Deep learning for HER2 IHC scoring with quantitative explainability across magnifications

Accurate assessment of human epidermal growth factor receptor 2 (HER2) status is critical for guiding targeted therapy in breast cancer; however, manual immunohistochemistry (IHC) scoring is subject to inter-observer variability, particularly in borderline cases. While deep learning approaches have shown strong performance in automated HER2 classification, their clinical applicability is often limited by insufficient interpretability and sensitivity to variations in imaging magnification. In this study, we present a deep learning-based framework for HER2 IHC scoring that integrates a ResNet50 backbone with a hybrid layer unfreezing strategy to balance feature adaptation and computational efficiency. A single model is trained across datasets with different magnification levels (10 \(×\) and 40 \(×\) ), enabling consistent performance without requiring separate models for each resolution. To enhance interpretability, Score-CAM is incorporated alongside quantitatively defined metrics, including Membrane Activation Precision (MAP), Explanation Consistency (EC), and Relevance Coverage (RC), which evaluate the alignment and stability of model attention with respect to pathologist-annotated membrane regions. These annotations are used exclusively for evaluation to ensure objective validation of explanation quality. The proposed framework is evaluated on three publicly available HER2 IHC datasets, achieving up to 96% accuracy on patch-based data and maintaining strong performance on lower-magnification images. Experimental results demonstrate that the model provides consistent performance across the evaluated magnification settings while producing clinically meaningful and quantitatively validated explanations. Overall, this study highlights the potential of combining deep learning with structured explainability evaluation for HER2 scoring, while emphasizing the need for further validation across diverse clinical datasets before real-world deployment.