How Google’s Deep Mind Uses Artificial Intelligence to Speed Diagnosis and Treatment of Eye Diseases
Technological advances in eye research are resulting in more OCT scans than there are experts to read them, a problem that may be solved by a ground-breaking AI system.
by Todd Farley
The fields of ophthalmology and optometry have benefited greatly from OCT technology, which produces a detailed, three-dimensional map of the back of a patient’s eye. Because of the ease of producing these scans, however, there are more scans to read than experts to read them. For example, Moorfields Eye Hospital in the United Kingdom produces more than 1,000 OCT scans a day. A solution to that problem may be close at hand, as early indications imply that new artificial-intelligence (AI) technologies produced by Google’s Deep Mind are as accurate as human experts when it comes to interpreting all those OCT scans.
The research, published online in the journal Nature Medicine, shows that an AI system trained by reading more than 14,000 annotated OCT images is as precise as a human when it comes to interpreting scans from clinical practices. In addition, for more than 50 different eye diseases, the AI system was able to correctly recommend when and how people should be referred to treatment. Perhaps even more importantly, after the AI system quickly identifies the eye disease in question, its “instant triaging process” means it can recommend to eye doctors which patients need to be treated most urgently.
This triaging process entail two different networks. The first is called the “segmentation network,” which is a map of the eye that indicates the different types of tissue and different features of disease that were found (including unusual fluids, hemorrhages, or lesions). The second is called the “classification network,” which formulates a diagnosis based on the symptoms and then makes a referral. The results of the second network are given as a percentage, meaning clinicians are able to see how much confidence the AI system has regarding that particular analysis. Perhaps just as importantly, the researchers from Deep Mind believe the AI technology that they have used so successfully with OCT scans will be compatible with any model scanner, regardless of the hardware being used or even if the machine is in the process of being updated. That ensures that this technology can be used on innumerable patients around the world.
Delays in reading OCT scans can threaten patients’ sight, whether in those suffering from age-related macular degeneration or diabetic eye disease — or a person experiencing a more acute issue such as a lesion or hemorrhaging. Using this new AI technology should therefore cut down the time it takes for clinicians to read those OCT scans, not to mention the cost of having them read. And while Deep Mind concedes that these results are still “early,” their hope is that if these technologies continue to work in future trials there may one day be a time when these AI systems can become a routine part of clinical ophthalmic practice.
