A deep learning model utilizing H&E-stained images for predicting HER2-low expression in breast cancer

Abstract

The development and clinical application of novel HER2-targeted antibody-drug conjugates (ADCs) have significantly improved outcomes for breast cancer patients with HER2-low expression. The efficacy of such therapies critically depends on accurate assessment of HER2-low status. However, immunohistochemical (IHC) interpretation of HER2-low breast cancer faces multiple challenges due to tumor heterogeneity and interobserver variability among pathologists. This study aimed to develop a deep learning-based framework for analyzing hematoxylin and eosin (H&E) stained whole-slide images (WSIs) of breast cancer to achieve precise prediction of HER2-low status while providing interpretable evidence. We retrospectively collected 776 cases of invasive breast carcinoma diagnosed at the Affiliated Hospital of Zunyi Medical University between January 2019 and April 2023 to construct a HER2-low expression dataset. Leveraging an ImageNet-pretrained ResNet50 model for feature extraction and a CLAM (Clustering-constrained Attention Multiple Instance Learning) model with 10-fold cross-validation, our framework demonstrated robust performance on both validation and test sets. Critical HER2-low predictive regions were visualized using attention heatmaps to enhance model interpretability. The mean AUC of the model was 0.613 ± 0.118 on the validation set, and 0.608 ± 0.104 on the test set. The attention heatmap visualization provided biologically plausible explanations for model predictions, offering reliable decision-support tools for pathologists.