On August 13, 2018, Nature Medicine published “Clinically applicable deep learning for diagnosis and referral in retinal disease,” the result of a research partnership between DeepMind and Moorfields Eye Hospital in London that had begun in 2016. The system read optical coherence tomography (OCT) scans - three-dimensional images of the back of the eye that are detailed but hard to interpret and normally require an expert - and recommended how patients should be referred for treatment.
The headline finding was that the AI could correctly recommend referrals for over 50 sight-threatening eye conditions, including age-related macular degeneration and diabetic eye disease, as accurately as world-leading ophthalmologists and optometrists. It was benchmarked against the real, de-identified referral decisions used at Moorfields, so the comparison was against actual clinical practice rather than a curated test set.
The design addressed two practical obstacles to deploying medical AI. First, the system was built in two stages - one network segmented the scan into tissue types, a second made the referral decision - so it could be retrained for different OCT scanners without relabeling everything, a step toward generalizing beyond a single device. Second, it surfaced the segmentation map and a confidence score, letting clinicians see what the model was reacting to rather than receiving an opaque verdict, a partial answer to the “black box” objection that dogs medical AI.
The study mattered because it took a deep-learning model the full distance toward clinical relevance: real-world data, expert-level performance on a broad range of conditions, and a workflow that fit how an eye clinic actually triages patients. It became one of the most-cited demonstrations that AI could match specialists on a complex diagnostic task.