TECHNOLOGY 4.0

Google’s AI Eye Doctor, an artificial Intelligence algorithm, is being employed in a pilot project in Indian hospitals to detect diabetic retinopathy among patients, writes Alnoor Peermohamed

Business Standard is delighted to launch a special page on how technologies of the fourth industrial revolution are impacting governments, companies and citizens. The page will appear every alternate Thursday.

Picture this. A woman in her mid-40s walks into a diabetic clinic in a small town in India. Apart from the plethora of tests the physician runs on her, he also photographs her eyes. This picture is then uploaded onto a system that grades the condition of the patient’s retinas and spits out one of two verdicts — visit an ophthalmologist immediately or follow up with another test in the next 12 months.

What the doctor is looking for are signs of diabetic retinopathy (DR), a disease which accounts for 10 per cent of the cases where patients suffering from diabetes lose their vision. With 70 million diabetics, India is known as the diabetic capital of the world. So far, the process of grading retinal images to determine the presence of the disease was done manually, which takes hours. With the use of an artificial intelligence (AI) algorithm built by internet giant Google, we may soon get a diagnosis in a few minutes.

Over the past four years researchers at the Google Brain AI have worked with three hospitals in India — Aravind Eye Hospital, Sankara Nethralaya and Narayana Nethralaya — to feed the algorithm with images to train the AI in grading photographs of patients’ retinas.

Now India is also taking the lead in validating the algorithm’s performance in a real world scenario. Google is running a pilot at Aravind Eye Hospital in Madurai and planning a similar one at Sankara Nethralaya to grade photos of patients’ retinas via AI. The exercise began after researchers at Google published a paper saying that their algorithm had achieved an accuracy of 98.6 per cent in detecting DR, on a par with the performance of ophthalmologists and retinal specialists in the US.

“Our work and that of many other researchers have shown that deep learning can be used to train very accurate algorithms for detecting diseases like diabetic retinopathy. However, training an algorithm is just the beginning.

The next step is to better understand how to implement these algorithms in partnership with doctors and health care systems,” says Lily Peng, product manager at Google Brain Team.

Google’s AI is being used alongside the manual grading process that these hospitals already have. “We’ve been taking pictures of patients’ retinas for the past 15 years and we have a software that semiautomates the grading process for detection of diabetic retinopathy. We’ve now added the AI component which looks at the photos after the manual grading is done,” says Dr R Kim, chief medical officer at Aravind Eye Hospital in Madurai. “When there’s a big difference in the results from the manual process and the AI, we refer it to a senior retina specialist.”

Kim says that in the three months since the hospital began using AI to supplement its DR grading process, it has already proved to be more accurate than the manual grader. While, they are not replacing the manual grading process yet, it isn’t stopping the hospital from increasing the base of patients who will now be graded by the Google algorithm.

The Madurai-based hospital chain will roll out the use of the AI system at its primary vision centres starting next month, where, apart from undergoing general eye examination, diabetic patients will be checked for DR. The AI will grade the photos of a patient’s retina and the results will be shown to an ophthalmologist who would be available via teleconference.

“Doing the grading in high volume is a problem, but for AI it’s quite easy; the answer comes in a few seconds and the data is right there. Once we are comfortable with the accuracy, the AI will replace the manual grader and it will then be escalated to detecting other conditions in the retina,” adds Kim.

However, Google still needs to prove how well the software works in the real world where the quality of the photo will vary widely. From a regulatory perspective, April marked a historic moment when the US Food and Drug Administration (FDA) approved the sale of the first medical device using AI software. IDx-DR, a product that detects diabetic retinopathy, was able to achieve an accuracy of 87.4 per cent in detecting the presence of ‘more than mild’ diabetic retinopathy.

Unlike Google’s solution, IDx’s software is locked to working with specific hardware — a retinal camera called Topcon NW400. In India, that’s a problem given that retinal cameras vary widely in price (from ~500,000 to ~4 million) and in quality. If Google’s AI Eye Doctor is to work here, it needs to be able to detect the presence of DR even from lower-resolution photographs.

For Dr Rohit Shetty, vice-chairman at Narayana Nethralaya, this is a sticking point. When studies are commissioned, they’re done using high-quality images which are then processed heavily to give the most accurate view of the condition of a patient’s eye. In India and other countries in south Asia and Africa where this algorithm can be of great use, access to high-end cameras is an issue.

The bigger roadblock for the introduction of AI in medicine is the lack of laws surrounding this space. India is working on a telemedicine law to regulate the practice of doctors using tech to diagnose disease remotely, but it is a long way from being tabled.

“Technologies like Google’s AI Eye Doctor can be used in peripheral vision centres in rural areas. But if a doctor signs on the AI’s diagnosis, is he liable for whatever the machine says?” asks Shetty of Narayana Nethralaya. “Google certainly isn’t liable since every report they send comes with a disclaimer.”

Microsoft too is working on using AI to detect DR. The company has partnered with Forus Health, an eyecare device manufacturer in India, and has integrated its AI-based retinal imaging software into these machines. “This will help Forus technicians identify eye fundus (interior of the eye) images as well as disease conditions better,” Microsoft writes in a statement.

Given the potential of AI in this field of medicine, it may not be long before artificial intelligence is routinely used to diagnose diabetic retinopathy, and much more.