Introduction:
– Ocular fundus abnormalities like diabetic retinopathy (DR), age-related macular degeneration (ARMD), retinal vein occlusion (RVO), and pathological myopia (PM) are major causes of blindness worldwide.
– Early detection is crucial for preventing irreversible vision loss from these conditions.
– AI systems based on fundus imaging like color fundus photography (CFP) have potential for screening in primary care settings.
– This study aimed to investigate the real-world application of an AI-based fundus screening system in a clinical environment and population screening.
Methods:
– 637 CFP images from an ophthalmic clinic and 20,355 images from a physical exam center were collected.
– An AI system trained to detect 7 conditions (normal, ARMD, DR, RVO, referable glaucoma, PM, other abnormalities) using CFP was applied.
– Diagnostic performance was evaluated against clinician consensus as the gold standard.
– Sensitivity, specificity, accuracy, positive/negative predictive values were calculated.
Results:
– The AI system showed high diagnostic performance (sensitivity, specificity, accuracy all >80%) for DR, RVO, and PM.
– Performance was lower for ARMD (sensitivity 92.9%, PPV 44.8%) and referable glaucoma (sensitivity 65.5%, PPV 48.7%).
– Condition percentages were similar between the clinic and screening populations.
Discussion:
– The AI system demonstrated good screening potential for DR, RVO and PM in real-world settings.
– Lower ARMD and glaucoma performance may relate to disease characteristics and imaging limitations.
– Larger training datasets and multimodal imaging could improve the system further.
Conclusion:
– The studied AI-based fundus screening system showed promising clinical utility for early detection of major blinding eye diseases, especially DR, RVO and PM.
– Further refinement is needed to improve screening for ARMD and glaucoma.
In summary, the standard elements of an introduction, methods, results, discussion, and conclusion are covered, presenting the research rationale, approach, key findings, interpretation, and implications. The review highlights the study’s novel real-world evaluation and outlines strengths, limitations, and potential improvements for the AI screening system.
