Semantic fluency doesn't guarantee factual accuracy in RadVLM chest X-ray reports

A radiology vision-language model (RadVLM) generated easier-to-read chest X-ray impressions, but readability weakly correlated with factual accuracy (association coefficient 0.154; 95% CI 0.124-0.183), so a fluent report could hide incorrect findings.

• Entity‑relation fidelity was low (mean RadGraph F1 0.251), indicating substantially imperfect structural correctness in AI‑generated Findings.
• AI‑generated Impressions were more readable than reference reports (mean Flesch‑Kincaid 11.72 vs. 15.33), but surface fluency did not reliably signal factual correctness.
• The evaluation is limited to a single model on a single public dataset; findings require multi‑model, multi‑dataset replication before generalization.

Automated summary

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