Leveraging Expert Input for Robust and Explainable AI-Assisted Lung Cancer Detection in Chest X-rays

Expert collaboration improves AI trustworthiness in medicine: The study found that standard explainable AI techniques often failed to match how radiologists actually diagnose lung cancer, but when researchers worked directly with a radiologist to build explanations around clinically relevant concepts, the AI system's reasoning became aligned with expert practice and more trustworthy for clinical use.

Current "black-box" explanations may mislead clinicians: Existing methods for explaining AI decisions in medical imaging showed "inconsistencies and divergence from expert radiologist assessments," suggesting that without proper validation, these explanations could give false confidence in AI recommendations rather than genuine insight into how decisions are made.

Expert-informed AI is more robust to manipulation: The ClinicXAI system, which incorporated radiologist input, demonstrated "significantly greater resilience to adversarial perturbations" compared to the standard deep learning model, meaning it was less likely to be fooled by subtle image manipulations—an important safety consideration for clinical deployment.

Explainability and accuracy don't have to trade off: The expert-driven approach maintained "high diagnostic accuracy" while providing clinically meaningful explanations, challenging the assumption that making AI more interpretable requires sacrificing performance.

AI adoption in healthcare requires clinical input from the start: Rather than treating explainability as an afterthought, this research demonstrates that involving domain experts in designing AI systems from the beginning produces tools that clinicians can actually understand and trust enough to use in practice.