Deep learning yields macro-AUC 0.971 for automated HER2 scoring on breast IHC images

In a dual-center retrospective study of 118 breast cancer patients, a patch-level AlexNet model achieved macro-AUC 0.971 for three-class HER2 IHC score prediction, with most errors between adjacent grades.

• Retrospective dual-center study using 135 IHC whole-slide images; six pretrained models compared; AlexNet performed best (macro-AUC 0.971; per-class AUCs 0.980 for 1+, 0.955 for 2+, 0.979 for 3+).
• Grad-CAM visualizations aligned with tumor-rich regions, supporting prediction interpretability.
• Limitations include patch-level (not slide-level) validation, lack of external prospective testing, and baseline group differences (clinical symptoms, lymph node status); generalizability remains unproven.

RadPigeon summaries are original and for information only. They are not clinical advice.