Evaluations of Deep Learning Approaches for Glaucoma Screening Using Retinal Images from Mobile Device

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Early detection through screening is crucial for managing glaucoma.
• Deep learning offers potential for widespread glaucoma screening using retinal images.

Purpose of the Study:

• To compare deep learning classification and segmentation methods for glaucoma screening.
• To evaluate performance using images from both retinography and mobile devices.
• To assess the feasibility of glaucoma screening with low-cost mobile imaging.

Main Methods:

• Utilized Xception, ResNet152 V2, and Inception ResNet V2 for classification.
• Employed U-Net with Inception ResNet V2/V3 backbones for segmentation and cup-to-disc ratio (CDR) estimation.
• Analyzed model activation maps for classification interpretability.

Main Results:

• Deep learning models achieved performance comparable to state-of-the-art methods.
• Classification and segmentation tasks showed similar results between retinography and mobile device images.
• The classification method demonstrated effectiveness on low-quality mobile datasets.

Conclusions:

• Deep learning models are viable for glaucoma screening with both standard and mobile retinal imaging.
• Mobile devices with low-cost lenses can facilitate increased glaucoma screening frequency.
• AI-powered screening using accessible technology can improve early glaucoma detection rates.