Diffusion-based virtual multi-stain staining for pituitary adenoma histopathology

Immunohistochemistry (IHC) for lineage-defining transcription factors-such as Pit1, Tpit, and SF1-together with markers like Ki67 and ACTH, is central to classifying pituitary neuroendocrine tumors and guiding management. Yet preparing multiple IHC slides beyond hematoxylin-eosin (H&E) is costly and inconsistent across sites. We aim to produce clinically reliable virtual IHC stains from a single H&E image under unpaired supervision.

We develop a diffusion-based one-to-many virtual staining framework that transforms a single H&E image into five key IHC stains. Our approach fuses multi-scale structural features into a unified conditional representation, processed by a shared backbone with stain-specific decoders to enable mutually reinforcing learning across stains. The model incorporates perception-prioritized, timestep-aware diffusion loss and a stain-consistency term during training, combined with classifier-guided sampling during inference, to achieve exceptional fidelity in both cellular structure and chromogenic style. For clinical validation, we further introduced the nuclear positivity rate as a slide-level endpoint to quantitatively evaluate agreement between virtual and real IHC.

Compared to the best-performing baseline, our method achieves superior pixel and perceptual fidelity (PSNR ↑ +0.94 dB, SSIM ↑ +0.015; FID ↓-7.8%, LPIPS ↓-7.9%). For the clinical endpoint, averaged across all stains, it delivers substantial improvements in mean absolute error, correlation, and accuracy relative to the baseline average-reducing MAE by approximately 28.7%, increasing correlation by 0.07, and boosting accuracy by 0.05. Coupled with roughly 80% fewer parameters and comparable throughput, these outcomes position our approach as a unified, clinically focused diffusion model for virtual multi-stain synthesis.