Towards a Device Agnostic AI for Diabetic Retinopathy Screening: An External Validation Study

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients.
• Early detection and screening are crucial for preventing vision impairment.
• Artificial intelligence (AI) offers potential for automated screening of retinal images.

Purpose of the Study:

• To assess the performance of a validated AI algorithm, originally developed for smartphone cameras, on images acquired with a standard desktop fundus camera for diabetic retinopathy screening.

Main Methods:

• The AI algorithm analyzed 233 eyes from 135 patients with diabetes using desktop fundus images.
• Performance was evaluated for detecting any DR, referable DR (RDR), and sight-threatening DR (STDR).
• AI performance was compared against consensus grading by specialists and clinical examination.

Main Results:

• High sensitivity (98.3%) and specificity (83.7%) for RDR detection against image grading.
• Sensitivity for any DR detection was 97.6% against both image grading and clinical examination.
• Sensitivity for STDR detection was 99.0% against image grading and 100% against clinical examination.

Conclusions:

• The AI algorithm shows robust performance in screening for referable DR and any DR using desktop fundus camera images.
• The AI algorithm's effectiveness extends beyond its original smartphone camera optimization.
• This suggests broader applicability of AI tools in diabetic retinopathy screening programs.