Training Set Design for Uneven Illumination Correction in High-Resolution Whole Slide Images.

Area of Science:

• Digital Pathology
• Computational Imaging
• Machine Learning in Microscopy

Background:

• Uneven illumination correction is crucial for whole-slide imaging (WSI) preprocessing.
• Current deep learning methods struggle with generalization and computational demands.
• Patch-based training may miss global illumination patterns, causing inconsistencies.

Purpose of the Study:

• Identify limitations in deep learning for uneven illumination correction.
• Propose a training set design strategy for improved performance and resource utilization.
• Enhance generalization and consistency in WSI illumination correction.

Main Methods:

• Developed a novel training set design strategy for neural networks.
• Focused on preserving original image resolution.
• Incorporated a global view of illumination patterns during training.

Main Results:

• Achieved more uniform illumination correction across entire WSI slides.
• Reduced artifacts and improved image consistency.
• Enhanced model robustness and scalability for practical applications.

Conclusions:

• The proposed training strategy addresses key limitations in deep learning for WSI illumination correction.
• This approach makes deep learning-based correction more viable for clinical and research use.
• Optimized training improves model performance and resource efficiency.