Saleh Alzughaibi
In this evaluation of a single model on a single public dataset, surface-level fluency and readability did not reliably indicate factual correctness. These findings apply primarily to RadVLM in the experimental setting studied and should not be generalized to AI report generation as a whole without multi-model, multi-dataset replication. Even so, they reinforce a broader principle: clinical deployment of such systems requires explicit entity-relation validation, structured accuracy assessment, and radiologist oversight to protect patient safety.
BACKGROUND: Artificial intelligence-generated radiology reports may reduce clinician burden, but their factual accuracy remains uncertain. This study evaluated whether surface-level semantic similarity in AI-generated Impression sections reliably reflects structural fidelity in the underlying Findings.
METHODS: RadVLM, a radiology-focused vision-language model, generated Findings and Impression sections for 3,000 public chest radiograph studies. A section-aware evaluation framework assessed Findings-level entity-relation fidelity using RadGraph F1 and Impression-level quality using BERTScore-F1 (semantic similarity), CheXbert cosine similarity (label-space agreement), and readability metrics. Generalized linear models examined associations between Impression-level metrics and Findings-level fidelity.
RESULTS: Mean RadGraph F1 was 0.251 (95% CI, 0.245-0.257), indicating substantially imperfect structural correctness. Mean BERTScore-F1 was 0.498 and CheXbert cosine similarity was 0.394. AI-generated Impressions were more readable (mean Flesch-Kincaid 11.72 vs. 15.33 for reference). A weak positive association existed between semantic similarity and entity-relation fidelity (coefficient 0.154; 95% CI, 0.124-0.183), and this association attenuated substantially with longer Impressions. In practical terms, the association was too weak to be useful clinically: a large gain in semantic similarity corresponded to only a small change in structural fidelity, so a fluent, readable Impression offered little assurance that the underlying Findings were factually correct.
CONCLUSIONS: In this evaluation of a single model on a single public dataset, surface-level fluency and readability did not reliably indicate factual correctness. These findings apply primarily to RadVLM in the experimental setting studied and should not be generalized to AI report generation as a whole without multi-model, multi-dataset replication. Even so, they reinforce a broader principle: clinical deployment of such systems requires explicit entity-relation validation, structured accuracy assessment, and radiologist oversight to protect patient safety.