Low-magnification deep learning model for rapid HER2 status prediction from H&E whole-slide images

Background: HER2 (human epidermal growth factor receptor 2) overexpression is a pivotal biomarker for breast cancer prognosis and targeted therapy selection. Conventional assessment requires immunohistochemistry (IHC) and/or in situ hybridization (ISH), which are costly, time-consuming, and constrained by limited tissue and resource availability. In contrast, hematoxylin-and-eosin (H&E) slides are routinely acquired for diagnosis. Leveraging deep learning to infer HER2 status directly from H&E whole-slide images (WSIs) could substantially streamline the diagnostic workflow and reduce cost. However, existing deep learning models typically operate at high magnifications, resulting in slow slide-level processing and high computational costs, which hinder scalability and limit their integration into real-time clinical workflows. Methods: To predict HER2 overexpression directly from routine H&E whole-slide images (WSIs), we developed a streamlined deep-learning model tailored for low-magnification pathology images. The approach extracts meaningful histologic features from each slide and integrates them at the whole-slide level to generate a binary prediction of HER2 status (positive vs negative). Model development and validation were performed using the TCGA-BRCA dataset, applying a five-fold cross-validation strategy to ensure robustness and generalizability. In each fold, 467 WSIs were used for training and 145 for testing. HER2 status in TCGA was determined primarily by IHC, supplemented with ISH results. Results: Our model achieved an AUC of 0.7280.029 and an F1-score of 0.6530.054. In comparison, the state-of-the-art deep learning model UNI2 achieved an AUC of 0.7150.010 and F1-score of 0.6270.036. Despite comparable accuracy, our model demonstrated markedly higher efficiency, processing 8.8 whole-slide images per minute-approximately 30 faster than UNI2-while requiring significantly less computational and storage resources. Conclusions: This study highlights the potential of our deep learning model to predict HER2 status from low-magnification H&E WSIs. Our model's efficiency and scalability highlight its potential for integration into digital pathology workflows, enabling near real-time molecular screening without the need for additional staining or cloud-based computation. This level of efficiency further strengthens its suitability for real-world deployment, particularly in settings with high volume or limited computational resources. These findings support the feasibility of deep learning-driven virtual biomarker prediction as a practical step toward accessible, AI-assisted precision oncology.