GHMS-CycleGAN: Graph-Based Hierarchical Multi-stain CycleGAN for Stain Normalization and Classification in Digital...

Area of Science:

• Digital pathology and medical imaging analysis
• Artificial intelligence in healthcare
• Computational pathology

Background:

• Digital pathology (DP) offers computer-aided diagnosis (CAD) tools using deep learning (DL).
• Staining variations and scanner distortions degrade DL-based CAD system performance.
• Robust image normalization is crucial for reliable DL models in pathology.

Purpose of the Study:

• To introduce novel stain and scan normalization methods for histopathology, RBC, and WBC image classification.
• To develop a framework using hierarchical multi-stain CycleGAN (HMS-CycleGAN) and graph-based HMS-CycleGAN (GHMS-CycleGAN).
• To evaluate the impact of normalization on downstream image classification tasks.

Main Methods:

• Proposed HMS-CycleGAN and GHMS-CycleGAN for multi-stain and color normalization.
• Exploited hierarchical and network structures in DP images for normalization mapping.
• Validated on diverse histopathology, RBC, and WBC datasets from multiple scanners and institutions.

Main Results:

• HMS-CycleGAN demonstrated superior robustness and consistency across staining variations compared to state-of-the-art methods.
• Normalization methods consistently improved classification performance over unnormalized classifiers.
• Achieved 84.05% accuracy on Wilds Camelyon and 51.06% F1-score for WBC classification.
• HMS-CycleGAN and GHMS-CycleGAN reached 80-91% F1-scores for cross-scanner RBC classification.

Conclusions:

• Image normalization is vital for building robust and generalizable DL models in digital pathology.
• The proposed HMS-CycleGAN and GHMS-CycleGAN methods significantly enhance DL-based image classification performance.
• These normalization techniques address key challenges in digital pathology, paving the way for improved diagnostic tools.