Improving stain normalization for digital histological image analysis based on the cycle generative adversarial...

Area of Science:

• Digital pathology
• Computational imaging
• Machine learning in medicine

Background:

• Stain color variations in histopathological images present a significant challenge for automated analysis.
• Inconsistent staining environments and scanning devices lead to unreliable deep learning model performance.
• Preserving structural integrity during stain normalization is critical for accurate interpretation.

Purpose of the Study:

• To develop a robust stain normalization framework for digital histopathology images.
• To enable stable color-domain conversion across diverse staining conditions.
• To maintain structural information during the normalization process for deep learning applications.

Main Methods:

• A generative adversarial network (GAN)-based framework, I-GAN, integrating StainGAN and Stain-to-Stain Translation (STST).
• Incorporation of identity loss within an RGB-grayscale training strategy.
• Application of RGB images during testing to preserve original stain characteristics.

Main Results:

• I-GAN achieved high performance on the MITOS-ATYPIA 14 dataset with SSIM of 0.980, PSNR of 29.579, and DeltaE-ITP of 46.284.
• Demonstrated superior structural preservation and color fidelity.
• Achieved excellent results on downstream classification tasks, including 0.964 average precision on Camelyon17 and 0.87 accuracy on the ICIAR2018 BACH dataset.

Conclusions:

• The I-GAN framework effectively addresses stain normalization challenges in hematoxylin and eosin-stained images.
• It ensures structural integrity and accurate color-domain conversion for digital histopathology.
• The approach shows robustness and practical applicability for medical image analysis.