Generative adversarial network based digital stain conversion for generating RGB EVG stained image from hyperspectral...

Area of Science:

• Digital pathology
• Computational imaging
• Biomedical image analysis

Background:

• Elastic fiber quantification is crucial for disease diagnosis.
• Hematoxylin and eosin (H&E) staining is cost-effective but cannot differentiate elastic fibers.
• Conventional Verhoeff's van Gieson (EVG) staining is expensive and time-consuming.

Purpose of the Study:

• To develop a deep learning-based computerized method for generating RGB EVG stained images from hyperspectral H&E stained images.
• To reduce the time and cost associated with conventional EVG staining procedures.

Main Methods:

• Utilized hyperspectral H&E stained images and RGB EVG stained whole slide images of human pancreatic tissue.
• Employed a CycleGAN-based deep learning model for digital stain conversion between different modalities (hyperspectral and RGB) and channel dimensions.
• Introduced a set of three basis functions to calculate a loss component, preserving EVG image features within the reduced channel dimension of H&E images.

Main Results:

• A set of three basis functions, including linear discriminant function and transmittance spectra of eosin and hematoxylin, effectively retained elastic fiber properties for discrimination from collagen.
• The proposed training method required fewer paired training data to generate realistic EVG stained images with precise elastic fiber identification.
• The model achieved simultaneous hyperspectral to RGB and H&E to EVG image conversion, demonstrating effectiveness in generating realistic RGB EVG images.

Conclusions:

• The developed deep learning method successfully generates realistic RGB EVG stained images from hyperspectral H&E stained images.
• The intentionally designed set of three basis functions proved effective in retaining relevant information for accurate elastic fiber identification.
• This digital approach offers a significant advancement in cost-effective and time-efficient tissue analysis for pathological diagnosis.