PathNarratives: Data annotation for pathological human-AI collaborative diagnosis.

Area of Science:

• Digital pathology
• Medical artificial intelligence (AI)
• Explainable AI (XAI)

Background:

• Pathology is crucial for clinical diagnosis, but AI adoption is limited by a lack of explainability.
• Pathologists require clear rationale behind AI diagnostic suggestions for effective integration into practice.

Purpose of the Study:

• To introduce PathNarratives, a novel annotation framework for explainable AI in pathology.
• To develop a colorectal pathology dataset (CR-PathNarratives) using this framework.
• To evaluate the impact of explainable AI on diagnostic accuracy and human-AI collaboration.

Main Methods:

• Developed PathNarratives: a hierarchical decision-to-reason data structure with a narrative annotation process and interactive tool.
• Recruited 8 pathologists to annotate 174 whole-slide images (WSIs) for the CR-PathNarratives dataset.
• Conducted classification and captioning experiments using the dataset to assess AI performance and collaboration.

Main Results:

• Fine-grain classification improved accuracy from 79.56% to 85.26%.
• Pathologist trust and confidence in AI increased from 3.88 to 4.63 after receiving detailed explanations.
• Classification and captioning tasks showed improved results with reason labels, offering explainable clues.

Conclusions:

• PathNarratives facilitate the creation of explainable AI models in pathology.
• Explainable AI enhances diagnostic accuracy and fosters greater trust and confidence in human-AI collaboration.
• The developed dataset and methods support the advancement of AI in pathological diagnosis.