Offline computer-aided diagnosis for Glaucoma detection using fundus images targeted at mobile devices

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness, often asymptomatic in early stages.
• Timely diagnosis and treatment are crucial for slowing glaucoma progression.
• Conventional diagnostic methods are expensive, require expert interpretation, and are not scalable for widespread screening.

Purpose of the Study:

• To develop an interpretable computer-aided diagnosis (CAD) pipeline for glaucoma detection.
• To enable glaucoma diagnosis using fundus images on mobile devices for offline use.
• To create an affordable and accessible tool for early glaucoma screening.

Main Methods:

• A pipeline integrating Convolutional Neural Networks (CNNs) for segmentation and classification of fundus images was developed.
• The pipeline performs optic disc and optic cup segmentation and glaucoma classification.
• The system was optimized for mobile devices, assessing time and space complexity for offline functionality.

Main Results:

• Achieved 0.91 Intersection over Union (IoU) for optic disc segmentation and 0.75 IoU for optic cup segmentation.
• Attained 0.87 accuracy, 0.85 sensitivity, and 0.93 Area Under the Curve (AUC) for glaucoma classification.
• The pipeline operates in under two seconds on average Android smartphones.

Conclusions:

• The developed pipeline demonstrates significant potential for early glaucoma diagnosis.
• The system offers comparable or superior performance to existing CAD systems while operating on resource-constrained devices.
• This technology can facilitate the development of accurate, affordable CAD systems for large-scale glaucoma screenings.