What it does

H&E staining is cheap and ubiquitous; IHC biomarker staining is expensive and slow. The premise of "virtual staining" is to use a generative model to predict the IHC stain directly from an H&E slide. Existing approaches need the two slides to be co-registered at sub-cell resolution — a brittle preprocessing step that fails on real-world hospital data.

This project removes that constraint. The model is trained on unaligned slide pairs and learns the cell-level mapping without explicit registration.

Results

The system outperformed four published baselines on cell-level pathology metrics for biomarker quantification — the metrics that actually matter clinically, not just pixel-level similarity.

Stack

PyTorch, custom multi-resolution sampling, OpenSlide for whole-slide IO, evaluation harness in scikit-image and a custom cell-detection wrapper.