Bridging the interpretability gap for medical artificial intelligence models using class-association manifold...

Area of Science:

• Medical Artificial Intelligence (AI)
• Explainable AI (XAI)
• Machine Learning

Background:

• Intelligent medical devices increasingly require explainability.
• Current medical AI models face an 'interpretability gap', limiting trust and adoption.
• Existing methods struggle to balance explainability with diagnostic accuracy.

Purpose of the Study:

• To propose a novel generative approach, class-association manifold learning, to enhance the explainability of medical AI models.
• To address the 'interpretability gap' in medical AI while maintaining high diagnostic performance.
• To facilitate the discovery of novel medical knowledge and clinical rules.

Main Methods:

• Class-association manifold learning: A generative method to decouple decision patterns from background data.
• Low-dimensional mapping: Representing global class-associated knowledge efficiently.
• AI-generated sample modifications and differential diagnosis rule visualization.
• Topology map: Modeling decision rule sets for intuitive explication and virtual contrastive examples.

Main Results:

• Preserved near-perfect diagnostic accuracy while enhancing model explainability.
• Successfully decoupled common decision-related patterns from individual backgrounds.
• Extracted medical-compliant knowledge unknown during model training.
• Demonstrated higher accuracy in explaining medical AI model behavior compared to existing methods.

Conclusions:

• Class-association manifold learning offers a powerful solution to the interpretability challenge in medical AI.
• The method enables intuitive understanding of black-box models through topology maps and virtual examples.
• This approach has the potential to assist in clinical rule discovery and advance AI-driven medical knowledge generation.