Unsupervised stain augmentation enhanced glomerular instance segmentation on pathology images.

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Computational Pathology

Background:

• Supervised deep learning models for glomerular instance segmentation perform poorly on unseen stains due to stain-specific feature highlighting.
• Acquiring multi-stain annotated pathology datasets is labor-intensive and time-consuming, limiting model generalizability.

Purpose of the Study:

• To propose an unsupervised stain augmentation method for robust glomerular instance segmentation across diverse staining techniques.
• To enhance the staining diversity of training datasets without manual annotation.

Main Methods:

• Utilized contrastive unpaired translation (CUT) to convert between Periodic Acid-Schiff (PAS), Masson's Trichrome (MT), and Periodic Acid-Silver Methenamine (PASM) stains.
• Replaced the Mask R-CNN backbone with Swin Transformer for improved feature extraction in instance segmentation.

Main Results:

• The proposed stain augmentation method significantly outperformed existing approaches across all metrics for PAS, PASM, and MT stains.
• Ablation experiments confirmed the effectiveness of the individual components of the proposed method.
• The model was validated on a dataset comprising 216 whole slide images (WSIs).

Conclusions:

• Unsupervised stain augmentation shows significant potential for improving glomerular segmentation in digital pathology.
• This approach can be extended to other complex segmentation tasks within pathology image analysis.