Using generative AI to investigate medical imagery models and datasets.

Area of Science:

• Medical Imaging AI
• Explainable AI (XAI)
• Computer Vision in Healthcare

Background:

• AI models show promise in medical imaging but lack explainability.
• Understanding AI decision-making is crucial for clinical trust and novel discoveries.
• Current AI explanations are insufficient for complex medical tasks.

Purpose of the Study:

• To develop a workflow for generating hypotheses about visual signals AI models learn for medical image classification.
• To enhance trust in AI by explaining its predictions and uncovering new insights.
• To enable scientific discovery by identifying unknown signals in medical data.

Main Methods:

• A 4-step workflow involving classifier training, StyleGAN-based image generation (StylEx), automatic attribute visualization, and interdisciplinary expert review.
• Generating counterfactual visualizations by modifying detected attributes to understand their impact on predictions.
• Presenting discovered attributes and visualizations to experts to formulate hypotheses on underlying mechanisms.

Main Results:

• Demonstrated applicability across eight tasks and three medical imaging modalities (retinal, eye, chest).
• Identified clinically known features, automatically-learned confounders (e.g., X-ray underexposure, eye makeup), and novel, physiologically plausible attributes (e.g., sex-related fundus differences).

Conclusions:

• The approach facilitates hypothesis generation for better understanding, assessment, and knowledge extraction from AI models.
• Highlights the importance of interdisciplinary perspectives to interpret attributes reflecting real-world healthcare delivery and socio-cultural factors.
• Code will be released to enable researchers to train StylEx models and responsibly interpret revealed attributes.