INFORMER framework uses saliency-map graph analysis to flag unreliable chest X-ray AI predictions

A new quality control framework for multi-label chest X-ray classification improved mean F1 score (0.574 vs 0.563) and sensitivity (0.752 vs 0.700) over output-based methods, with up to 21.92% boost in bootstrapped F1 under noise. Evaluated on CheXpert, the INFORMER approach use…

• The retrieval-based extension offers case-based explanations for flagged outputs; clinicians evaluated their relevance for interpretability.
• A graph-based class-distinctiveness method analyzes saliency-derived information to identify unreliable multi-label predictions.
• In bootstrapped evaluation with added image noise (0.001–0.005), the retrieval variant improved F1 by 21.92% over baseline, while the non-retrieval variant improved by 13.53%.

Automated summary

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