AI screening for retinal disease: from algorithm to clinical evidence

Why it matters

Most retinal disease is first seen outside specialist clinics, by optometrists and in community eye care. If AI can flag the right biomarkers on an OCT scan at that point, patients who need a specialist get referred sooner and those who don’t are spared the trip.

Step 1: building the algorithm (ARVO 2025)

I led development of a deep learning algorithm that detects retinal biomarkers on Topcon Maestro OCT scans: RPE loss, intraretinal and subretinal fluid, and pigment epithelial detachment. A built-in image-quality module keeps false positives down on messy real-world data.

• Trained and validated on 11,000 expert-labelled B-scans from the Medical University of Vienna.
• Average F1 score of 0.83 on an independent dataset, and a false-positive rate of 0.02% in 1,000 patients under 35.
• Applied to 81,903 real-world scans from 57,963 patients. Biomarkers were found in 2.1% of patients, rising to 4.7% in those over 50, and were present even in 40–50-year-olds (0.9%), a case for screening younger populations.

Step 2: does it help clinicians? (ARVO 2026)

Building a model is half the job. The other half is showing that it improves decisions. In a crossover reader study, two graders assessed 112 real-world OCT volumes from primary eye care twice: once alone and once with the AI report. An independent clinical expert set the reference.

• Overall detection performance improved by 42% (Cohen’s κ from 0.41 to 0.58, p < 0.01).
• Sensitivity for RPE loss rose from 0.64 to 0.96, and for subretinal hyporeflective spaces from 0.65 to 0.95.
• Graders were 11% faster with AI support.
• Every false negative co-occurred with another detected biomarker, which limits the risk of a missed referral.

Presentation

I presented the reader study in Denver on 3 May 2026, in one of the largest sessions at ARVO 2026.