Uncertainty-Aware Information Pursuit for Interpretable and Reliable Medical Image Analysis

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• AI systems require human-interpretable decisions for safety-critical domains like medical image analysis.
• Variational Information Pursuit (VIP) provides interpretable-by-design AI by querying human-understandable concepts.
• Existing VIP methods lack sample-specific uncertainty handling, leading to suboptimal query selection and reduced robustness.

Purpose of the Study:

• To develop an interpretable and uncertainty-aware AI framework for medical imaging.
• To address limitations in existing VIP methods by accounting for upstream uncertainties in concept predictions.
• To enhance the reliability and robustness of AI decisions in medical image analysis.

Main Methods:

• Proposed two uncertainty-aware models: EUAV-IP and IUA-VIP, integrating uncertainty estimates into the VIP querying process.
• EUAV-IP utilizes masking to skip uncertain concepts.
• IUA-VIP implicitly incorporates uncertainty into query selection for more informed decisions.

Main Results:

• The proposed IUA-VIP model achieved state-of-the-art accuracy among interpretable-by-design approaches on four out of five medical imaging datasets.
• The models demonstrated reliable decision-making based on sample-tailored concept subsets.
• Generated more concise explanations by selecting fewer, more informative concepts.

Conclusions:

• The developed framework enhances AI trustworthiness and supports safer AI deployment in healthcare.
• Accounting for sample-specific uncertainty improves the clinical alignment and reliability of interpretable AI models.
• The approach enables AI to make reliable decisions using a subset of concepts specific to each sample, without human intervention.