A systematic evaluation of grayscale conversion methods for mitigating color variation in deep learning-based...

Area of Science:

• Digital pathology
• Machine learning in healthcare
• Computational imaging

Background:

• Deep learning (DL) for histopathological analysis faces challenges due to color variations from staining and scanning.
• Morphological features in hematoxylin and eosin slides are crucial for diagnosis.
• Standardizing input data is essential for robust DL model performance.

Purpose of the Study:

• To investigate grayscale conversion as a method to standardize histopathological images for DL analysis.
• To evaluate the performance of various grayscale algorithms against RGB color images.
• To introduce and assess a novel attention-based grayscale conversion method (ACSRM).

Main Methods:

• Evaluated six grayscale algorithms and RGB across single-center, mixed-center, cross-scanner, and cross-center generalization tests.
• Introduced and tested an attention-based grayscale conversion method (ACSRM) using transformer attention mechanisms.
• Utilized statistical tests (Wilcoxon signed-rank, McNemar's) to compare model performance and decision behaviors.

Main Results:

• Grayscale methods achieved performance comparable to RGB in homogeneous settings.
• In mixed-center training, grayscale algorithms outperformed RGB in several models, showing distinct decision behaviors.
• ACSRM and Luster demonstrated superior generalization under distribution-shift scenarios (cross-scanner and cross-center tests), significantly improving F1-scores.

Conclusions:

• Grayscale conversion is an effective strategy for standardizing histopathological images, mitigating color variations.
• The novel ACSRM and Luster methods show significant promise for enhancing DL generalization in digital pathology.
• These findings facilitate the reliable clinical deployment of deep learning in histopathology.