A Hybrid Feature Fusion Network for Predicting HER2 Status on H &E-Stained Histopathology Images

Human epidermal growth factor receptor-2 (HER2) is critical in regulating cellular growth, development, and differentiation processes in normal cells. Breast cancer is the most common and lethal cancer among women worldwide, and about 25% of breast cancer patients have HER2 overexpression/amplification. At present, the commonly evaluating HER2 status methods are tissue-consuming and prone to analysis or interpretation errors. Therefore, this study proposes a hybrid feature fusion method (HM-HER2) based on deep learning to predict the HER2 status of breast cancer patients by using the microscopic morphology of tissues on hematoxylin and eosin (H &E) slides. The HM-HER2 model includes a Global representation module (GRM) and a Local representation module (LRM) for global and local features extraction, respectively, and then fuses global and local features for final classification. The experimental results show the effectiveness and potential of the proposed HM-HER2 model in the field of H &E-stained whole slide images (WSIs) classification of breast cancer.