.png)
Moorfields / Google DeepMind
Healthcare · UK
94% referral accuracy
Moorfields Eye Hospital triages 1,000+ daily eye scans at specialist level with two neural networks
The AI system does not just detect disease. It recommends how urgently a patient should be referred, across more than 50 sight-threatening conditions, at a level matching world-leading ophthalmologists.
Friction
Moorfields received over 1,000 OCT scan referrals per day. 3D retinal scans require specialist interpretation, and delays between scan and treatment can lead to irreversible vision loss. There were not enough specialists to keep up.
Breakthrough
Two neural networks in sequence: the first segments the scan into an interpretable map of retinal tissue features. The second uses that map to generate a diagnosis and referral urgency recommendation, with a confidence percentage attached. The physician sees which image features the model used, not just the outcome.
Impact
94% accuracy in correct referral recommendations for 50+ eye conditions, matching or exceeding world-leading ophthalmologists. Model trained on 14,884 OCT scans. Segmentation layer proved scanner-independent, meaning accuracy held when tested on a different device type.
Unlock the full analysis with breakthrough, impact, what made it smart and its technical approach below!
Problem
OCT scans are detailed 3D images of the retina that require expert interpretation. The volume at Moorfields was growing faster than specialist capacity. Delays meant patients with urgent conditions could lose vision while waiting.
What made it smart
The two-stage architecture separates "what do I see" from "what should happen next." This makes the model explainable: the referral decision can be traced back to specific retinal features, not just a black-box score. That explainability is what makes it defensible in clinical practice.
Technical approach
Neural network 1 segments the raw OCT scan into a tissue map with labelled features. Neural network 2 takes that map as input and produces a diagnosis and referral urgency recommendation across 50+ conditions. The tissue map also acts as a scanner-agnostic intermediate representation, enabling generalisability across hardware.
Strategic lesson
AI in high-stakes decisions earns trust not by being a black box with a good score, but by showing its work and making the reasoning visible to the human who acts on it.
Reflection question
In your organisation, where are expert judgements being made that are currently opaque, time-consuming, or bottlenecked on scarce human capacity?