国际医学放射学杂志2024，Vol.47Issue(3)：288-293,6.DOI:10.19300/j.2024.L21305
基于超声的深度学习模型对乳腺癌HER2表达的预测价值
The predictive value of an ultrasound-based deep learning model for HER2 expression of breast cancer
摘要
Abstract
Objective To explore the potential value of an ultrasound-based deep learning model for preoperatively predicting HER2-Zero and HER2-Low expression in breast cancer.Methods Clinical data and preoperative ultrasound images of 243 patients pathologically diagnosed with breast cancer were retrospectively collected from 2 hospitals.The patients were divided into a training set(124 cases)and a validation set(119 cases)based on their source hospitals.Lesions in the ultrasound images were manually delineated using ImageJ software.Three pre-trained basic deep learning methods(VGG16,ResNet50,DenseNet121)were used to construct the ultrasound-based deep learning model in the training set,and the most efficient model was selected.Multivariate logistic regression analysis was used in the training set to identify clinical independent predictors related to HER2 status and construct a clinical model.These clinical independent predictors were combined with the optimal deep learning model to create a comprehensive model.The area under the receiver operating characteristic curve(AUC)was used to the predictive performance of each model,while calibration curve and the decision curve were used to evaluate the fitting effect and clinical net benefit of the comprehensive model.Results Multivariate logistic regression analysis showed that Ki-67 expression status was an independent clinical predictor of HER2 status in breast cancer.Among the three deep learning models,the DenseNet121 model performed the best.Therefore,Ki-67 and DenseNet121 were combined to construct the comprehensive model.In both the training and validation sets,the predictive performance of the comprehensive model(AUCs values of 0.865 and 0.848,respectively)was higher than that of the clinical model and the three deep learning models.Calibration curves showed that the comprehensive model had good calibration fitting effects in both data sets(Hosmer-Lemeshow test:P=0.267,0.398),and decision curve indicated a wide range of clinical net benefit for the comprehensive model.Conclusions The ultrasound-based deep learning model has significant value in preoperatively predicting HER2-Zero and HER2-Low expression in breast cancer patients,providing a basis for clinical treatment decision-making.关键词
超声/深度学习/乳腺癌/人表皮生长因子受体2Key words
Ultrasound/Deep learning/Breast cancer/Human epidermal growth factor receptor 2分类
医药卫生引用本文复制引用
余娜芊,刘宇,姚梦霞,黄春旺,吴磊,王瑛..基于超声的深度学习模型对乳腺癌HER2表达的预测价值[J].国际医学放射学杂志,2024,47(3):288-293,6.基金项目
国家自然科学基金面上项目(82272088) （82272088）
国家自然科学基金青年基金(82102019) （82102019）
