Data-driven color augmentation for H&E stained images in computational pathology.

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Whole Slide Images (WSIs) are crucial for computational pathology but suffer from stain color heterogeneity due to varied acquisition parameters across institutions.
• This heterogeneity hinders the robustness of Convolutional Neural Networks (CNNs), the leading algorithms for WSI analysis.
• Existing methods like Hue-Saturation-Contrast (HSC) and stain augmentation partially address this but lack comprehensive reliability.

Purpose of the Study:

• To introduce Data-Driven Color Augmentation (DDCA), a novel method to enhance the reliability of training data for computational pathology models.
• To improve the efficiency and robustness of color augmentation techniques by filtering unrealistic augmented samples.
• To validate DDCA's effectiveness in improving CNN performance on WSI classification tasks, particularly in the presence of stain color variations.

Main Methods:

• DDCA utilizes a large reference database of over 2 million Hematoxylin and Eosin (H&E) color variations from diverse datasets.
• Augmented data with unrealistic color distributions are discarded during Convolutional Neural Network (CNN) training.
• DDCA is integrated with HSC color augmentation, stain augmentation, and H&E-adversarial networks for colon and prostate cancer classification.

Main Results:

• DDCA significantly improves classification performance on unseen Whole Slide Images (WSIs) with heterogeneous color variations.
• Comparative analysis against 11 state-of-the-art methods demonstrates DDCA's superior ability to handle stain color heterogeneity.
• The method enhances the reliability of augmented data, leading to more robust computational pathology models.

Conclusions:

• Data-Driven Color Augmentation (DDCA) effectively addresses the challenge of stain color heterogeneity in Whole Slide Images (WSIs).
• DDCA improves the performance and robustness of Convolutional Neural Networks (CNNs) in computational pathology tasks.
• This approach offers a reliable strategy for generating high-quality training data, advancing the field of digital pathology.