Mammography-based deep learning radiomics for predicting HER2 expression in breast cancer: a dual-center study with focus on HER2-low status

Background Accurate preoperative assessment of HER2 status is critical for guiding targeted therapy, but conventional biopsy-based methods are invasive and limited by tumor heterogeneity. Mammography is a widely accessible and cost-effective imaging modality routinely performed in the preoperative setting. This study aimed to develop and validate a mammography-based deep learning radiomics model for noninvasive prediction of HER2 expression status in breast cancer. Methods A total of 546 breast cancer patients from two centers were retrospectively enrolled: 420 from Center 1 (294 in the training cohort and 126 in the internal validation cohort) and 126 from Center 2 as the external validation cohort. Two classification tasks were defined: Task 1 distinguished HER2-positive from HER2-negative breast cancer, and Task 2 distinguished HER2-low from HER2-zero breast cancer. Radiomic features were extracted using the Visformer architecture. A weighted voting ensemble method was then employed to construct the prediction model, with weights optimized solely within the training cohort using cross-validated AUC. Model performance for both tasks was evaluated using receiver operating characteristic (ROC) curves. Results For Task 1 (HER2-positive vs. HER2-negative), the radiomics model achieved AUCs of 0.836, 0.763, and 0.714 in the training, internal validation, and external validation cohorts, respectively. For Task 2 (HER2-low vs. HER2-zero), the corresponding AUCs were 0.890, 0.785, and 0.713. Clinicopathological analysis revealed that histological grade (OR = 2.28), Ki-67 index (OR = 4.71), and BI-RADS category (OR = 2.49) were independent predictors of HER2-positive status, while PR status (OR = 4.48), BI-RADS category (OR = 1.75), and breast density (OR = 0.32) were independent predictors of HER2-low status (all P < 0.05). Conclusion The mammography-based deep learning radiomics model demonstrated meaningful discriminative ability for predicting HER2 expression status in breast cancer and may serve as a potential noninvasive adjunct for preoperative HER2 assessment, warranting further prospective validation. Clinical trial number Not applicable.