Multimodal foundation model assistance for differentiating primary open-angle glaucoma in highly myopic eyes

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Differentiating primary open-angle glaucoma (POAG) from non-glaucoma in highly myopic eyes presents diagnostic challenges.
• Artificial intelligence (AI) offers potential to augment diagnostic capabilities in ophthalmology.

Purpose of the Study:

• To develop and evaluate a multimodal foundation-model-assisted system for POAG detection in highly myopic eyes.
• To assess the impact of AI assistance on ophthalmologist diagnostic performance.

Main Methods:

• A retrospective study utilizing a multimodal classifier (RETFound) integrating fundus photographs, OCT, and OCT/RNFL variables.
• Six ophthalmologists evaluated 150 cases twice: once unaided and once with AI assistance.
• Performance metrics included sensitivity, specificity, accuracy, F1 score, confidence, reading time, and inter-reader agreement.

Main Results:

• The multimodal model achieved an area under the receiver operating characteristic curve (AUC) of 0.911.
• AI assistance numerically improved mean sensitivity (68.0% to 75.3%), accuracy (74.2% to 79.6%), and F1 score (0.790 to 0.844).
• Mean specificity remained unchanged (91.3%), while inter-reader agreement (Fleiss kappa) improved from 0.583 to 0.709.

Conclusions:

• AI assistance demonstrated numerically improved diagnostic performance and inter-reader agreement in differentiating POAG in highly myopic eyes.
• Statistically significant differences at the reader level were not observed after multiplicity adjustment and are considered exploratory.
• Prospective, multicenter validation is essential prior to clinical implementation of the AI system.