CellMix: A General Instance Relationship-Based Method for Data Augmentation Toward Pathology Image Classification.

Area of Science:

• Digital pathology
• Computational imaging
• Machine learning for healthcare

Background:

• High-quality annotated pathology samples are crucial but labor-intensive to obtain.
• Existing data augmentation methods struggle with pathology images' unique characteristics like local specificity and instance relationships.

Purpose of the Study:

• To introduce CellMix, a novel framework for pathology image data augmentation.
• To address limitations of current methods by preserving and introducing instance relationships.

Main Methods:

• CellMix utilizes a distribution-oriented in-place shuffle approach, dividing images into patches and shuffling them within batches.
• A curriculum learning-inspired, loss-driven strategy controls relationship augmentation for adaptive instance exploration.
• The method preserves locational instance relationships while introducing novel ones.

Main Results:

• CellMix achieved state-of-the-art (SOTA) performance on pathology image classification tasks.
• Demonstrated superior results across seven diverse datasets.
• Effectively handles distribution-related noise and varying difficulties.

Conclusions:

• CellMix offers an innovative instance relationship-centered approach for pathology image classification.
• This method advances general data augmentation techniques in digital pathology.
• The framework shows significant potential for improving automated analysis of histopathological images.