In-Person Verification of Deep Learning Algorithm for Diabetic Retinopathy Screening Using Different Techniques...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) screening is crucial for preventing vision loss in diabetes patients.
• Conventional fundus examinations are resource-intensive and prone to human error.
• Automated DR detection models offer a potential solution to improve screening efficiency and accessibility.

Purpose of the Study:

• To assess the clinical performance of an automated diabetic retinopathy screening model.
• To evaluate the model's ability to detect referable DR cases in a real-world clinical setting.
• To compare the automated model's performance with gold-standard dilated fundus examinations.

Main Methods:

• Retrospective analysis of fundus photographs (Eidon and Nidek) for DR staging.
• Prospective enrollment of diabetic patients for automated DR detection.
• Comparison of automated detection results against dilated fundus examinations.

Main Results:

• The automated model demonstrated high sensitivity and specificity for detecting referable DR across both image types.
• Sensitivities ranged from 0.86 to 0.93, and specificities from 0.84 to 0.92.
• Eidon 60°-field images showed a more favorable performance in differentiating referable DR.

Conclusions:

• The deep learning algorithm for referable DR detection performs favorably as an alternative to conventional screening.
• Variations in image quality and angle can impact the model's sensitivity and specificity.
• Automated DR screening holds promise for improving healthcare resource allocation in diabetes management.