Performance of General-Purpose Vision Language Models and Ophthalmology Foundation Models in Glaucoma Detection and...

Area of Science:

• Ophthalmology and Artificial Intelligence
• Medical Imaging Analysis
• Computer Vision in Healthcare

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Accurate detection and monitoring of glaucoma progression are crucial for effective management.
• Optical coherence tomography (OCT) provides high-resolution cross-sectional images of the optic nerve head and retinal nerve fiber layer.

Purpose of the Study:

• To evaluate the performance of vision-language models (VLMs) for glaucoma detection.
• To assess the capability of VLMs in predicting visual field (VF) mean deviation (MD) using OCT images.
• To compare the efficacy of fine-tuned VLMs against specialized foundation models and traditional methods.

Main Methods:

• Utilized 27,610 SPECTRALIS OCT images from 1025 participants across DIGS and ADAGES studies.
• Fine-tuned LLaVA, PaliGemma, RETFound, and ResNet-50 models for glaucoma classification and VF MD prediction.
• Trained models on OCT circle scans centered on the optic nerve head, evaluating performance using AUC and MAE metrics.

Main Results:

• The fine-tuned LLaVA model achieved the highest performance: AUC of 0.92 for glaucoma classification and MAE of 1.79 dB for VF MD prediction.
• RETFound and PaliGemma also demonstrated strong performance with AUCs of 0.91 and 0.90, respectively.
• Frozen vision encoders resulted in reduced accuracy; stratified analysis indicated better performance in older individuals and moderate-to-advanced glaucoma cases for classification.

Conclusions:

• Fine-tuned VLMs exhibit high performance in glaucoma detection and VF MD prediction, comparable to or exceeding specialized models.
• These general-purpose AI models can be effectively adapted for clinical glaucoma care.
• VLMs show significant potential for scalable decision support systems utilizing OCT imaging.