Yet another STAin NORmalization Method: Point Set Registration for Color Space Alignment in Histological Images.

Area of Science:

• Digital pathology
• Computational imaging
• Machine learning in medicine

Background:

• Machine learning and deep learning significantly advanced automatic histology image analysis.
• Domain shift, where models encounter images with different visual characteristics than training data, challenges model generalization.
• Stain normalization techniques are preprocessing methods used to mitigate domain shift in histological image analysis.

Purpose of the Study:

• To introduce YSTANORM (Yet another STain NOrmalization Method), a novel color normalization algorithm for histology images.
• To evaluate the impact of YSTANORM on the performance of state-of-the-art histological image segmentation models.
• To compare YSTANORM against existing stain normalization methods and unnormalized data.

Main Methods:

• YSTANORM aligns pixel data points in color or optical density space.
• The method uses point set registration to compute transformations between reference and target image color distributions.
• Segmentation models were evaluated on unnormalized images, images normalized with various methods, and images processed with YSTANORM.

Main Results:

• Most traditional stain normalization methods negatively impacted model performance.
• YSTANORM demonstrated an ability to improve the performance of histological segmentation models.
• The effectiveness of YSTANORM was dependent on selecting an appropriate reference image.

Conclusions:

• YSTANORM offers a promising approach to mitigate domain shift in histology image analysis.
• Careful selection of reference images is crucial for optimal performance with YSTANORM.
• The findings suggest YSTANORM can enhance the reliability and generalization of deep learning models in digital pathology.