Prediction of Expression of Ki-67 Status in Breast Cancer via Deep Learning-Based Radiomics Model
10.3969/j.issn.1005-5185.2025.10.005
- VernacularTitle:基于深度学习的影像组学模型预测乳腺癌表达Ki-67状态
- Author:
Hanmin XIE
1
;
Jialing CHENG
1
;
Yuelong LI
1
;
Chengwei LI
1
;
Chaoxiang YANG
1
;
Ruoxian ZHANG
1
Author Information
1. 广东省妇幼保健院放射科,广东 广州 511400
- Publication Type:Journal Article
- Keywords:
Breast neoplasms;
Magnetic resonance imaging;
Ki-67 antigen;
Radiomics;
Deep learning;
Machine learning;
Forecasting
- From:
Chinese Journal of Medical Imaging
2025;33(10):1049-1055
- CountryChina
- Language:Chinese
-
Abstract:
Purpose To analyze the value of a deep learning(DL)radiomics model based on dynamic contrast-enhanced MRI images in predicting the expression of Ki-67 status in breast cancer.Materials and Methods A retrospective analysis of 152 breast cancer patients confirmed by pathological results at Guangdong Women and Children Hospital,MRI images and clinical pathological data were reviewed,and based on postoperative immunohistochemistry results,the images of the high and low expression groups of Ki-67 were randomly sampled in a ratio of 8∶2 to form a training set of 122 cases and a validation set of 30 cases.Single-factor and multi-factor Logistic regression analyses of clinical data were performed to select independent predictors of breast cancer expressing Ki-67 status.The ResNet-18 model was used as the basic model for DL feature extraction.Hand-crafted radiomic features and DL features were extracted.Eight machine learning models were constructed based on clinical features,hand-crafted radiomic features,DL features,and their combinations.The area under the receiver operating characteristic curve was used to evaluate the predictive performance of the models,and the best model was determined as the output model.Results The progesterone receptor status(OR=0.764,P=0.040)and human epidermal growth factor receptor-2 status(OR=1.187,P=0.046)were independent clinical predictors of breast cancer expressing Ki-67 status.The combined feature models demonstrated superior performance over the individual feature models,and the support vector machine algorithm had the highest prediction performance in the validation set,with an area under the curve of 0.847.Conclusion The DL radiomics model based on dynamic contrast-enhanced MRI images can effectively predict the expression of Ki-67 status in breast cancer.The support vector machine algorithm combined with feature model is the best,which can help the clinical diagnosis and treatment of breast cancer and prognosis evaluation.
