• Home  
  • Retinal Image AI Flagged Uncertain Cases for Specialist Review in Tests
- Research Briefs

Retinal Image AI Flagged Uncertain Cases for Specialist Review in Tests

A retinal image AI model deferred uncertain predictions, with 91.3% accuracy among non-deferred cases in a simulated referral test.

An experimental diabetic retinopathy grading model estimated uncertainty alongside disease severity. In a simulated referral test, it achieved 91.3% accuracy among cases that were not deferred for specialist review.

What researchers studied

Researchers developed EviProto-DR, a system that uses ordered disease-grade representations and evidential uncertainty estimates when interpreting retinal photographs. It was evaluated on image datasets including APTOS 2019 and DDR, with an external assessment on MESSIDOR-2. The team examined grade agreement, confidence calibration and simulated referrals. These were retrospective computational tests rather than a trial involving patients and clinicians in a live screening service.

What they found

The model achieved quadratic weighted kappa of 0.918 on APTOS 2019 and 0.897 on DDR, indicating agreement that accounts for differences in disease severity. Expected calibration error on APTOS was 0.061. In the simulated referral workflow, accuracy among retained, non-deferred predictions was 91.3%. Importantly, this percentage excludes images marked uncertain and sent for review, so it is not accuracy across the full original set.

What it means

Automated screening systems may be more useful when they can identify predictions that deserve human scrutiny. This work explores an approach to pairing classification with a referral signal rather than producing a confident answer in every case. Actual value would depend on the chosen uncertainty threshold, available specialists and patient outcomes. The reported benchmark does not establish clinical benefit or justify independent deployment.

Important limitation

Retrospective datasets may not capture variation in cameras, populations, image quality or disease prevalence encountered in practice. The referral analysis was simulated, not prospectively validated in a clinic. Additional clinical studies would be required to assess safety, workflow consequences and whether the system improves care. No patient outcome improvement was demonstrated.

Source Information

Study: Ordinal prototype learning with evidential uncertainty for uncertainty-aware diabetic retinopathy grading

Authors: M. Kirubakaran and V. Vijayarajan

Journal / Source: Frontiers in Artificial Intelligence

Published: 19 November 2026

DOI: 10.3389/frai.2026.1941228

Contact Us

Research Today is a South African digital publication that makes credible research easier to understand.

TERMS OF USE & PRIVACY POLICY

follow us