Contrastive virtual staining enhances deep learning-based PDAC subtyping from H&E-stained tissue cores

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in oncology

Background:

• Pancreatic ductal adenocarcinoma (PDAC) subtyping traditionally uses labor-intensive immunohistochemistry (IHC).
• Manual IHC staining introduces variability and is time-consuming.
• Virtual staining generates synthetic IHC images from H&E slides, but often lacks diagnostic feature assessment.

Purpose of the Study:

• To develop and validate a novel virtual staining method for improved PDAC subtyping.
• To enhance diagnostic accuracy using AI-generated synthetic IHC images from H&E slides.
• To assess the clinical utility of contrastive virtual staining in PDAC diagnostics.

Main Methods:

• Implemented a novel cycleGAN framework with a contrastive-inspired approach.
• Trained the model on semipaired datasets from consecutive tissue sections.
• Generated synthetic IHC images from standard H&E slides for PDAC subtyping.

Main Results:

• Significantly improved PDAC subtyping accuracy for KRT81 (F1-score from 0.66 to 0.77) and HNF1A (F1-score from 0.61 to 0.73).
• Outperformed baseline CycleGAN models in generating diagnostically relevant synthetic IHC images.
• Demonstrated enhanced classification performance compared to direct H&E image analysis.

Conclusions:

• Contrastive virtual staining shows significant clinical potential for PDAC diagnostics.
• This AI-driven approach streamlines subtyping and improves diagnostic robustness.
• The method offers a promising alternative to manual IHC for PDAC classification.