Machine learning outperformed human graders on eye images
For left eyes, a score measuring how well estimates distinguished glaucoma status was 88% for machine learning versus 79% for human graders.
Stake: Cancer Research UK funded the study.
Edited by Boris Djordjevic · AgingResearch.News
A cross-sectional observational study compared eye images from 6,304 EPIC-Norfolk Eye Study participants, whose mean age was 68; 57% were women. Machine learning and trained human graders independently estimated the relative size of a central hollow where the optic nerve leaves the eye. Specialists assessed glaucoma status. Of the participants, 696 had glaucoma or suspected glaucoma in at least one eye.
The researchers compared how well the image-based estimates distinguished glaucoma status. Machine learning scored 88% versus 79% for human graders in left eyes, and 90% versus 81% in right eyes. These were ranking scores, not percentages of correct diagnoses. The authors concluded that the results offered promise for machine-learning-supported population screening.
Why it matters
Glaucoma is an age-related eye disease. The study addressed whether automated image assessment could support screening in a population-based setting.
Caveats
This cross-sectional comparison assessed classification performance, not whether screening improved health. It included only participants whose images from both eyes were usable by both humans and machine learning.
- Machine learning
- Deep learning
- Glaucoma
- Human-derived vertical cup-disc ratio
- Machine learning-derived vertical cup-disc ratio
- Humans
The paper
Moorfields Eye Hospital NHS Foundation Trust
Ophthalmology, Aug 10, 2026, Open access
Comments
Be the first to comment.