Stain transfer using Generative Adversarial Networks and disentangled features.

Area of Science:

• Digital pathology
• Computational imaging
• Machine learning in histopathology

Background:

• Histopathology image color variation hinders machine learning performance.
• Existing stain transfer methods have limitations like manual feature extraction and single-style transfer.

Purpose of the Study:

• To develop novel Generative Adversarial Network (GAN) based models for histopathology stain transfer.
• To overcome limitations of existing methods, enabling many-to-one and many-to-many stain transformations.

Main Methods:

• Utilized GANs with feature disentanglement to extract color and structural features automatically.
• Developed two models: one for single transformations and another for many-to-many transformations.
• Evaluated models on the Mitosis-Atypia Dataset and three additional datasets.

Main Results:

• Both proposed models demonstrated strong performance in stain transfer.
• The second model achieved superior results, outperforming six state-of-the-art methods on the Mitosis-Atypia Dataset.
• The second model achieved a Histogram Intersection Score (HIS) of 0.88 (L-channel), 0.85 (a-channel), and 0.75 (b-channel), and 90.3% classification accuracy.

Conclusions:

• The proposed GAN-based models effectively address color variation in histopathology images.
• The many-to-many stain transformation model shows significant promise for improving machine learning applications in digital pathology.