CoD-MIL: Chain-of-Diagnosis Prompting Multiple Instance Learning for Whole Slide Image Classification

Area of Science:

• Digital Pathology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Multiple Instance Learning (MIL) is key for analyzing large whole slide images (WSIs) in digital pathology.
• Existing MIL methods struggle with generalization and interpretability due to reliance on image data and fixed labels.
• Vision Language Models (VLMs) show promise for improving MIL performance and transferability.

Purpose of the Study:

• To develop an advanced MIL framework for whole slide image classification that addresses limitations of current methods.
• To enhance the interpretability and generalization of MIL models by incorporating natural language processing techniques.
• To improve the accuracy of tumor region localization in WSIs.

Main Methods:

• Proposed a novel Chain-of-Diagnosis prompting Multiple Instance Learning (CoD-MIL) framework.
• Utilized a chain-of-diagnosis text prompt to break down complex WSI diagnostic processes into progressive stages.
• Introduced a text-guided contrastive masking module for precise tumor region localization using discriminative instance masking and normal tissue text guidance.

Main Results:

• CoD-MIL demonstrated significant effectiveness in whole slide image classification tasks.
• The framework showed superiority over existing methods in experiments on three real-world subtyping datasets.
• The text-guided contrastive masking module successfully improved tumor region localization accuracy.

Conclusions:

• The proposed CoD-MIL framework offers a powerful new approach for whole slide image classification in digital pathology.
• Integrating chain-of-thought principles and text guidance enhances the diagnostic capabilities of MIL models.
• CoD-MIL provides a more interpretable and generalizable solution for pathological diagnosis using WSIs.