Author(s) :

Aneesh Dantuluri.

Volume/Issue :

Abstract :

HER2-status is a vital biomarker for breast cancer diagnosis and treatment, typically assessed using immunohistochemistry (IHC), a technique that is expensive and demanding of laboratory experience. Hematoxylin and eosin (H&E) staining, in contrast, is widely available and inexpensive, motivating approaches that can computationally translate H&E images into IHC. While previous work has explored translating H&E stains of breast tissue into IHC using purely unsupervised methods, this study introduces a hybrid CycleGAN framework that combines unsupervised cycle-consistency with supervised paired reconstruction objectives. By leveraging the paired structure of the BCI dataset, this approach significantly improves quantitative metrics (PSNR: 16.203 → 17.807 (Adam); SSIM: 0.373 → 0.4061 (AdamW) and visual fidelity compared to unsupervised-only baselines, narrowing the performance gap with supervised-only architectures while maintaining CycleGAN’s flexibility. These f indings show that incorporating limited supervision into cycle-consistent adversarial training enhances H&E-to-IHC translation quality, offering a more affordable and accessible pathway to HER2 screening.