Automated iridocorneal angle classification using a multimodal large language model

To evaluate the diagnostic performance of a general-purpose vision-language model (GPT-4o) in interpreting gonioscopic images of the anterior chamber angle, with a focus on angle configuration classification and Shaffer grading.

A total of 80 slit-lamp goniophotographs representing a range of anterior chamber angle (ACA) configurations were collected from a digital image repository and a specialized textbook. Each image was evaluated by three glaucoma specialists (reference standard), three comprehensive ophthalmologists, and GPT-4o across three trials. GPT-4o was prompted with a standardized diagnostic instruction and shown one example per Shaffer grade prior to evaluation. Performance metrics included accuracy, sensitivity, specificity, Quadratic Weighted Kappa (QWK), and Fleiss' Kappa. Descriptive analyses and confusion matrices were used to assess misclassification patterns.

GPT-4o demonstrated high accuracy in classifying angle configuration (91.3%, 91.3%, and 90.0% across trials) with perfect or near-perfect specificity (> 98%) but moderate sensitivity (56.3%) for detecting closed angles. In Shaffer grading, GPT-4o showed moderate agreement with expert consensus (QWK = 0.471-0.573) and fair inter-trial consistency (Fleiss' Kappa = 0.235). Misclassifications occurred predominantly between adjacent grades, particularly underestimating wide angles. Comprehensive ophthalmologists achieved higher sensitivity and agreement with the reference standard in both classification tasks.

GPT-4o was able to produce structured, clinically relevant assessments of gonioscopic images without domain-specific training. Although outperformed by comprehensive ophthalmologists and previous task-specific AI models, GPT-4o demonstrated consistent performance with potential utility as a supportive tool in settings with limited specialist access. Further refinement and clinical validation are needed to improve its reliability in grading tasks.