Detection of low expression of molecular subtype HER2 in breast cancer using ultrasound-based radiomics

Abstract

Purpose: HER2-low expression breast cancer (BC) accounts for nearly half of all breast cancers which may benefit from new antibody-drug conjugates targeted treatments. HER2-low BC is gradually being recognized as a distinct subtype. Therefore, we aimed to use ultrasound-based radiomics（USR）to develop an efficient evaluation approach of HER2-low status. Methods: 222 patients with a histologically diagnosis of BC were retrospectively analyzed and randomly divided into training and test cohort. Radiomics features were extracted from the preoperative ultrasound images, followed by Lasso regression for dimension reduction.Based on the selected features, the optimal machine learning classifier was selected to construct a USR model to predict HER2-low expression. Multivariable logistic regression was used to identify independent clinical risk factors.Finally, a clinical-USR model incorporating the radiomics features and the clinical risk factors was constructed. Model performance was assessed using receiver operating characteristic curve and decision curve analysis. Results: The USR model and clinical-USR model had good predictive ability in the training cohort ( AUC 0.91; 0.86 , respectively) and test cohort ( AUC 0.83; 0.78 , respectively).Both of them outperformed the clinical model (p < 0.05, DeLong test).Decision curve analysis confirmed that the model had clinical utility. Conclusion: The machine learning model based on ultrasound images had high prediction value for HER2-low BC.