Detecting Eye Disease Using Vision Transformers Informed by Ophthalmology Resident Gaze Data

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Computer Vision

Background:

• Glaucoma detection accuracy is critical for preventing vision loss.
• Current AI models like the Vision Transformer (ViT) show promise but can be further optimized.
• Integrating human expertise, such as ophthalmologist gaze patterns, may enhance AI performance.

Purpose of the Study:

• To develop and evaluate two novel approaches for enhancing the Vision Transformer (ViT) model using ophthalmology resident gaze data.
• To assess the impact of gaze-informed training on the accuracy and computational efficiency of ViT for glaucoma detection.
• To explore a new paradigm for human-AI collaboration in ophthalmic diagnostics.

Main Methods:

• Two proof-of-concept models were developed: Fixation-Order-Informed ViT and Ophthalmologist-Gaze-Augmented ViT.
• These models were trained by integrating ophthalmology resident gaze data.
• The performance of the gaze-informed ViTs was compared against the standard ViT for glaucoma detection.

Main Results:

• The Fixation-Order-Informed ViT and Ophthalmologist-Gaze-Augmented ViT demonstrated superior accuracy compared to the standard ViT.
• The enhanced ViT models also exhibited improved computational efficiency.
• The results indicate that incorporating gaze data significantly benefits AI performance in glaucoma detection.

Conclusions:

• Integrating ophthalmology resident gaze data enhances ViT model performance for glaucoma detection.
• Gaze-informed ViTs offer a pathway towards more accurate and computationally efficient AI diagnostic tools.
• This approach establishes a new paradigm for direct medical expert-AI interfacing, fostering more interpretable AI teammates in clinical settings.