IBO: Inpainting-Based Occlusion to Enhance Explainable Artificial Intelligence Evaluation in Histopathology

Better explanations require realistic image manipulation: Traditional methods for explaining AI decisions in cancer diagnosis create artificial, unrealistic images when hiding parts of tissue samples. This paper shows that using AI to fill in hidden areas with realistic tissue (rather than blurring or masking) produces explanations that are nearly twice as accurate, improving prediction precision from 42% to 71%.

"Black box" AI in medical imaging needs trustworthy evaluation methods: Deep learning models excel at analyzing histopathology images for cancer detection, but doctors and regulators need to understand why the AI makes specific decisions. This work addresses a critical gap: current methods for testing whether explanations are reliable can themselves be unreliable, potentially giving false confidence in AI systems.

Out-of-distribution artifacts undermine AI interpretability: When researchers test AI explanations by hiding parts of images, the artificial replacements (like solid colors or blurs) don't look like real tissue, causing the AI to behave abnormally. This paper demonstrates that this technical problem has real consequences—it makes it harder to accurately evaluate whether an explanation method actually works.

Generative AI can improve responsible AI practices: Using a diffusion model to realistically reconstruct tissue samples represents a practical application of generative AI beyond content creation—it enhances the ability to audit and validate medical AI systems, supporting the broader goal of trustworthy AI deployment in healthcare.