Hybrid Deep Learning Framework for Multi-Class Breast Cancer Scoring Using Grad-CAM++.

This study presents a novel hybrid deep learning model that combines the ResNet-18 and EfficientNet-B0 architectures. It focuses on automating human epidermal rowth factor receptor 2 (HER2) immunohistochemistry (IHC) scoring in breast cancer. The framework tackles important challenges in classifying subtle HER2 expression patterns (0, 1+, 2+, 3+), by integrating ResNet-18’s spatial feature extraction with EfficientNet-B0’s multi-scale texture analysis. The model achieves 94% classification accuracy after being trained on an annotated patch-level images dataset known as HER2-IHC-40x. It shows a strong ability to distinguish borderline cases (1+/2+) with a macro F1-score of 0.93, while also maintaining efficiency by using frozen pretrained backbones. Quantitative evaluation confirms its reliability with ROC-AUC scores of 0.994 and high precisionrecall metrics. Additionally, Grad-CAM++ was mployed to visualize class-specific discriminative regions and enhance model interpretability. These results are validated through ablation studies against six baseline architectures. This work improves standardized HER2 scoring by connecting morphological and textural feature learning, which can help minimize differences in interpretation among clinical practitioners.

Recommended citation: Nabi, M. S., Fauzi, M. F. A., Karim, H. B. A. (2025). Hybrid Deep Learning Framework for Multi-Class Breast Cancer Scoring Using Grad-CAM++. In Proceedings of the International Conference on Information Technology (InCIT 2025). IEEE (Scopus Indexed)