Related Experiment Video
Updated: Jun 28, 2026
09:44
Recording and Analyzing Multimodal Large-Scale Neuronal Ensemble Dynamics on CMOS-Integrated High-Density Microelectrode Array
Published on: March 8, 2024
Automated iridocorneal angle classification using a multimodal large language model
Giovanni Rubegni1, Dario Romano2, Benedetta Colizzi2
1Ophthalmology Unit, Department of Medicine, Surgery and Neurosciences, University of Siena, Siena, Italy. giovannirubegni@gmail.com.
Summary
A general-purpose AI model, GPT-4o, showed high accuracy in classifying anterior chamber angle configurations from gonioscopic images but had moderate sensitivity for detecting closed angles. While useful as a supportive tool, it was outperformed by human experts.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Gonioscopy is crucial for diagnosing glaucoma by assessing anterior chamber angle (ACA) configuration.
- Accurate ACA grading is essential for timely treatment and management of glaucoma.
Purpose of the Study:
- To evaluate the diagnostic performance of GPT-4o, a vision-language model, in interpreting gonioscopic images.
- To assess GPT-4o's ability in classifying ACA configuration and Shaffer grading.
Main Methods:
- 80 gonioscopic images were analyzed by glaucoma specialists, ophthalmologists, and GPT-4o.
- GPT-4o received standardized instructions and one example per Shaffer grade.
- Performance was measured using accuracy, sensitivity, specificity, and Kappa statistics.
Main Results:
- GPT-4o achieved high accuracy (90-91.3%) in angle configuration classification but moderate sensitivity (56.3%) for closed angles.
- Shaffer grading showed moderate agreement (QWK 0.471-0.573) with experts, with predominant misclassification between adjacent grades.
- Comprehensive ophthalmologists demonstrated superior performance compared to GPT-4o.
Conclusions:
- GPT-4o can generate clinically relevant assessments of gonioscopic images without specific training.
- While not matching expert performance, GPT-4o shows potential as a supportive tool in resource-limited settings.
- Further validation is required to enhance its reliability for grading tasks.
