Multi-domain stain normalization for digital pathology: A cycle-consistent adversarial network for whole slide images

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Computer-Aided Diagnosis

Background:

• Histologic staining variations across institutions challenge computer-aided diagnosis (CAD) reliability.
• Disparities in pathological whole slide images (WSIs) reduce algorithm performance and hinder cancer diagnosis applications.
• Staining differences introduce biases, negatively impacting model performance during domain shifts.

Purpose of the Study:

• To develop a multi-domain stain normalization technique for pathological WSIs.
• To address the challenge of varying histologic staining in medical centers.
• To improve the reliability and applicability of CAD systems.

Main Methods:

• Proposed MultiStain-CycleGAN, a novel multi-domain stain normalization approach based on CycleGAN.
• Modified CycleGAN to normalize images from different origins without retraining or separate models.
• Conducted extensive evaluations using domain classification, tumor classification performance, Structural Similarity Index (SSIM), and Fréchet Inception Distance (FID).

Main Results:

• MultiStain-CycleGAN demonstrated multi-domain capability, achieving high image quality.
• The method effectively fooled a domain classifier, indicating successful normalization.
• Maintained high tumor classifier performance while reducing domain shift.
• Achieved superior performance compared to other multi-domain capable methods.

Conclusions:

• MultiStain-CycleGAN is a robust solution for stain normalization across diverse datasets.
• The approach enhances CAD reliability by mitigating staining-induced biases.
• Improved normalization facilitates better patient data privacy by disguising image origins.