Background: Automated chest X-ray report generation has advanced significantly with the emergence of vision–language models (VLMs), offering the potential to enhance clinical workflow efficiency. However, evaluating generated reports remains a funda...
Background: Automated chest X-ray report generation has advanced significantly with the emergence of vision–language models (VLMs), offering the potential to enhance clinical workflow efficiency. However, evaluating generated reports remains a fundamental challenge. Conventional metrics—whether lexical, semantic, or entity-based—often fail to capture clinical correctness, overlooking subtle yet critical errors in findings, anatomical locations, severity assessments, or temporal comparisons. Large language model (LLM)-based evaluators achieve stronger alignment with expert judgment but are computationally intensive and slow, limiting scalability in real-world clinical settings.
Purpose: This study introduces a lightweight, regression-based metric designed to provide rapid and interpretable evaluation of chest X-ray reports by explicitly modeling clinically meaningful error types.
Methods: The proposed metric processes a reference–candidate report pair using a compact, domain-specific BERT-based encoder and jointly predicts six fine-grained error categories, including false findings, omissions, mislocalizations, and incorrect statements of severity or comparison. A dual-head architecture simultaneously regresses error counts and detects the presence of each error type, mitigating class imbalance and improving count accuracy. Training was conducted on 32,000 synthetic report pairs derived from the publicly available MIMIC-CXR dataset. Candidate reports were generated using a specialized vision–language foundation model, and category-specific error annotations were obtained through structured comparisons with a commercial LLM under secure governance. Evaluation was performed on multiple radiologist-annotated benchmarks, including ReXVal—a dataset containing expert error counts across six predefined clinical categories—as well as ReFiSco-v0, RadEvalX, and RaTE-Eval.
Results: On ReXVal, the metric achieved a Kendall’s τ of 0.786 (95% CI: 0.749–0.816) and a Spearman’s ρ of 0.933 (95% CI: 0.907–0.949), significantly outperforming traditional metrics (p < 0.001) while showing no statistically significant difference from the computationally intensive GREEN error count metric (p > 0.05). On RadEvalX, it attained τ = 0.580 (95% CI: 0.477–0.669) and ρ = 0.745 (95% CI: 0.625–0.826), significantly surpassing the GREEN error count metric (p = 0.013), while reducing mean absolute error by 33% compared with the LLM-based metric on RaTE-Eval. Inference time averaged 9.53 ms per sample, approximately 280 times faster than GREEN (2,642 ms). Robustness testing under class-imbalance filtering (ReXVal*) showed minimal performance degradation (Δτ = –0.033).
Conclusions: A BERT-based multi-target regression approach can efficiently and accurately quantify fine-grained clinical errors in generated radiology reports. The method demonstrates performance with no statistically significant difference from, and in out-of-distribution scenarios superior to, state-of-the-art LLM-based metrics, while offering significantly higher computational efficiency. This makes it a practical and scalable alternative to direct LLM-based evaluation in medical AI workflows.