DeepMind's eye study shows medical AI pays for data with product access

DeepMind matched retina specialists on more than 50 eye diseases and paid Moorfields in free use, not cash. Clinical data now buys a stake in the product.
A deep learning system trained on 14,884 eye scans now recommends the right referral for more than 50 sight-threatening diseases as well as retina specialists do. DeepMind and Moorfields Eye Hospital published the peer-reviewed result on 13 August: on 1,000 test scans the system made the correct referral 94% of the time and matched or beat eight experts, four retina specialists and four specialist optometrists, without missing a single serious case.
The science is the smaller story. The contract around it shows how medical AI will pay for its scarcest input. Moorfields receives no cash. If clinical trials validate the system, its clinicians use it free across all 30 of its UK hospitals and community clinics for an initial five years, and the hospital keeps the cleaned, labelled database as its own asset. Clinical data is now bought with a share of the product, and that sets the price every machine learning company in medicine will face when it asks a hospital for images.
Volume was never the bottleneck. Image classifiers for medicine have usually needed millions of annotated pictures; this one needed fewer than 15,000 because the labels were careful and the architecture split the task in two. One network maps tissue and disease features, a second reads that map and recommends a referral. Because the map looks the same whatever machine took the scan, a hospital with a different scanner needs only the first network retrained. The software therefore travels across hardware makers, while the curation stays with whoever did the slow work of cleaning a hospital archive.
The money sits in the queue. Moorfields receives more than 1,000 scan referrals a day across clinics that serve 300,000 patients a year. In 2016 it received about 7,000 urgent referrals for suspected wet macular degeneration, and roughly 800 of those patients had the disease. Every false alarm costs a specialist appointment, so a system that sorts referrals reliably is priced in clinician hours saved, not in diagnoses made. Founders should build unit economics on that hour and treat the data agreement as part of the product, with consent and governance settled before training starts. The cost of skipping that step is on record: in July 2017 British regulators found that the Royal Free hospital broke data protection law when it shared 1.6 million patient records with DeepMind for a kidney app.
Against this stands the free licence itself: the largest early customer pays nothing, and the product does not yet exist. It still needs prospective trials and regulatory approval, which can take years. The signal to watch is a regulator's clearance, of the kind IDx won in April for diabetic eye screening, and the first hospital outside the partnership that pays. Fanaura looks for medical AI teams whose plan already names both, and our venture thesis favours companies that own the path to a cleared product, a point we raise early in how we work with founders.
Training took 14,884 scans. The business will take a hospital willing to be paid in software.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building machine learning for clinical imaging with a clear path to regulatory approval? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.