Deeply supervised two stage generative adversarial network for stain normalization.

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Histopathological image color variations pose challenges for computational pathology and deep learning methods.
• Existing stain normalization techniques often suffer from low texture retention, poor performance on small datasets, or limited generalization.

Purpose of the Study:

• To propose a novel Deep Supervised Two-stage Generative Adversarial Network (DSTGAN) for robust stain normalization.
• To enhance the learning capacity and generalization of generative adversarial networks through deep supervision and semi-supervised strategies.
• To improve texture retention in normalized histopathological images.

Main Methods:

• Developed DSTGAN incorporating deep supervision and model regularization for enhanced learning.
• Implemented a novel two-stage staining strategy leveraging semi-supervised concepts for effective training.
• Designed a generator capable of capturing long-distance semantic relationships to preserve texture information.

Main Results:

• Achieved state-of-the-art performance on multiple benchmark datasets (TUPAC-2016, MITOS-ATYPIA-14, ICIAR-BACH-2018, MICCAI-16-GlaS).
• Improved classification and segmentation precision by 5.2% and 4.2%, respectively.
• Demonstrated superior image quality and texture retention compared to existing methods.

Conclusions:

• DSTGAN effectively reduces the impact of staining variations on computational pathology.
• The proposed method significantly enhances the performance of downstream classification and segmentation tasks.
• DSTGAN offers a promising solution for improving the accuracy of histopathological image analysis.