Radiology’s AI Writing Boom: 265 Studies Map the LLM Research Surge

Large language models have swept into radiology with remarkable speed, and a new bibliometric analysis has now mapped exactly where that surge is concentrated. The study, published in the Egyptian Journal of Radiology and Nuclear Medicine, analyzed 265 journal articles indexed in the OpenAlex database and found that research on LLMs in radiology reporting, summarization, and workflow tasks effectively did not exist before 2023, then exploded into a fast-growing field dominated by a handful of journals, countries, and institutions.

The numbers tell a striking story. No eligible studies were identified between 2020 and 2022. The field emerged in 2023 with 26 publications, roughly 9.8 percent of the final dataset. Output jumped to 68 articles in 2024, a year-over-year increase of 161.5 percent, and peaked at 134 articles in 2025, representing 50.6 percent of all included studies and a further growth rate of 97.1 percent. By April 11, 2026, 37 additional articles had been indexed, bringing the total to 265 studies.

The methodological design matters for interpreting those figures. The researchers searched OpenAlex on April 11, 2026, using a predefined strategy that combined terms for large language models, radiology, and reporting or workflow tasks, while explicitly excluding primarily image-based artificial intelligence applications such as image classification, segmentation, and detection. After date restrictions and document-type filtering reduced an initial 3,511 records to 1,722, two reviewers independently screened titles and abstracts, ultimately excluding 1,457 records and leaving the final dataset of 265 studies. The reporting followed the BIBLIO guideline for bibliometric reviews of biomedical literature.

What the analysis found was a literature firmly anchored in text rather than pixels. Thematic mapping of the top 50 co-occurring terms from titles and abstracts revealed dominant clusters centered on structured reporting, report simplification, impression generation, information extraction, patient-centered communication, and workflow-oriented uses of LLMs. One cluster revolved around large language models and radiology reports, including terms like GPT-4, free-text radiology reports, and fine-tuned and open-source models. Another organized around ChatGPT, encompassing patient understanding, patient-friendly summaries, and report simplification. Further clusters captured MRI reports, impression generation, and report generation alongside prompt engineering and retrieval-augmented generation.

The geographic concentration is equally pronounced. The United States led all contributing countries, appearing in 89 studies, or 33.6 percent of the dataset, with 55 of those being intra-national collaborations and 34 internationally collaborative. Germany ranked second with 46 studies, and China third with 33. Japan contributed 17 studies, while the United Kingdom, South Korea, and Türkiye tied for fifth place with 13 studies each. Notably, the collaboration profiles differed: the UK’s output was heavily international, with 9 of its 13 studies involving co-authors from other countries, whereas Japan’s was predominantly domestic, with 14 of 17 studies intra-national.

Institutional productivity showed a similar concentration. Humboldt-Universität zu Berlin ranked first with 12 publications, followed by Mayo Clinic in Arizona and University Hospital Cologne with 9 each, and Charité—Universitätsmedizin Berlin and Deutsches Herzzentrum München with 8 each. Journal output was likewise concentrated: Radiology published the most studies with 18, followed by the Journal of Imaging Informatics in Medicine with 15 and European Radiology with 12. The top five journals collectively accounted for 61 publications, just 23 percent of the included studies.

Citation patterns reinforce the picture of a young field shaped by early landmark papers. The 265 studies accumulated 4,577 citations, with a mean of 17.3 citations per article but a median of only 3, and an H-index of 32. Eighty-four articles, 31.7 percent, had received no citations at all, while just 10 articles, 3.8 percent, had reached at least 100 citations. The top five most cited articles were all published in 2023 and focused on report simplification, plain-language translation, structured reporting, and impression generation. The most cited was an exploratory case study on simplified radiology reports published in European Radiology, with 495 citations, followed by work on translating radiology reports into plain language using ChatGPT and GPT-4, with 331 citations, and a multilingual feasibility study on transforming free-text reports into structured reporting, with 307.

Open access is a defining feature of this literature. Of the 265 studies, 207, or 78.1 percent, were available through some form of open access, while 58 remained closed. Gold open access was the most frequent model at 80 studies, followed by hybrid at 55, diamond at 38, green at 26, and bronze at 8. Open access remained predominant across publication years, ranging from 72.1 percent of 2024 publications to 82.8 percent in 2025.

The authors interpret the sharp post-2023 rise as the result of a technical inflection point meeting a pre-existing workflow need. Unlike earlier waves of radiology AI that required image-model development and annotation pipelines, much of the early LLM literature could be built directly on existing report corpora and evaluated against practical tasks such as zero-shot information extraction, report labeling, impression streamlining, and patient-centered simplification. Because these applications are text-based and relatively easy to prototype against retrospective datasets, they created unusually low barriers to early publication. The analysis also notes that the field’s focus has broadened from feasibility toward quality assurance, privacy, and reporting standards, which the authors read as a sign of early consolidation rather than mere novelty-driven expansion.

Important caveats temper the findings. The analysis relied solely on OpenAlex, without parallel searches in PubMed, Scopus, or Web of Science, so the dataset represents the OpenAlex-indexed subset rather than a comprehensive census. Only English-language journal articles were included, excluding conference proceedings, preprints, and non-English publications. Citation counts are inherently time-sensitive in such a young literature, and the authors caution that most included studies are task-specific and retrospective, meaning publication growth should not be equated with clinical readiness. Bibliometric concentration, they note, reflects visibility and output rather than implementation depth or long-term clinical impact. Looking forward, the study highlights the need for multicenter validation studies, prospective workflow evaluations, privacy-aware deployment strategies, and outcome-oriented research demonstrating whether these tools genuinely improve efficiency, report quality, and patient comprehension without introducing clinically silent failure modes.

Subject of Research: Bibliometric mapping of research on large language models in radiology reporting, summarization, and workflow

Article Title: Global research landscape of large language models in radiology reporting, summarization, and workflow: a bibliometric and knowledge-mapping analysis

Article References: Awawdeh, M., & Awawdeh, I. (2026). Global research landscape of large language models in radiology reporting, summarization, and workflow: a bibliometric and knowledge-mapping analysis. Egyptian Journal of Radiology and Nuclear Medicine, 57(1), Article 231. https://doi.org/10.1186/s43055-026-01898-7

Image Credits: AI Generated

DOI: 10.1186/s43055-026-01898-7

Keywords: large language models, radiology reporting, bibliometrics, structured reporting, report summarization, OpenAlex, VOSviewer, open access, citation analysis, radiology workflow, ChatGPT, patient communication