EUVIMEDEUVIMED entdecken
Crossref

Feasibility of a multimodal large language model in interpreting plain radiographs of bone tumors: a pilot study

Samet Genez, Hamza Özer

Journal of Radiology in Medicine · 09.03.2026 · 3 (1) · 5-8

DOI: 10.51271/jrm-0042 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Radiology in Medicine, online erschienen am 09.03.2026. Verfasst von Samet Genez und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Aims

To evaluate the diagnostic accuracy and clinical feasibility of a multimodal large language model (ChatGPT-5) in interpreting plain radiographs of bone tumors and differentiating between benign and malignant lesions.

Methods

This retrospective pilot study utilized 50 verified bone tumor cases (27 benign and 23 malignant) sourced from the Radiopaedia database. Anonymized radiographs were processed by ChatGPT-5 using a standardized zero-shot prompt in independent sessions to prevent contextual bias. Model performance was assessed based on the accuracy of the most likely diagnosis, the inclusion of correct diagnoses within the top three differentials, and benign–malignant classification metrics. Statistical analysis included the Clopper–Pearson binomial method for confidence intervals and McNemar’s exact test to evaluate improvements in diagnostic accuracy and potential systematic error asymmetry.

Results

The model achieved 100% accuracy in identifying the imaging modality and the affected bone. The accuracy for the single most likely diagnosis was 56.0% (95% CI: 41.3–70.0), which significantly increased to 70.0% (95% CI: 55.4–82.1) when two differential diagnoses were included (p=0.016). For benign–malignant classification, the model demonstrated an overall accuracy of 76.0%, with a high specificity of 96.3% but a notably limited sensitivity for malignancy at 52.2%. A statistically significant error asymmetry indicated a systematic tendency toward benign classification (p=0.006).

Conclusion

While ChatGPT-5 demonstrates proficiency in foundational radiographic identification, its low sensitivity for malignancy remains a critical limitation for independent clinical use. The results suggest that while multimodal LLMs may serve as promising educational or triage aids, they currently require rigorous human expert oversight to maintain diagnostic safety in image interpretation.

Zitieren

APA

Genez, S., & Özer, H. (2026). Feasibility of a multimodal large language model in interpreting plain radiographs of bone tumors: a pilot study. Journal of Radiology in Medicine, 3(1), 5-8. https://doi.org/10.51271/jrm-0042

Vancouver

Genez S, Özer H. Feasibility of a multimodal large language model in interpreting plain radiographs of bone tumors: a pilot study. Journal of Radiology in Medicine. 2026;3(1):5-8. doi:10.51271/jrm-0042

BibTeX
@article{Genez2026,
  author = {Genez, Samet and Özer, Hamza},
  title = {Feasibility of a multimodal large language model in interpreting plain radiographs of bone tumors: a pilot study},
  journal = {Journal of Radiology in Medicine},
  year = {2026},
  volume = {3},
  number = {1},
  pages = {5--8},
  doi = {10.51271/jrm-0042}
}
Kostenlos starten →