Deep generative virtual staining for breast cancer: from H&E histology to HER2 immunohistochemistry

Abstract Immunohistochemistry (IHC) remains the clinical gold standard for molecular profiling in breast cancer; however, traditional workflows are hindered by high costs, reagent dependency, and tissue exhaustion. Virtual staining has emerged as a computational alternative; however, prevailing virtual staining models often fail to preserve high-frequency structural details critical for perceptual and histological fidelity. We present a novel chemically-informed, two-stage conditional Generative Adversarial Network (cGAN) framework designed for high-fidelity H&E-to-IHC translation. Our approach uniquely decouples chemical stain extraction from high-resolution textural refinement through a specialized StainMapper (G₁) and a score-conditioned RefineGenerator (G₂). By integrating multi-scale hierarchical score conditioning with a hybrid loss function including adversarial, perceptual, Sobel edge, and stain-consistency terms, the framework enforces both global distributional realism and local morphological precision. Evaluated on the BCI and HER2match benchmarks, our model achieves state-of-the-art performance, reaching a Fréchet Inception Distance (FID) of 59.67 , PSNR of 22.50 dB , SSIM of 0.4773 , LPIPS of 0.5279 and KID of 0.0264 ± 0.0086 on BCI; PSNR of 20.66 dB , LPIPS of 0.4486 , FID of 20.42 , and KID of 0.0124 ± 0.0092 on HER2match, outperforming all baselines on distributional and perceptual metrics. These results demonstrate that incorporating histopathological priors and structural attention mechanisms improves the histological and perceptual fidelity of virtual IHC synthesis, providing a methodological foundation for potential secondary applications such as stain standardization, dataset augmentation, and pathology education in settings where HER2 score is already known.