Artificial Intelligence In Corneal Maps Interpretation: A Comparative Evaluation Of Four Large Language Models

Published 2026 - 44th Congress of the ESCRS

Reference: FP06.07 | Type: Free paper | DOI: 10.82333/5x9r-pb73

Authors: Dr Filippo Lixi* ¹ , Dr Mario Troisi ² , Dr Andrea Taloni ³ , Dr Costanza Rossi ¹ , Dr Claudia Corda ¹ , Dr Valerio Calabresi ¹ , Dr Antonio Budroni ¹ , Dr Camilla Bianchi ¹ , Prof. Dr Giulia Coco ⁴ , Prof. Dr Giuseppe Giannaccare ¹

¹Department of Surgical Sciences,University of Cagliari,Cagliari,Italy, ²Federico II University,Naples,Italy, ³University of Ferrara,Ferrara,Italy, ⁴Department of Clinical Sciences and Translational Medicine,University of Rome Tor Vergata,Roma,Italy

Purpose

To evaluate and compare the diagnostic performance of four multimodal large language models (LLMs) in interpreting refractive corneal maps across a broad spectrum of corneal conditions.

Setting

Single-center, comparative cross-sectional study conducted at the Complex Ophthalmology Unit, University Hospital of Cagliari (Cagliari, Italy).

Methods

De-identified corneal tomography images acquired using the Oculus Pentacam (Oculus Optikgeräte GmbH, Wetzlar, Germany) were analyzed by four multimodal LLMs (ChatGPT-5.2, Gemini 3.0 Pro, Claude Opus 4.6, Grok 4.1) using a standardised prompt. Outputs were compared with reference diagnoses established by corneal specialists. Stage-stratified analysis was performed for keratoconus [Topographic Keratoconus Classification (TKC) grades 1–4].

Results

Diagnostic performance differed significantly across LLMs (p < 0.001). ChatGPT-5.2 achieved the highest diagnostic performance (85.8%), followed by Gemini 3.0 Pro (81.6%), Claude Opus 4.6 (61.5%), and Grok 4.1 (51.4%). ChatGPT-5.2 and Gemini 3.0 Pro were statistically equivalent (p = 0.142). Overall, KC detection did not differ across models (82.8–89.9%; p = 0.392); however, for TKC grade 1, significant differences emerged (p = 0.027), with ChatGPT-5.2 outperforming Gemini 3.0 Pro (83.3% vs. 25.0%). Greater KC severity was associated with higher diagnostic performance for most models. Metric-extraction accuracy was highest for ChatGPT-5.2 (99.7%) and Gemini 3.0 Pro (97.9%), exceeding that of Claude Opus 4.6 (93.8%), and Grok 4.1 (17.4%).

Conclusions

Multimodal LLMs show variable but potentially clinically relevant capability in interpreting refractive corneal maps. ChatGPT-5.2 and Gemini 3.0 Pro achieved superior and equivalent overall accuracy, while ChatGPT-5.2 showed a unique advantage for early KC detection. Grok 4.1 demonstrated near-complete metric-extraction failure. These systems may function as adjunctive decision-support tools, but prospective validation and integration studies are required before routine clinical implementation.