M2OTCA: Multiple-magnification optimal transport-based cross-attention learning for whole slide image classification

Area of Science:

• Digital pathology
• Computational oncology
• Machine learning in medicine

Background:

• Accurate whole slide image (WSI) classification is crucial for early cancer diagnosis.
• Existing multiple instance learning (MIL) methods struggle with overfitting due to limited slide-level labels and single-magnification analysis.
• There's a need to leverage multi-magnification information for more robust WSI classification.

Purpose of the Study:

• To propose a novel multiple-magnification optimal transport-based cross-attention (M2OTCA) MIL framework for WSI classification.
• To address the limitations of single-magnification feature mining and weak supervision in existing MIL approaches.
• To enhance feature consistency across different magnifications for improved diagnostic accuracy.

Main Methods:

• Developed an M2OTCA framework utilizing optimal transport (OT) to match feature distributions across magnifications.
• Introduced region structural prototypes to condense instances and reduce computational complexity of OT learning.
• Derived region structural prototype gradients for cross-magnification explanation and feature co-expression.

Main Results:

• The M2OTCA framework demonstrated superior performance in WSI classification compared to state-of-the-art methods.
• Experiments on four WSI datasets validated the effectiveness of the proposed multi-magnification approach.
• The method successfully integrated information from different magnifications for more accurate predictions.

Conclusions:

• The M2OTCA framework offers a significant advancement in automated WSI classification for cancer diagnosis.
• Leveraging cross-magnification feature consistency through optimal transport enhances model robustness and accuracy.
• This approach provides a promising direction for improving computational pathology tools.