Facilitating diabetic retinopathy screening using automated retinal image analysis in underresourced settings

Area of Science:

• Ophthalmology
• Medical Imaging
• Diabetology

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in Indigenous Australians.
• Access to timely DR screening is limited in remote Australian primary care settings.
• Automated Retinal Image Analysis (ARIA) offers a potential solution for DR screening.

Purpose of the Study:

• To evaluate the performance of an automated retinal image analysis (ARIA) system.
• To compare ARIA's classification of diabetic retinopathy (DR) status against human grading.
• To assess ARIA's utility for DR screening in Indigenous Australian adults.

Main Methods:

• A cohort of 410 Indigenous Australian adults with type 2 diabetes from remote Northern Territory primary care services participated.
• Retinal fundus photographs were analyzed by both a UK human grader and an ARIA system.
• Sensitivity, specificity, and agreement (Kappa statistic) were calculated for DR classification.

Main Results:

• ARIA demonstrated high agreement with human grading for 'Any DR' (88.0%, Kappa=0.76) and proliferative DR (98.2%, Kappa=0.89).
• The ARIA system achieved 91.4% sensitivity and 85.0% specificity for detecting 'Any DR'.
• For proliferative DR, ARIA showed 96.8% sensitivity and 98.3% specificity.

Conclusions:

• The ARIA software exhibits high sensitivity for detecting diabetic retinopathy, making it a valuable triage tool for human graders.
• ARIA's performance in classifying proliferative DR is also highly sensitive.
• Future ARIA versions incorporating maculopathy and referable DR detection could enhance diabetes care in underserved regions.