Truncated Normal Mixture Prior Based Deep Latent Model for Color Normalization of Histology Images.

Area of Science:

• Digital Pathology
• Computational Imaging
• Medical Image Analysis

Background:

• Color variation in Hematoxylin and Eosin (H&E) stained histological images poses a significant challenge for computer-aided diagnosis.
• Inconsistent staining can lead to misinterpretation and affect diagnostic accuracy.

Purpose of the Study:

• To develop a novel deep generative model for reducing color variation in H&E stained histological images.
• To improve the reliability and accuracy of computer-aided diagnosis by addressing color disagreement.

Main Methods:

• A deep generative model employing a color appearance encoder and a stain density encoder to extract disentangled information.
• Utilizing a generative and reconstructive module with objective functions to capture latent color and stain information.
• A discriminator model that evaluates image samples and joint distributions, incorporating a mixture of truncated normal distributions to handle overlapping stain information.

Main Results:

• The proposed model demonstrated superior performance in stain separation (91.67% of cases) and color normalization (69.05% of cases) compared to state-of-the-art methods.
• Validation was performed on multiple public datasets of H&E stained histological images.

Conclusions:

• The novel deep generative model effectively reduces color variation in H&E stained images.
• This approach significantly enhances stain separation and color normalization, offering a promising solution for improving digital pathology diagnostics.