Uncertainty-aware chest X-ray report generation using conformal prediction

Applying conformal prediction to vision-language model report drafting for chest X-rays, high-confidence outputs showed significantly better agreement with radiologist impressions (P<0.001) and higher classification AUROC (P<0.05).

• Label-based classification on ChestX-Det10 (n=3001): the high-confidence subgroup achieved significantly higher AUROC across multiple thoracic pathologies (P<0.05).
• Sentence-based generation on Open-I (n=3660): certain outputs had greater semantic similarity to ground-truth impressions (P<0.001), validated by an independent LLM evaluator.
• CONRep is a model-agnostic conformal prediction framework, applicable to any VLM to flag uncertain outputs and improve trustworthiness.

Automated summary

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