2928 - Pretreatment Mammography Deep Learning to Predict Luminal and TNBC Phenotypes in Early Breast Cancer: A Timely Decision-Support Tool When CNB-IHC is Discordant or Unavailable for Retesting
Presenter(s)
M. Zhang1, X. Yu2, J. Wang2, Z. Lyu2, Z. Dong2, and W. Hu2; 1Fudan University Shanghai Cancer Center, Shanghai, China, 2Department of Radiation Oncology, Fudan University Shanghai Cancer Center; Department of Oncology, Shanghai Medical College, Fudan University; Shanghai Clinical Research Center for Radiation Oncology; Shanghai Key Laboratory of Radiation Oncology, Shanghai, China
Purpose/Objective(s): For breast cancer, several molecular subtypes corresponding to distinct responses to treatments and prognoses have been recognized. In current practice, immunohistochemistry (IHC) surrogates (ER/PR/HER2/Ki67) obtained from core needle biopsy (CNB) are widely used for preoperative subtyping, but are vulnerable to sampling error from tumor heterogeneity and technical factors (fixation, antibody choice, thresholds). Large paired CNB–surgical specimen series report discordance rates of 3.9% (ER), 4.8% (PR), and 1.2% (HER2), and subtype discordance can be amplified when biomarkers are combined. Retesting IHC on surgical specimens is time-consuming and not routinely performed; moreover, for neoadjuvant patients—especially those achieving pathologic complete response (pCR)—no residual tumor may remain for postoperative retesting, making accurate pretreatment stratification crucial. We developed a deep learning (DL) model using pretreatment standard-view mammography to output Luminal/non-Luminal and TNBC/non-TNBC probabilities to support timely decisions on whether additional IHC workup (e.g., deeper sections/restaining on original CNB blocks or selective retesting when feasible) is warranted.
Materials/Methods: We retrospectively collected 848 mammograms (CC and MLO; 424 breasts) from 424 patients with pathology-confirmed early-stage (I–III) breast cancer treated at our institution (Feb 2015–Dec 2021). Ground truth subtype labels were derived from definitive pathology/IHC. A ConvNeXt-Small backbone extracted view-specific features; breast-level probability was the mean of CC and MLO outputs. We initialized with pretrained weights from the RSNA Screening Mammography Breast Cancer Detection dataset and fine-tuned on our cohort. To mitigate overfitting and improve robustness, we additionally incorporated the public Chinese Mammography Database (CMMD; subtype-labeled mammograms) into the training pipeline. Performance was evaluated at breast level using four-fold cross-validation and an independent internal test cohort; primary metric was AUC.
Results: For Luminal vs non-Luminal classification, mean AUC was 0.818 across cross-validation validation folds. For TNBC vs non-TNBC, mean AUC was 0.812 in validation.
Conclusion: A pretreatment mammography DL model can noninvasively predict clinically actionable phenotypes (Luminal/non-Luminal and TNBC/non-TNBC) in early breast cancer prior to surgery. By providing an independent, rapid probabilistic signal, this approach may help flag cases at higher risk of CNB-IHC discordance/false negatives and support timely decisions on whether additional IHC workup (e.g., restaining/deeper sections on CNB blocks or selective retesting when feasible) is warranted—particularly relevant in neoadjuvant pathways where postoperative retesting may be impractical or impossible in pCR.