Preoperative prediction of Ki-67 expression status in breast cancer based on dynamic contrast enhanced MRI radiomics combined with clinical imaging features model
10.3760/cma.j.cn112149-20211224-01137
- VernacularTitle:基于MRI多期增强影像组学联合临床影像特征模型术前预测乳腺癌Ki-67表达状态
- Author:
Shunan CHE
1
;
Mei XUE
;
Jing LI
;
Yuan TIAN
;
Jiesi HU
;
Sicong WANG
;
Xinming ZHAO
;
Chunwu ZHOU
Author Information
1. 国家癌症中心 国家肿瘤临床医学研究中心 中国医学科学院北京协和医学院肿瘤医院影像诊断科，北京100021
- Keywords:
Breast neoplasms;
Magnetic resonance imaging;
Ki-67;
Radiomics;
Prediction model
- From:
Chinese Journal of Radiology
2022;56(9):967-975
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the value of preoperative prediction of Ki-67 expression status in breast cancer based on multi-phase enhanced MRI combined with clinical imaging characteristics prediction model.Methods:This study was retrospective. A total of 213 breast cancer patients who underwent surgical treatment at Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College between June 2016 and May 2017 were enrolled. All patients were female, aged 24-78 (51±10) years, and underwent routine breast MRI within 2 weeks prior to surgery. According to the different Ki-67 expression of postoperative pathological results, patients were divided into high expression group (Ki-67≥20%, 153 cases) and low expression group (Ki-67<20%, 60 cases). The radiomic features of breast cancer lesions were extracted from phase 2 (CE-2) and phase 7 (CE-7) images of dynamic contrast enhanced (DCE)-MRI, and all cases were divided into training and test sets according to the ratio of 7∶3. The radiomic features were first selected using ANOVA and Wilcoxon signed-rank test, followed by the least absolute shrinkage and selection operator method regression model. The same method of parameters selection was applied to clinical information and conventional imaging features [including gland classification, degree of background parenchymal enhancement, multifocal/multicentric, lesion location, lesion morphology, lesion long diameter, lesion short diameter, T 2WI signal characteristics, diffusion-weighted imaging (DWI) signal characteristics, apparent diffusion coefficient (ADC) values, time-signal intensity curve type, and axillary lymph nodes larger than 1 cm in short axis]. Support vector machine (SVM) was then used to construct prediction models for Ki-67 high and low expression states. The predictive performance of the models were evaluated using receiver operating characteristic (ROC) curves and area under cueve(AUC). Results:Totally 1 029 radiomic features were extracted from CE-2 and CE-7 images, respectively, and 9 and 7 best features were obtained after selection, respectively. And combining the two sets of features for a total of 16 features constituted the CE-2+CE-7 image best features. Five valuable parameters including lesion location, lesion short diameter, DWI signal characteristics, ADC values, and axillary lymph nodes larger than 1 cm in short axis, were selected from all clinical image features. The SVM prediction models obtained from the radiomic features of CE-2 and CE-7 images had a high AUC in predicting Ki-67 expression status (>0.70) in both the training set and the test set. The models were constructed by combining the CE-2, CE-7, and CE-2+CE-7 radiomic features with clinical imaging features, respectively, and the corresponding model performance in predicting Ki-67 expression status was improved compared with the models obtained by using the CE-2, CE-7, and CE-2+CE-7 radiomic features alone. The SVM prediction model obtained from CE-2+CE-7 radiomic features combined with clinical imaging features had the best prediction performance, with AUC of 0.895, accuracy of 84.6%, sensitivity of 87.9%, and specificity of 76.2% for predicting Ki-67 expression status in the training set and AUC of 0.822, accuracy of 70.3%, sensitivity of 76.1%, and specificity of 55.6% in test sets.Conclusion:The SVM prediction model based on DCE-MRI radiomic features can effectively predict Ki-67 expression status, and the combination of radiomic features and clinical imaging features can further improve the model prediction performance.
