Dr Ramasamy Kim, head of retina services at an eye hospital in southern India, and his team at the Aravind eye hospital in Madurai have examined about 15,000 images from across the country showing the interior surface of the eyeball, known as the fundus.
By grading each image, marking abnormal spots, lesions and indications of bleeding, they have contributed to a database of retinal images from all over the world, compiled by Google and its health tech subsidiary, Verily.
The data has been assembled into an algorithm, a set of instructions built into a computer program to identify eye complications arising from diabetes.
Kim is now checking to see whether the diagnosis the program has made is accurate. It is. He points to several yellow spots that look like pin-pricks of light emerging from the reddish-brown orb. Over time, when severe diabetes damages the tiny blood vessels of the retina, he says, it causes them to leak blood and cholesterol. In many cases, the patient does not even realize this is happening.
There are no symptoms until retinal tissue begins to swell, vision becomes cloudy and eyesight begins to fail. At this point, it is difficult to treat and in some cases, irreversible. And today, Kim says, it is remarkable that a machine can be trained to identify even the earliest stages of the diabetic retinopathy and grade it with as much accuracy as a human doctor.
“It’s been very exciting to watch this take shape,” says Dr Renu Rajan, a retinal surgeon also involved in the project. “When a machine is trained to be capable of identifying abnormal patterns in this way, it saves a doctor so much time in diagnosis; time that could be better spent helping the patient manage the condition and in aftercare.”
“Big data” and artificial intelligence has come in for stinging criticism in recent years, over issues of privacy, fairness, accountability and transparency.
But in healthcare, especially in developing countries, AI technology that employs data sets for machine learning is now expected to make rapid strides as an effective screening and diagnostic tool. In India, for example, it has the potential to save millions of lives.
According to the National Health Profile report of 2017, the country has a little over 1 million doctors to treat its population of 1.3 billion people. Of these medics, only about 10% work in the public health sector. There is an acute shortage of healthcare providers and infrastructure in rural areas. In addition, diabetes is a growing problem in India, where an estimated 72 million people have the condition. Though eye care may not seem like a priority for those with diabetes, as many as 10% could develop diabetic retinopathy, which can result in blindness.
Since 2016, Aravind hospital has been involved in a pilot project to test the algorithm. At this stage, the results it generates is being checked against a manual grading process.
“If there is a big difference between the results generated from the AI software and the manual diagnosis by an ophthalmologist, a senior retina specialist will make the final decision on the grading,” says Kim. These discrepancies are analyzed, helping to improve and fine tune the data that will be fed into it.
“With a 97.5% accuracy rate, it hasn’t been too far off the mark,” says Rajan.
Two other major ophthalmic centres in India have been involved in building the algorithm – Narayana Nethralaya in Bengaluru and Shankara Nethralaya hospital in Chennai.
Kim hopes the future could include AI-powered machines set up like vending booths across India’s remotest villages, capable of taking photographs of people’s inner eye and offering a digital diagnosis.
“Think of it as a screening and referral tool, that could tell you with a great deal of precision whether you needed to see a specialist or not and if so, how urgently you needed to see one,” says Dr H Parida, a retina specialist. Parida assisted the team early on, helping to analyze retinal images to build the algorithm.
AI algorithms in development have the potential to work with specialized attachments to ordinary smartphone cameras, making them cost-effective.
“AI technology is exciting and has made considerable progress in eye care possible, but there are areas where we need to exercise caution,” says Dr Rohit Shetty, corneal and refractive surgeon and vice-chairman of Narayana Nethralaya. The quality of the algorithm and how well it works will depend on the quality of the images that are being fed into it, he adds. Regulating the quality of these images as AI becomes mainstream and establishing accountability in case of errors then becomes critical.
Google and Verily say they have addressed this issue in India by working with well-established eye care centres to build the raw data that their algorithm needs.
“By partnering with well-known institutions like Aravind eye hospital and Shankara Nethralaya, we can continue our research and pilot studies in implementing AI-powered screening technology, and then extend it to clinical practice,” says Dr Lily Peng, a Google product manager.
