AI-Assisted Detection of Macular OCT Abnormalities by Optometrists: A Retrospective Reader Study

Optical coherence tomography (OCT) is a key imaging modality for diagnosing retinal disease, but accurate interpretation may require specialist expertise. Artificial intelligence (AI)-assisted OCT has the potential to enhance detection of retinal abnormalities in clinical practice. However, evidence on how AI decision-support tools influence optometrist diagnostic performance in real-world OCT interpretation remains limited.

This study aimed to evaluate the impact of an AI-integrated OCT interpretation tool, CIRRUS® PathFinder™ (PF), on optometrists' ability to detect retinal abnormalities on macular OCT scans compared with an expert ophthalmologist reference standard.

This retrospective diagnostic accuracy reader study analyzed 200 anonymized macular OCT B-scans from 200 patients at a retinal clinic in Singapore. All scans met pre-specified signal strength criteria. Two optometrists independently classified scans as normal or abnormal under unaided and PF-assisted conditions, and classifications were then compared with an expert ophthalmologist reference standard. Diagnostic performance metrics, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy, were calculated with 95% confidence intervals (CI), and agreement with the expert grader was assessed using Cohen's kappa (κ). Changes in paired classifications were evaluated using McNemar's exact test.

Of the 200 scans, 73 (36.5%) were classified as abnormal by the expert. PF assistance increased abnormality detection and substantially reduced false negative classifications. Sensitivity improved from 61.6% (95% CI 49.5-72.8) to 98.6% (95% CI 92.6-100.0) for Optometrist A (p<2.0x10^-8) and from 87.7% (95% CI 77.9-94.2) to 95.9% (95% CI 88.5-99.1) for Optometrist B (p=0.031). Agreement with expert grading increased from moderate to substantial for Optometrist A (κ=0.528 to 0.729) and remained almost perfect for Optometrist B (κ=0.817 to 0.842). Overall diagnostic accuracy improved from 79.0% to 86.5% and from 91.5% to 92.5%, respectively, with modest reductions in specificity.

AI-assisted OCT interpretation with PF improved optometrists' sensitivity for detecting expert-confirmed macular OCT abnormalities, primarily by reducing false-negative classifications. Further prospective studies in broader clinical settings are needed to determine the impact on referral appropriateness, workflow efficiency, reader behavior and patient outcomes.