Tackling stain variability using CycleGAN-based stain augmentation.

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Pathology exhibits significant inter- and intra-laboratory stain variability, hindering the development and application of deep learning (DL) models.
• Manual annotation to address all stain variability is impractical, necessitating automated solutions.
• Investigated unsupervised DL approaches to mitigate stain variability in kidney pathology.

Purpose of the Study:

• Improve the applicability of a pretrained DL segmentation model to external multi-centric cohorts with substantial stain variability.
• Compare data augmentation using CycleGANs against traditional stain normalization methods for DL models.

Main Methods:

• Augmented training data with stain variability using CycleGANs.
• Retrained a pretrained DL segmentation model on the stain-augmented dataset.
• Compared the stain-augmented model against an unmodified pretrained model and CycleGAN-based stain normalization.

Main Results:

• The proposed stain-augmented model achieved the highest mean segmentation accuracy across all external cohorts.
• Performance on the training cohort remained comparable.
• CycleGAN-based stain normalization slightly degraded performance due to imperceptible encoded information.

Conclusions:

• Data augmentation with stain variability is more effective than stain normalization for improving DL model performance in pathology.
• The practical applicability of this approach requires balancing the slight performance increase against the computational cost (carbon footprint).