Human-AI Collaboration and Explainability for 2D/3D Registration Quality Assurance

AI alone is accurate, but human-AI teams are more reliable for safety-critical decisions: While the AI-only model achieved the highest accuracy, collaborative approaches (Human+AI and Human+XAI) demonstrated superior sensitivity, precision, and specificity—metrics that matter more in surgical contexts where missing a problem is particularly dangerous.

Explainability helps humans understand AI, but doesn't necessarily increase trust: When the AI provided explanations for its predictions (Human+XAI condition), participants reported significantly better understanding compared to AI without explanations (Human+AI). However, this improved understanding did not translate into measurably higher trust between the two collaborative conditions, suggesting that understanding and trust are distinct factors requiring separate attention.

Human-AI collaboration reduces cognitive burden without sacrificing safety: Participants experienced significantly lower workload when working with AI assistance compared to making decisions alone (Human-only condition), while maintaining or improving the quality of decisions—a critical finding for preventing human fatigue and error in high-stakes surgical environments.

Current visualization tools alone are insufficient for patient safety: The research identifies that existing methods of displaying surgical alignment information cannot reliably help humans detect misalignments, establishing a clear need for AI-assisted decision support tools in this domain rather than relying on human visual inspection alone.