Unpaired virtual histological staining using prior-guided generative adversarial networks.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence

Background:

• Fibrosis is a critical indicator of chronic liver disease progression.
• Histopathology, particularly Masson trichrome staining, is the gold standard for fibrosis assessment.
• Current methods face challenges with cost, time, and staining variability.

Purpose of the Study:

• To develop an AI-driven method for virtual Masson trichrome staining from H&E images.
• To overcome limitations of existing image translation techniques for fibrosis evaluation.
• To provide a cost-effective and efficient alternative for fibrosis staging.

Main Methods:

• A prior-guided generative adversarial network (GAN) was employed for unpaired image-to-image translation.
• The GAN utilized prior knowledge for improved encoder-decoder constraints on a small dataset.
• A finetuning strategy was implemented for color standardization and computational efficiency.

Main Results:

• The proposed method accurately generates virtual Masson trichrome stained images from H&E images.
• Results showed a high correlation between real and virtual Masson trichrome staging (ρ=0.82).
• The method demonstrated superior performance compared to existing approaches.

Conclusions:

• The AI-based virtual staining method offers a reliable and efficient tool for fibrosis evaluation in chronic liver disease.
• This technique can standardize staining and reduce diagnostic burdens.
• The approach holds promise for clinical application in liver disease assessment.