3SGAN: Semi-Supervised and Multi-Task GAN for Stain Normalization and Nuclei Segmentation of Histopathological Images

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Staining variations in digital pathology challenge automated cell segmentation.
• Manual nucleus annotation is labor-intensive, limiting dataset availability.
• A need exists for robust segmentation methods that handle staining variability with minimal annotations.

Purpose of the Study:

• To develop a novel framework mitigating staining variability and enabling high-accuracy nucleus segmentation.
• To achieve robust performance using minimal annotation data.
• To provide a practical and scalable solution for digital pathology.

Main Methods:

• A multi-task dual-branch generative adversarial network (GAN) named 3SGAN was developed.
• The framework employs a semi-supervised, teacher-student paradigm (AttCycle and TransCycle models).
• Trained and evaluated on 1408 Whole-Slide Images (WSIs) with 101 staining styles, using annotations for only 5% of the data.

Main Results:

• 3SGAN achieved superior nucleus segmentation accuracy (F1-score: 0.8140, mean IoU: 0.8201, AJI: 0.6915).
• Significant improvements in stain normalization quality were observed (RMSE: 0.0908, PSNR: 21.0615, SSIM: 0.8556).
• External validation confirmed strong generalizability across diverse datasets and staining protocols.

Conclusions:

• High-performance nucleus segmentation and stain normalization are achievable with minimal annotations.
• The 3SGAN framework offers a practical and scalable solution for digital pathology.
• The approach is suitable for diverse clinical settings and staining protocols.