Deep learning versus human graders for classifying diabetic retinopathy severity in a nationwide screening program

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients.
• Early detection and treatment are crucial to prevent severe vision impairment.
• Automated DR detection using artificial intelligence (AI) shows promise.

Purpose of the Study:

• To validate a deep learning algorithm for DR detection in a large clinical population.
• To compare the algorithm's performance against human graders in a nationwide screening program.
• To assess the algorithm's accuracy in identifying referable DR and diabetic macular edema (DME).

Main Methods:

• Analysis of 25,326 retinal images from Thailand's nationwide DR screening program.
• Deep learning algorithm used for DR severity grading and DME detection.
• Comparison of algorithm performance with grades from a panel of international retinal specialists and human graders.

Main Results:

• The deep learning algorithm demonstrated significantly higher sensitivity (97%) versus human graders (74%) for detecting referable DR (moderate NPDR or worse).
• Specificity was slightly lower for the algorithm (96%) compared to human graders (98%).
• The algorithm showed higher sensitivity for severe NPDR, proliferative DR (PDR), and DME, with a 23% reduction in false negatives.

Conclusions:

• Deep learning algorithms offer a valuable, highly sensitive tool for DR screening.
• AI can significantly improve the detection of referable DR, potentially reducing missed diagnoses.
• The algorithm's performance suggests its utility in large-scale, community-based DR screening programs.