[A multi-scale attention and federated learning-based method for human epidermal growth factor receptor 2...

Immunohistochemistry (IHC) images for human epidermal growth factor receptor 2 (HER2) in breast cancer face challenges in cross-institutional sharing and suffer from insufficient representation of boundary categories. To address these issues, this paper proposes an automatic grading method that integrates multi-scale attention mechanisms with federated learning. Specifically, we construct a multi-scale attention residual network 50 (MSA-ResNet50) model, which enhances the model's discriminative capability for membrane staining intensity, staining continuity, and tissue morphological structures by fusing channel attention, spatial attention, and global attention branches in parallel. For the federated learning strategy, we incorporate client-side validation set accuracy and macro-averaged F1 score to perform quality-weighted aggregation based on the federated averaging (FedAvg) algorithm. Experimental results demonstrate that the proposed method achieves an accuracy of 87.78%, outperforming comparative methods. Furthermore, visualization results indicate that our model effectively strengthens feature responses in membrane staining regions that are critical for HER2 grading discrimination, thereby improving classification accuracy for the HER2 2+ category, which is clinically challenging to interpret. This work provides a technical reference for subsequent validation on real-world multi-center clinical data.