CoChat
Sign Up
PreprintCited by 8Artificial Intelligence in Healthcare and EducationMultimodal Machine Learning ApplicationsRadiology practices and education

“Large language models in radiology reporting—a systematic review of performance, limitations, and clinical implications”

Open access
Artsi, Yaara, Klang, Eyal, Collins, Jeremy D., Glicksberg, Benjamin S., Korfiatis, Panagiotis, Nadkarni, Girish N, Sorin, Vera·2025·medRxiv
DOI 10.1101/2025.03.18.25324193
Open this paper in CoChat — free accountOpen PDF

Abstract

Abstract Background Large language models (LLMs) have emerged as potential tools for automated radiology reporting. However, concerns regarding their fidelity, reliability, and clinical applicability remain. This systematic review examines the current literature on LLM-generated radiology reports. Methods We conducted a systematic search of MEDLINE, Google Scholar, Scopus, and Web of Science to identify studies published between January 2015 and February 2025. Studies evaluating LLM-generated radiology reports were included. The study follows PRISMA guidelines. Risk of bias was assessed using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool. Results Nine studies met the inclusion criteria. Of these, six evaluated full radiology reports, while three focused on impression generation. Six studies assessed base LLMs, and three evaluated fine-tuned models. Fine-tuned models demonstrated better alignment with expert evaluations and achieved higher performance on natural language processing metrics compared to base models. All LLMs showed hallucinations, misdiagnoses, and inconsistencies. Conclusion LLMs show promise in radiology reporting. However, limitations in diagnostic accuracy and hallucinations necessitate human oversight. Future research should focus on improving evaluation frameworks, incorporating diverse datasets, and prospectively validating AI-generated reports in clinical workflows.

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1101/2025.03.18.25324193

Cites

29 of 29 references verified0 citation errors · 0 not checkable

Statistics

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Key barriers to translation include the generation of clinically plausible but factually incorrect statements (hallucinations) and the lack of structured frameworks that decompose generation into auditable, verifiable steps amenable to clinical governance.

    “Key barriers include the generation of clinically plausible but factually incorrect statements—so-called hallucinations (11, 12)—and the lack of structured frameworks that decompose generation into auditable, verifiable steps amenable to clinical governance.”

    From Swin-Qwen3: a three-stage vision–language framework for automated radiology report generation with multi-agent verification (2026) · cites this paper as [11]

    Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

    Experimental — accuracy not yet measured

  • Many existing vision-language models rely on simplistic feature concatenation strategies that fail to capture fine-grained correspondences between localised radiographic findings and their semantic descriptions.

    “these models often rely on simplistic feature concatenation strategies that fail to capture fine-grained correspondences between localised radiographic findings and their semantic descriptions (11, 16).”

    From Swin-Qwen3: a three-stage vision–language framework for automated radiology report generation with multi-agent verification (2026) · cites this paper as [11]

    The Discussion addresses input quality and limitations in radiology reporting, but not feature concatenation or fine-grained image–text correspondence. Study Selection and Characteristics describes one model combining regional image features with knowledge-graph features; it does not say this was simplistic concatenation or that it failed to link localised findings to descriptions. Searches for those terms found no supporting passage. Page numbers were unavailable in our copy. (LLM, conf=0.75)

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Cites

Flagged first, then by citations

Research works printed but not linked: 3

  1. Zhang T, Kishore V, Wu F, Weinberger KQ, Artzi Y. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:190409675. 2019.
  2. Zhang Y, Wu J, Li Y, et al. Leveraging professional radiologists' expertise to enhance LLMs' evaluation for AI-generated radiology reports. In: Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE; 2023:1456-1465. doi:10.1109/ICCV.2023.00156. · DOI
  3. Yu F, Endo M, Krishnan R, et al. Evaluating progress in automatic chest X-ray radiology report generation. Patterns (N Y). 2023;4(9):100802. Published 2023 Aug 3. doi:10.1016/j.patter.2023.100802 · DOI · Paper page

Cited by

Sorted by citations

Related

Authors
Artsi, Y., Klang, E., Collins, J. D., Glicksberg, B. S., Korfiatis, P., Nadkarni, G. N., & Sorin, V.
Affiliations
  • Bar-Ilan University (IL)
  • Mount Sinai Hospital (US)
  • Mayo Clinic (US)
Corresponding author
Artsi, Yaara
ORCID
Venue
medRxiv
Publisher
openRxiv
Published
2025-03-19
DOI
10.1101/2025.03.18.25324193
OpenAlex
W4408648898
Cited by
7
References
26
Citations by year
2026: 4 · 2025: 3
Type
posted content
Language
en
Open access
Licence
www.medrxiv.org/about/FAQ#license (vor)
Locations
medRxiv (acceptedVersion)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › Radiology practices and education
Keywords
large language models, radiology reporting, automated medical report generation, hallucinations, fine-tuned models, diagnostic accuracy, human oversight

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review