Beyond Post hoc Explanations: A Comprehensive Framework for Accountable AI in Medical Imaging Through Transparency,...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Explainable AI

Background:

• Deep learning in medical imaging offers diagnostic advances but faces clinical adoption barriers due to its "black-box" nature.
• Current explainable AI (XAI) methods often provide post hoc explanations, potentially misrepresenting AI model confidence and impacting clinical decisions.

Purpose of the Study:

• To systematically review and meta-analyze the fidelity, stability, and performance trade-offs of XAI methods across medical imaging modalities.
• To evaluate the impact of noise perturbation on XAI stability in radiology, pathology, and ophthalmology.
• To propose an accountability framework for developing and deploying trustworthy AI in clinical settings.

Main Methods:

• Systematic review and meta-analysis of 67 studies evaluating XAI in 23 radiology, 19 pathology, and 25 ophthalmology applications.
• Analysis of 847 identified studies to compare the fidelity of LIME, SHAP, and Grad-CAM.
• Assessment of XAI stability under noise perturbation across different medical imaging modalities.

Main Results:

• LIME demonstrated superior fidelity (0.81) compared to SHAP (0.38) and Grad-CAM (0.54) across all modalities.
• Post hoc explanations exhibited poor stability under noise; SHAP degraded by 53% in ophthalmology versus 11% in radiology.
• Interpretable models incurred a consistent 5-7% AUC performance penalty, with modality-specific stability patterns observed.

Conclusions:

• Tailored XAI approaches are essential due to modality-specific stability patterns.
• A three-pillar accountability framework is proposed, emphasizing transparency, interpretable design, and cautious deployment of post hoc explanations with uncertainty quantification.
• This framework aims to enhance clinical decision-making quality and patient safety by fostering genuinely accountable AI systems.