Recent Advances in Artificial Intelligence for Radiology Report Generation: A Brief Review

Purpose: Rising imaging volumes combined with radiologist workforce shortages have created workflow bottlenecks that may be addressed by artificial intelligence (AI). This review evaluates recent advancements in AI vision-language models (VLMs) for radiology report generation.

Methods/Materials: A focused literature search of PubMed, IEEE Xplore, arXiv, and Google Scholar identified studies published from 2020 to 2024. Of 8,772 studies, 14 met inclusion criteria requiring explicit report generation evaluation. Extracted data included model architectures, training datasets, and metrics, including Bi-Lingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), Metric for Evaluation of Translation with Explicit Ordering (METEOR), Consensus-based Image Description Evaluation (CIDEr), CheXbert F1, RadGraph F1. All included studies used de-identified publicly available datasets, complying with HIPAA regulations.

Results: Of the 14 included studies, 11 VLMs were trained primarily on chest radiograph datasets, most commonly MIMIC-CXR, with limited representation of other imaging modalities. The various optimization strategies demonstrated trade-offs between linguistic fidelity and clinical accuracy. For example, masked modeling led to higher natural language generation performance, especially in METEOR and CIDEr-D, but showed lower clinical entity-relation accuracy as measured by RadGraph F1. Meanwhile, contrastive alignment increased entity-relation accuracy but underperformed in n-gram precision. Models incorporating multiple optimization strategies demonstrated more balanced performances. Human evaluation was in 4 studies, which showed that although AI reports were rated equal or better than human-written reports in 40–60% of cases, AI reports had higher rates of omission and errors, especially for complex findings.

Conclusions: Although AI VLMs show potential for automated radiology report generation, they remain limited by a lack of multimodal training datasets, unstandardized evaluation metrics, and errors. If limitations are addressed and VLMs are properly integrated with human oversight, VLMs could serve as assistive tools that enhance efficiency and patient care.