Virtual Histological Staining as a Tool for Extending Renal Segmentation Across Stains.

Area of Science:

• Digital Pathology
• Computational Histopathology
• Artificial Intelligence in Medicine

Background:

• Clinical use of diverse histological stains complicates deep learning analysis of whole-slide kidney images.
• Hematoxylin and eosin (H&E) staining is common, but other stains like periodic acid-Schiff (PAS) are crucial for specific diagnoses.

Purpose of the Study:

• To develop and validate an in silico method for virtually restaining H&E-stained kidney tissue to PAS.
• To enable stain-specific deep learning tools for renal histopathology using a single H&E stain.

Main Methods:

• Cycle-consistent generative adversarial neural networks were trained on a large UK renal dataset.
• A virtual staining model inferred PAS from H&E, and its realism was assessed by pathologists.
• A glomerular segmentation model was trained on multiple datasets and applied to virtually stained images.

Main Results:

• Pathologists could not reliably distinguish virtual PAS staining from real PAS staining (52.5% and 75.8% accuracy).
• The virtual staining approach successfully inferred PAS staining on unseen H&E images.
• The PAS-specific glomerular segmentation model performed effectively on virtually stained images.

Conclusions:

• Virtual staining offers a viable solution to overcome limitations of stain-specific deep learning in renal histopathology.
• This technique enhances the utility of digital pathology by enabling multi-stain analysis from a single H&E slide.
• Further research can expand this method to incorporate additional histological stains.