Tumor-Aware Unsupervised Digital HER2 Immunohistochemistry via Dual Contrastive Learning

Breast cancer is the most common cancer and the second leading cause of cancer-related death among women worldwide. For accurate diagnosis, pathologists use immunohistochemical (IHC) staining for biomarkers such as human epidermal growth factor receptor 2 (HER2) as an auxiliary test, in addition to hematoxylin and eosin (H&E) staining for evaluating tissue morphology. Traditional IHC staining is limited by high cost, time, and labor, with a shortage of pathologists to meet demand. In this work, we propose a digital HER2 IHC staining algorithm based on tumor masks. The proposed model builds on the Dual Contrastive Learning Generative Adversarial Network (DCLGAN), adding tumor mask loss, structural similarity index measure (SSIM) loss, and a modified identity loss that differentiates between active and inactive regions. Performance was quantitatively assessed by comparing HER2 scores using the Fréchet Inception Distance (FID) and a grade classification model. Compared to DCLGAN, the proposed model achieved a 27.2% decrease in FID and a 2.57% increase in soft accuracy. Visual evaluation further showed that it prevents yellow discoloration and suppresses nonspecific region staining seen in unsupervised learning. These results indicate that effective learning is achievable even with limited and unaligned data, thereby accelerating digital pathology workflows.