Explainable AI for Breast Cancer Diagnosis Using EfficientNetB3 with Attention Mechanism
Md Serajun Nabi, Mohammad Faizal Ahmad Fauzi, Hezerul Bin Abdul Karim, Ahmad Shahrafidz Khalid, Tong Boon Tang, Normy Norfiza Abdul Razak · 2025
Accurate classification of HER2 immunohistochemistry (IHC) scores is essential for determining effective breast cancer treatment, yet it remains challenging due to subjective manual interpretation, especially for borderline scores ($\mathbf{1}+$and$\mathbf{2}+$). This study proposes an interpretable deep learning framework that combines EfficientNetB3 with a Convolutional Block Attention Module (CBAM) to strengthen feature extraction and attention to regions of interest that are diagnostically significant. To facilitate clinical trust, explainable AI (XAI) is performed using Gradient-weighted Class Activation Mapping (Grad-CAM). Evaluated on a HER2-IHC-40xWSI dataset of 10,997 image patches distributed over four HER2 classes ($0,1+, 2+, 3+$), the proposed model achieved an overall accuracy of 96% and a macro-averaged F1-score of$\mathbf{9 3 \%}$, demonstrating strong performance, particularly in borderline cases. The system demonstrates promising performance within the dataset, particularly in borderline cases, suggesting potential for broader generalization. These results highlight the potential of applying attention mechanisms with explainable AI for stable and interpretable HER2 IHC scoring in digital pathology.