Abstract Accurate and rapid determination of tumor histopathological features and molecular subtypes is critical for breast cancer prognosis and treatment strategies. This study evaluates the feasibility of a “digital biopsy” approach that uses machine learning models to predict hormone receptor status from ultrasound radiomic features non-invasively. A retrospective analysis was conducted on 353 breast tumors from 311 patients who were evaluated between February 2019 and December 2024. Radiomic features were extracted from two-dimensional segmentations of pre-biopsy ultrasound images. The dataset was divided into training (80%) and testing (20%) sets. Of the initial 245 radiomic features, the top 10 features were selected for model development. Two primary tree-based ensemble algorithms, XGBoost and ExtraTrees, were utilized to predict the estrogen receptor (ER), progesterone receptor (PR), C-erbB2, Ki-67, histological grade, and molecular subtype. Model interpretability and feature importance were analyzed using SHap (Shapley Additive Explanations). The majority of tumors were hormone receptor-positive (84.1% ER-positive and 74.8% PR-positive), with Luminal B being the most prevalent molecular subtype (49.6%). Predicting ER status demonstrated the most consistent performance, achieving accuracy rates between 0.79 and 0.83. The ExtraTrees model exhibited the greatest discriminatory ability for ER, with an area under the curve (AUC) of 0.70. Prediction of PR status achieved an accuracy of 0.73, while the XGBoost model yielded an AUC of 0.61. Furthermore, the predictive performance of C-erbB2, Ki-67, histological grade, and overall molecular subtype was promising across a range of scenarios and model modifications. Combining ultrasound-based radiomic features with machine learning algorithms can effectively identify complex image patterns associated with hormone receptor status. This approach offers a promising, non-invasive alternative to traditional tissue sampling by enabling a “digital biopsy.” Integrating these models into clinical workflows could reduce the diagnostic burden on patients and facilitate personalized treatment planning. However, further validation in larger, prospective cohorts is necessary.
Kuzan et al. (Tue,) studied this question.