Related Experiment Video
Updated: Oct 25, 2025
05:22
Author Spotlight: Enhanced Multiplex Immunofluorescent Microscopy Protocol for Neuroscience Research
Published on: June 21, 2024
574
Normalization of HE-stained histological images using cycle consistent generative adversarial networks.
Marlen Runz1,2, Daniel Rusche3, Stefan Schmidt4
1Institute of Pathology, University Medical Centre Mannheim, Heidelberg University, Mannheim, Germany. marlen.runz@medma.uni-heidelberg.de.
Diagnostic Pathology
|August 7, 2021
Summary
CycleGAN effectively normalizes histological images by reducing staining variations, improving downstream analysis and classifier performance. This method enhances image consistency across different acquisition and staining protocols.
Area of Science:
- Digital pathology
- Computational imaging
- Machine learning in medicine
Background:
- Histological images exhibit significant variance due to differences in acquisition, processing, and staining.
- These variations can hinder accurate downstream analyses like staining intensity evaluation and classification.
- Image normalization techniques are crucial for mitigating these inconsistencies.
Purpose of the Study:
- To investigate the efficacy of CycleGAN (cycle-consistent Generative Adversarial Network) for color normalization of hematoxylin-eosin (HE) stained histological images.
- To address variability in internal staining protocols and image acquisition using daily clinical data.
- To evaluate the impact of CycleGAN normalization on the performance of a pre-trained ResNet classifier.
Main Methods:
- Employed a CycleGAN architecture with generator (G) and discriminator (D) networks for image-to-image translation between source (A) and target (B) domains.
- Ensured cycle consistency (G_A(G_B(X_A)) ≈ X_A) to maintain image integrity during bidirectional mapping.
- Validated the approach on breast cancer and follicular thyroid carcinoma datasets, assessing quality with similarity measures and applying normalization to lymph node data for classifier evaluation.
Main Results:
- Achieved up to 96% improvement in similarity to target domain training images after mapping.
- Demonstrated high cycle consistency with similarity indices greater than 0.9.
- Increased the kappa-value of a ResNet classifier by over 50% when applied to normalized HE-stained lymph node images.
Conclusions:
- CycleGANs provide an efficient method for normalizing HE-stained histological images.
- The approach successfully compensates for variations from image acquisition and tissue staining protocols across different institutions.
- This overcomes staining inconsistencies, enhancing the reliability of digital pathology analyses.
