A human-in-the-loop explanation framework for morphologically transparent AI predictions from whole-slide images

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in medicine

Background:

• Deep learning models for whole-slide image (WSI) analysis lack transparency, functioning as 'black boxes'.
• This opacity hinders interpretability and clinical adoption of AI in diagnostics.
• Understanding the histomorphological patterns driving AI predictions is crucial for trust and verification.

Purpose of the Study:

• To develop and validate MorphoXAI, a human-in-the-loop framework for interpretable deep learning in WSI analysis.
• To provide both global and local explanations grounded in expert-interpreted morphology.
• To enhance pathologist-AI collaboration and facilitate transparent clinical deployment.

Main Methods:

• Proposed MorphoXAI, a framework integrating human-expert interpretations into deep learning models.
• Implemented global interpretability to identify patterns distinguishing WSI classes and confusions.
• Implemented local interpretability to pinpoint patterns and regions relevant to individual WSI predictions.

Main Results:

• MorphoXAI generated explanations accurately reflecting histomorphology at both global and local levels across diverse WSI tasks.
• Human evaluation confirmed explanations were easy to interpret, diagnostically rich, and aided decision-making.
• The framework successfully enhanced pathologist-AI collaboration.

Conclusions:

• Unifying global and local explanations grounded in expert morphology improves interpretability and verifiability of deep learning models.
• MorphoXAI facilitates transparent deployment of AI in clinical diagnostic settings.
• The framework supports enhanced pathologist-AI collaboration for improved diagnostic accuracy.