Explainable deep learning models for HER2 IHC scoring in breast cancer diagnosis

Abstract

Accurate and interpretable HER2 IHC scoring is crucial for guiding breast cancer treatment. However, manual evaluation remains inconsistent and subjective. This study proposes a deep learning framework that integrates both a custom Convolutional Neural Network (CNN) and a fine-tuned DenseNet121 model for automated HER2 scoring using the HER-IHC-40x dataset. Preprocessing involves HSV-based patch filtering and expert validation to ensure data relevance. To improve transparency and address the black-box nature of AI models, we employed explainable AI (XAI) techniques. Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) provide visual explanations at the pixel and region levels. These techniques enhance interpretability, ensuring clinical confidence by clearly visualizing and attributing model decisions, particularly in borderline HER2 cases (Class 1+ and 2+), where manual misinterpretations are common. The experimental results show that both CNN and DenseNet121 achieved 93% accuracy with excellent class-wise consistency. CNN, in particular, demonstrated higher prediction confidence and lower training loss, indicating superior calibration. The integration of explainability modules ensures improved clinical transparency and improves trust in AI-driven decision-making. Comparison with the existing literature confirms the strength of the proposed method in predictive capacity and interpretability, contributing to a robust AI-assisted breast cancer diagnosis.

Topics & Concepts