Eye scans: an AI that recommends where to send the patient

Why do eye scans pile up unread?

The number of medical imaging tests is growing faster than the number of specialists able to read them. One of these tests, the OCT scan (optical coherence tomography), makes a three-dimensional image of the retina, the tissue at the back of the eye that catches light.

The challenge is sorting: knowing which patient must see a specialist urgently, and which can wait or only needs monitoring.

Without help, reading times lengthen, and a patient whose sight is at risk may wait as long as one with nothing serious.

How did the team build its AI?

Until now, medical AI had mostly proved itself on two-dimensional photos, using collections of millions of labelled images. Reaching the level of expert clinicians on three-dimensional scans, in a real care pathway, remained unsolved.

The researchers designed a new deep learning architecture. It works in two steps: first it draws a map of the eye's tissues (a segmentation), then it uses that map to recommend a referral. It was tested on scans from patients referred to a major eye hospital.

The scans came from patients referred to a major eye hospital, so they reflect the mix of cases a real eye clinic sees, not a tidy test collection.

What did they find?

After training on 14,884 scans, the program reaches or exceeds experts at recommending patient referral, across a range of sight-threatening retinal diseases.

Another key result: the tissue map it produces does not depend on the brand of machine. Referral accuracy is maintained when segmentations from a different type of device are used. The tool therefore does not need retraining for every new piece of equipment.

The program was tested on scans from patients referred to a major eye hospital, a varied and realistic population, rather than on hand-picked cases.

What does this change for patients?

The authors say their work removes barriers to wider clinical use, without needing huge training collections and across several diseases at once. For a patient, the hope is faster and more consistent sorting of scans, so urgent cases see a specialist sooner.

This is a referral aid. Diagnosis and the choice of treatment stay in the hands of the eye doctor.

Another benefit: since the tissue map does not depend on the machine's brand, a hospital that changes equipment would not have to start from scratch. This is an important practical point for moving from research to daily use.

What are the limits?

The study compares the AI's recommendations with those of experts on scans already taken. It does not show that patients kept their sight better by using the tool.

The data come from one particular hospital. These results will need confirming in other centres, with other populations, before anyone can talk about routine use.