性较差的特征,Pearson用于去除冗余特征,最小绝对收缩与选择算法用于对保留特征的进一步降维,最终筛选出与PD-L1表达水平相关的最优特征。利用Logistic回归、支持向量机和极端梯度提升决策树(XGBoost)分别建立影像组学模型;单因素及多因素Logistic回归用于筛选临床危险因素并构建临床模型;基于影像组学评分和临床危险因素构建列线图模型。通过ROC曲线评估不同模型的诊断性能,并计算出AUC、灵敏度、特异度和准确度。结果:分化程度低、Ki-67高表达是术前预测PD-L1表达水平的临床危险因素。在训练集和测试集中,XGBoost模型均表现出最高的诊断性能,AUC分别为0.833、0.776。进一步基于分化程度、Ki-67表达和XGBoost影像组学评分构建列线图模型,结果显示,其在训练集中的AUC、灵敏度、特异度、准确度分别为0.902、80.39%、91.78%、86.73%,在测试集中的AUC、灵敏度、特异度、准确度分别为0.835、86.36%、77.42%、80.90%。结论:采用XGBoost建立的超声影像组学模型对乳腺癌PD-L1表达水平具有较高的预测价值,联合临床危险因素建立的列线图可以进一步提升诊断性能。
Objective:To explore the application value of machine learning-based ultrasound radiomics model in predicting the expression level of programmed death-ligand 1(PD-L1)in breast cancer patients before surgery.Methods:A total of 177 breast cancer patients who underwent routine ultrasound examinations before surgery who were confirmed by pathology at Lishui Central Hospital were randomly divided into a training set(124 cases)and a test set(53 cases)at a ratio of 7:3.The radiomics features of lesions on ultrasound images were extracted.The intraclass correlation coefficient was used to remove features with poor consistency.Pearson was used to remove redundant features,and the minimum absolute shrinkage and selection algorithm was used to further reduce the dimension of the retained features,and finally the optimal features related to the expression level of PD-L1 were screened out.Logistic regression,support vector machine and extreme gradient boosting decision tree(XGBoost)were used to establish radiomics models,respectively;univariate and multivariate logistic regression were used to screen clinical risk factors and construct clinical models;a nomogram model was constructed based on the radiomics score and clinical risk factors.The diagnostic performance of different models was evaluated by ROC curve,and the AUC,sensitivity,specificity and accuracy were calculated.Results:Low degree of differentiation and high expression of Ki-67 were clinical risk factors for preoperative prediction of PD-L1 expression level.In both the training and validation sets,the XGBoost model showed the highest predictive performance,with AUCs of 0.833 and 0.776,respectively.A nomogram model was further constructed based on the degree of differentiation,Ki-67 expression and XGBoost radiomics score.The results showed that its AUC,sensitivity,specificity and accuracy in the training set were 0.902,80.39%,91.78%and 86.73%,respectively,and its AUC,sensitivity,specificity and those in the test set were 0.835,86.36%,77.42%and 80.90%,respectively.Conclusion:The ultrasound radiomics model established by XGBoost has a high predictive value for the expression level of PD-L1 in breast cancer.The nomogram established by the combination of clinical risk factors can further improve the diagnostic performance,which is expected to assist in clinical decision-making.
作者
周柳荫
周毅
潘颖
蔡仕彬
吴爱芬
卢伟业
陈述政
ZHOU Liuyin;ZHOU Yi;PAN Ying;CAI Shibin;WU Aifen;LU Weiye;CHEN Shuzheng(Department of operating Room,the Fifth Affiliated Hospital of Wenzhou Medical University,Lishui Central Hospital,Lishui 323000,China;Department of Breast Surgery,the Fifth Affiliated Hospital of Wenzhou Medical University,Lishui Central Hospital,Lishui 323000,China;Department of Ultrasound,the Fifth Affiliated Hospital of Wenzhou Medical University,Lishui Central Hospital,Lishui 323000,China)
出处
《温州医科大学学报》
CAS
2024年第11期898-905,共8页
Journal of Wenzhou Medical University
基金
国家卫生健康委人才交流服务中心科研项目(RCLX2315134)。
关键词
影像组学
乳腺癌
超声
程序性死亡配体1
radiomics
breast cancer
ultrasound
programmed death-ligand 1
作者简介
第一作者:周柳荫,主管护师,Emai1:236298862@q.com。;通信作者:陈述政,主任医师,Email:dr.susan@163.com。
