Google's DeepMind AI business unit is hoping to teach computers to diagnose eye disease, using patient data from a U.K. hospital.
Using deep learning techniques, DeepMind hopes to improve diagnosis of two eye conditions: Age-related macular degeneration and diabetic retinopathy, both of which can lead to sight loss. If these conditions are detected early enough, patients' sight can be saved.
One way doctors look for signs of these diseases is by examining the interior of the eye, opposite the lens, an area called the fundus. They can do this either directly, with an ophthalmoscope, or by taking a digital fundus scan. Another diagnostic technique is to take a non-invasive three-dimensional scan of the retina using process called optical coherence tomography (OCT).
The project began when a consultant ophthalmologist at the hospital, Pearse Keane, asked DeepMind for help with the time-consuming process of analyzing the scans. If interpretation of the scans could be speeded up with computer assistance, then doctors would be able to treat more patients.
AI researchers are racing to apply machine learning techniques, in which computers figure out for themselves which pieces of data are significant, to medical diagnosis among many other tasks. DeepMind opened its health division in February, and is already working on a mobile app to help doctors and nurses diagnose acute kidney injury.
Meanwhile, researches at IBM have taught their Watson AI software to help doctors diagnose certain kinds of cancer.
DeepMind will process around one million digital fundus and OCT scans from patients of Moorfields Eye Hospital, along with anonymous information about the clinical diagnoses made by hospital staff, the treatment of any eye diseases detected, the model of scanner used and the patients' ages, it said Tuesday.
When researchers use anonymized data and have no way of identifying individual patients, explicit consent from patients is not required, the hospital, said.
Google has plenty of information, including GPS location tracks, online calendars and email records, that it could potentially correlate with, say, the scan file creation dates to perhaps identify patients.
However, DeepMind's agreement with the hospital forbids it from linking the research data with any other dataset. Indeed, the research protocol is so strict that, for now, the researchers are not even allowed to link up successive scans from the same person. They are, however, seeking permission to do just that, in order to study the evolution of disease and the effects of treatment over time.
As with the research already under way, all personally identifiable data will be removed before DeepMind receives the data, the hospital said.