Explainable AI in radiology: interpreting and evaluating AI predictions

Radiologists can learn to interpret and evaluate artificial intelligence predictions using explainable AI methods including saliency maps, perturbation, concept-based reasoning, and uncertainty quantification, as outlined in a practical review.

• Intended for practicing radiologists and physicians to improve AI oversight.
• Covers major categories: saliency maps, perturbation/feature-attribution, concept-based methods, example-based reasoning, and uncertainty quantification.
• Addresses common misconceptions and emerging regulatory obligations.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.