Of all games that show off human intelligence and imagination, the ancient stone-placing game of Go is the most complex. The world’s highest-ranked chess player, Garry Kasparov, had been defeated by IBM’s Deep Blue back in 1997, and now the top ranking for Go belonged to the circuits of Google’s AlphaGo.

Last week, the scientific publication Nature Medicine reported on a very different new talent for DeepMind. Working for the past 18 months at London’s Moorfields Eye Hospital—the oldest eye hospital in the world—the DeepMind program called DeepMind Health has demonstrated diagnostic skills that match the accuracy of expert eye doctors. The program can now recognize and recommend the correct decision for referral for more than 50 eye diseases with an accuracy of 94%. The program has learned to read and analyze 3-D retinal images produced by optical coherence tomography (OCT) scans.

A coalition of three partners has guided the project, including Moorfields, DeepMind Health, and the University College of London Institute of Ophthalmology. The value of the DeepMind program rests in its ability to do diagnostic screening, not provide direct medical care. Moorfields analyzes more than 1,000 OCT scans every day, and that creates a burden on the human specialists doing the reading—a burden that inevitably causes delays in diagnoses that can exacerbate symptoms.

DeepMind’s role is to process the scans, quickly reading and diagnosing with the same accuracy as most of its human counterparts. The cases that are most serious can be passed on to staff doctors for more immediate attention. Mustafa Suleyman, a cofounder and director of applied AI at DeepMind Health, told the press, “These incredibly exciting results take us one step closer to (faster diagnoses) and could, in time, transform the diagnosis, treatment, and management of patients with sight-threatening eye conditions, not just at Moorfields, but around the world.”


DeepMind’s system isn’t intended to produce a complete diagnostic end point. Instead, it provides diagnoses of conditions and then offers multiple possible explanations along with the confidence levels it has for its conclusions. It isolates those regions that attracted its attention, enabling the doctor to review the systems analysis and its conclusions.

Within its own systems, DeepMind uses a group of algorithms to shape its decisions. Each algorithm is trained separately so a decision about a problem is the result of a consensus, thereby eliminating mistakes possible with single-system failures.

A second unique feature of DeepMind’s system is that it uses two separate networks. A segmentation network converts the OCT raw data into a 3-D tissue map with color-coded segments. A senior clinician scientist for DeepMind explains, “That map doesn’t only describe the layers of the eye, but if there’s a disease in the eye, and where that disease is.”

The second network is a classification network that analyzes the segmentation map and assesses the type and location of any problems and the urgency of the care required.

The AI system was trained on nearly 15,000 OCT scans from approximately 7,500 Moorfields patients. The image below shows a sample DeepMind Health diagnosis. Specific regions are mapped in the images, and the top left of the display has the recommendations and confidence levels of those decisions.

Image: UCL, Moorfields, DeepMind, et al.

The AI system currently being tested at Moorfields is a research project, and DeepMind Health is looking forward to official approval for use in a clinical setting. At that point, the company will provide the system free of charge to Moorfields’ doctors for a period of five years. Once approved, though, DeepMind will be able to sell the technology to other hospitals in the United Kingdom and elsewhere.


The Moorfields project has brought attention to various problems of privacy in the public health sector. The scans analyzed today by the DeepMind system are anonymized before the files are put into the system so patients’ identities are protected. Moorfields controls the database produced by the work with those scans, and the hospital uses the data in nine other research studies in which it is involved. There are several outlets, including the hospital and DeepMind’s websites, currently informing the public about the projects and how the data is being used and by whom.

On its part, DeepMind has created a new research division, DeepMind Ethics & Society, to more efficiently address the various problems that can accompany AI research in general. It explains on the division’s web pages:

“We created DeepMind Ethics & Society because we believe AI can be of extraordinary benefit to the world, but only if held to the highest ethical standards. Technology is not value neutral, and technologists must take responsibility for the ethical and social impact of their work. In a field as complex as AI this is easier said than done, which is why we are committed to deep research into ethical and social questions, the inclusion of many voices, and ongoing reflection.”

Visit the DeepMind Ethics and Society website to get an idea of what DeepMind sees as the key ethical challenges of AI and read its list of five core principles.

About the Authors