Automated Identification of Different Severity Levels of Diabetic Retinopathy Using a Handheld Fundus Camera and...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness.
• Early detection and treatment are crucial for preventing vision loss.
• Current screening methods can be resource-intensive.

Purpose of the Study:

• To evaluate an AI system in a mobile retinal camera for detecting DR and more-than-mild DR (mtmDR).
• To assess the performance of a single-image protocol for DR screening.

Main Methods:

• A multicenter, cross-sectional diagnostic study involving 327 individuals with diabetes.
• Retinal fundus photographs were captured using a portable camera (Phelcom Eyer).
• Images were analyzed by deep learning algorithms (RAS and DRAS); ground truth was expert grading of 2-field images.

Main Results:

• The AI system demonstrated high sensitivity and specificity for detecting any DR (90.48% sensitivity, 90.65% specificity) and mtmDR (90.23% sensitivity, 85.06% specificity).
• Area under the ROC curve was 0.95 for any DR and 0.89 for mtmDR.
• Performance was evaluated against a rigorous clinical reference standard.

Conclusions:

• The AI-powered portable retinal camera achieves high accuracy in detecting DR severity with a single image per eye.
• This all-in-one solution has significant potential to increase screening coverage rates.
• The technology can contribute to the prevention of avoidable blindness through enhanced DR screening programs.