Automation of Macular Degeneration Classification in the AREDS Dataset, Using a Novel Neural Network Design

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Age-related macular degeneration (AMD) is a leading cause of vision loss.
• Early detection and risk stratification are crucial for managing AMD progression.
• Current diagnostic methods can be resource-intensive.

Purpose of the Study:

• To develop an ensemble of Convolutional Neural Networks (CNNs) for detecting and stratifying AMD risk.
• To utilize retinal photographs for automated AMD risk assessment.
• To predict the progression of AMD based on the Age-related Eye Disease Study (AREDS) Simplified Severity Scale.

Main Methods:

• Developed three individual CNNs to detect advanced AMD, drusen size, and pigmentary abnormalities.
• Arranged CNNs in a cascading architecture to calculate the AREDS Simplified 5-level risk severity score.
• Created a simplified binary classification of "low risk" and "high risk" for AMD progression.

Main Results:

• The CNN ensemble achieved 80.43% accuracy (quadratic kappa 0.870) against the 5-step AREDS scale.
• For binary classification, the ensemble reached 98.08% accuracy, with sensitivity ≥85% and specificity ≥99%.
• The model demonstrated high performance in classifying AMD risk levels.

Conclusions:

• An ensemble of CNNs accurately calculates the AREDS 5-step Simplified Severity Scale for AMD.
• This AI tool shows potential as a screening method to identify individuals benefiting from supplements.
• Improved health outcomes are possible by identifying asymptomatic individuals at risk for AMD progression.