Towards Visually Explaining Statistical Tests with Applications in Biomedical Imaging

Black-box AI models limit medical adoption: The paper identifies that while deep neural networks are powerful at detecting differences between patient groups in medical imaging, their lack of transparency has prevented wider use in clinical practice—highlighting a key barrier to responsible AI deployment in healthcare.

Explainability methods need to work without labels: Most existing explanation techniques require labeled data, but statistical testing often doesn't have labels. This paper addresses that gap by developing explanations that work in "label-free" settings, making the approach more practical for real-world medical analysis where such labels may not be available.

Visual explanations reveal which patients and regions matter: The framework identifies both "which individual samples" (specific patients) and "which input features" (anatomical regions) drive detected differences between groups, providing clinicians with actionable insights about disease-related variation rather than just a yes/no answer.

Anatomically meaningful results build trust: By highlighting "anatomically meaningful regions associated with disease-related variation," the method produces explanations that align with medical knowledge, making it easier for clinicians to validate whether the AI's reasoning makes sense—a crucial step for responsible medical AI.

Bridges two important fields: This work connects statistical inference with explainable AI, suggesting that responsible medical AI requires both rigorous statistical foundations and interpretability, not one or the other.