Glaucoma Detection and Feature Identification via GPT-4V Fundus Image Analysis

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Early detection and accurate diagnosis are crucial for effective management.
• Fundus image analysis is a key component in glaucoma diagnosis.

Purpose of the Study:

• To evaluate the diagnostic accuracy of GPT-4V (OpenAI) in identifying glaucoma.
• To assess GPT-4V's capability in detecting glaucoma-related features compared to expert ophthalmologists.
• To determine the consistency and agreement of GPT-4V's assessments.

Main Methods:

• 300 fundus images from public datasets (ACRIMA, ORIGA, RIM-One v3) were analyzed.
• GPT-4V assessed images for quality, gradability, cup-to-disc ratio, and glaucoma status.
• GPT-4V's assessments were compared against two independent expert graders and dataset labels.

Main Results:

• GPT-4V achieved overall accuracy slightly lower than expert graders across datasets (ranging from 0.68 to 0.81).
• Agreement between GPT-4V and experts varied, with Cohen kappa values from 0.08 to 0.72.
• GPT-4V demonstrated high consistency in image gradability (≥89%) and substantial agreement in rim thinning and cup-to-disc ratio assessments.

Conclusions:

• GPT-4V shows promise as a supplementary tool for glaucoma screening and detection via fundus image analysis.
• The model demonstrated good performance in identifying key glaucomatous features.
• Agreement varied across datasets, indicating a need for further refinement and validation.