EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Multimodal large language models for IOL power calculation in cataract surgery: A feasibility study

Hye Won Jun, Sun Young Ryu, Tae Keun Yoo

AJO International · 2025-12 · 2 (4) · 100205

DOI: 10.1016/j.ajoint.2025.100205 ↗

Kurz gefasst

Fachartikel in der Zeitschrift AJO International, 2025-12. Verfasst von Hye Won Jun und 2 weiteren Autor:innen. Die Arbeit wurde bisher 7-mal zitiert (Crossref) und nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

• Multimodal LLMs reproduced SRK/T using identical IOLMaster 700 biometry (180 eyes). • ChatGPT-5 showed near-reference agreement with SRK/T; others had larger errors. • Divergence was greatest in very long eyes and shallow anterior chambers, reflecting known limitations of SRK/T in extreme biometry. • Multimodal LLMs may serve as a portable, stepwise backup and educational secondary check on IOL calculations, but they should not replace validated biometry platforms. To assess whether multimodal large language models (MLLMs) can reproduce classical intraocular lens (IOL) power calculations and provide usable backup guidance when standard calculators are unavailable or for educational purposes. Methods-comparison feasibility study. From a public IOLMaster 700 dataset, 180 eyes were sampled and stratified by axial length and anterior chamber depth. IOL powers were computed by device Sanders–Retzlaff–Kraff/Theoretical (SRK/T) (reference), Barrett Universal II (APACRS), and three MLLMs (ChatGPT-5, Gemini-2.5-Pro, ChatGPT-4o) using identical biometry. Agreement metrics included mean absolute error (MAE) and the proportion within ±0.25, ±0.50, and ±1.00 D versus device SRK/T. ChatGPT-5 and Gemini-2.5-Pro implemented SRK/T using effective lens position estimation and a vergence approach, whereas ChatGPT-4o defaulted to SRK I unless tightly constrained. ChatGPT-5 showed near-reference agreement with SRK/T (MAE 0.30 D; 78.8% within ±0.50 D vs SRK/T), while Gemini-2.5-Pro and ChatGPT-4o had larger errors. Subgroup analyses across axial length–anterior chamber depth strata showed that ChatGPT-5 did not differ significantly from SRK/T after Bonferroni correction, whereas Gemini-2.5-Pro and ChatGPT-4o exhibited significant positive biases in long and short eyes. Agreement patterns were similar when Barrett Universal II was used as the secondary comparator. MLLMs, particularly ChatGPT-5, may provide a portable, stepwise backup for classical IOL power calculation (SRK/T). MLLMs may offer educational value and serve as a secondary check on conventional calculations but should not replace validated biometry platforms, especially in eyes with extreme biometry. Prospective, outcome-based validation and usability testing are warranted before any real-world clinical deployment.

Zitieren

APA

Jun, H. W., Ryu, S. Y., & Yoo, T. K. (2025). Multimodal large language models for IOL power calculation in cataract surgery: A feasibility study. AJO International, 2(4), 100205. https://doi.org/10.1016/j.ajoint.2025.100205

Vancouver

Jun HW, Ryu SY, Yoo TK. Multimodal large language models for IOL power calculation in cataract surgery: A feasibility study. AJO International. 2025;2(4):100205. doi:10.1016/j.ajoint.2025.100205

BibTeX
@article{Jun2025,
  author = {Jun, Hye Won and Ryu, Sun Young and Yoo, Tae Keun},
  title = {Multimodal large language models for IOL power calculation in cataract surgery: A feasibility study},
  journal = {AJO International},
  year = {2025},
  volume = {2},
  number = {4},
  pages = {100205},
  doi = {10.1016/j.ajoint.2025.100205}
}
Kostenlos starten →