Related Experiment Video
Updated: Apr 22, 2026
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
1.3K
Reporting efficiency in diagnostic imaging: Can plug-and-play general-purpose large language models outperform
Constance de Margerie-Mellon1, Loic Duron2, Laure Fournier3
1Université Paris Cité, PARCC UMRS 970, INSERM, AP-HP, Hôpital Saint-Louis, Department of Radiology, 75010, Paris, France. constance.de-margerie@aphp.fr.
European Radiology
|April 20, 2026
Summary
Large language models (LLMs) with speech recognition can generate radiology reports faster and with fewer errors than conventional speech recognition (CSR). However, time savings vary among radiologists, and new error types emerge with LLM use.
Area of Science:
- Medical Informatics
- Radiology Reporting
- Artificial Intelligence in Healthcare
Background:
- Conventional speech recognition (CSR) is widely used for radiology reporting.
- Efficiency and accuracy in report generation are critical for clinical workflow.
- General-purpose large language models (LLMs) offer potential advancements in text generation.
Purpose of the Study:
- To compare the efficiency and accuracy of LLMs versus CSR in radiology report generation.
- To evaluate generation times and error types associated with LLM and CSR use.
- To assess the clinical relevance and implementation of LLMs in radiology.
Main Methods:
- Prospective, multicenter study involving five radiologists.
- Comparison of 200 reports generated with CSR and 200 with a general-purpose LLM.
- Analysis of generation times and qualitative/quantitative error evaluation (Levenshtein distance).
Main Results:
- LLM group showed shorter median total generation time (238s vs 318s, p<0.01).
- LLM group had fewer grammar/spelling (79 vs 293) and transcription errors (225 vs 445, p<0.01).
- LLM use introduced new error patterns, including rewording, non-compliance, and confabulations; Levenshtein distance was higher (43 vs 20, p<0.01).
Conclusions:
- General-purpose LLMs can improve radiology report generation speed and reduce common errors compared to CSR.
- Time savings with LLMs are heterogeneous and depend on individual radiologist practices.
- While LLMs offer benefits, careful consideration of new error types and implementation strategies is necessary.
