Aravind Eye Hospital's Contribution to Deep Learning

Senthil Kumaran — Fri, 08 May 2026 11:55:47 GMT

I was reading the book, Genius Makers, written by Cade Metz. I was pleasantly surprised to come across the involvement and contributions of Madurai Aravind Eye Hospital to the field of Deep Learning, in year 2015, way before the onset of Large Language Models. The hospital's contributions were huge corpus of the retinal scan images used to identify a condition called diabetic retinopathy and by training the models with these dataset and then research, the models were able to identify the conditions with 90% accuracy.

Here is an excerpt from the chapter of the book.

The Aravind Eye Hospital sits at the southern tip of India, in the middle of a sprawling, crowded, ancient city called Madurai. Each day, more than two thousand people stream into this timeworn building, traveling from across India and sometimes other parts of the world. The hospital offers eyecare to anyone who walks through the front door, with or without an appointment, whether they can pay for care or not. On any given morning, dozens crowd into the waiting rooms on the fourth floor, as dozens more line up in the hallways, all waiting to walk into a tiny office where lab-coated technicians capture images of the backs of their eyes. This is a way of identifying signs of diabetic blindness. In India, nearly 70 million people are diabetic, and all are at risk of blindness. The condition is called diabetic retionopathy, and if detected early enough, it can be treated and stopped. Each year, hospitals like the Aravind scan millions of eyes, and then doctors examine each scan, looking for the tiny lesions, hemorrhages, and subtle discolorations that anticipate blindness. The trouble is that India doesn't produce enough doctors. For every 1 million people, there are only eleven ophthalmologists, and in rural areas, the ratio is even smaller. Most people never receive the screening they need. But in 2015, a Google engineer named Varun Gulshan hoped to change that. Born in India and educated at Oxford before joining a Silicon Valley start-up that was acquired by Google, he officially worked on a virtual-reality gadget called Google Cardboard. But in his "20 percent time," he started exploring diabetic retinopathy. His idea was to build a deep learning system that could automatically screen people for the condition, without help from a doctor, and so identify far more people that needed care than doctors ever could on their own. He soon contacted the Aravind Eye Hospital, and it agreed to share the thousands of digital eye scans that he would need to train his system. Gulshan didn't understand how to read these scans himself. He was a computer scientist, not a doctor. So he and his boss roped in a trained physician and biomedical engineer named Lily Peng, who happened to be working on the Google Search Engine. Others had tried to build systems for automatically reading eye scans in the past, but these efforts had never matched the skills of a trained physician. The difference this time was that Gulshan and Peng were using deep learning. Feeding thousands of retinal scans from the Aravind Eye Hospital into a neural network, they taught it to recognize signs of diabetic blindness. Such was their success, Jeff Dean pulled them into the Google Brain lab, around the same time that DeepMind was tackling the game of Go. The joke among Peng and the rest of her medically minded team was that they were a cancer that metastasized into the Brain. It wasn't a very good joke. But it was not a bad analogy. When Ilya Sutskever published the paper that remade machine translation—known as the Sequence to Sequence paper—he said it was not really about translation. When Jeff Dean and Greg Corrado read it, they agreed. They decided it was an ideal way of analyzing healthcare records. If researchers fed years of old medical records into the same kind of neural network, they decided, it could learn to recognize signs that illness was on the way. "If you line up the medical records data, it looks like a sequence you're trying to predict," Dean says. "Given a patient in this particular stage, how likely are they to develop diabetes in the next 12 months? If I discharge them from the hospital, will they come back in a week?" He and Corrado soon built a team inside Google Brain to explore the idea. It was in this environment that Lily Peng's blindness project took off—to the point where a dedicated healthcare unit was established inside the lab. Peng and her team acquired about one hundred and thirty thousand digital eye scans from the Aravind Eye Hospital and various other sources, and they asked about fifty-five American ophthalmologists to label them—to identify which included those tiny lesions and hemorrhages that indicated diabetic blindness was on the way. After that, they fed these images into a neural network. And then it learned to recognize telltale signs on its own. In the fall of 2016, with a paper in the Journal of the American Medical Association, the team unveiled a system that could identify signs of diabetic blindness as accurately as trained doctors, correctly spotting the condition more than 90 percent of the time, which exceeded the National Institutes of Health's recommended standard of at least 80 percent. Peng and her team acknowledged that the technology would have to clear many regulatory and logistical hurdles in the years to come, but it was ready for clinical trials.

They ran one trial at the Aravind Eye Hospital. In the short term, Google's system could help the hospital deal with the constant stream of patients moving through its doors. But the hope was that Aravind would also deploy the technology across the network of the more than forty "vision centers" it operated in rural areas around the country where few if any eye doctors were available. Aravind was founded in the late 1970s by a man named Govindappa Venkataswamy, an iconic figure known across India as "Dr. V." He envisioned a nationwide network of hospitals and vision centers that operated like McDonald's franchises, systematically reproducing inexpensive forms of eyecare for people across the country. Google's technology could play right into this idea—if they could actually get it in place. Deploying this technology was not like deploying a website or a smartphone app. The task was largely a matter of persuasion, not only in India but in the U.S. and the UK, where many others were exploring similar technology. The widespread concern among healthcare specialists and regulators was that a neural network was a black box. Unlike with past technologies, hospitals wouldn't have the means to explain why a diagnosis was made. Some researchers argued that new technologies could be built to solve this issue. But it was a far from trivial problem. "Don't believe anyone who says that it is," Geoff Hinton told the New Yorker in a sweeping feature story on the rise of deep learning in healthcare. Still, Hinton believed that as Google continued its work with diabetic retinopathy and others explored systems for reading X-rays, MRIs, and other medical scans, deep learning would fundamentally change the industry. "I think that if you work as a radiologist you are like Wile E. Coyote in the cartoon," he said during a lecture at a Toronto hospital. "You're already over the edge of the cliff, but you haven't yet looked down. There's no ground underneath." He argued that neural networks would eclipse the skills of trained doctors because they would continue to improve as researchers fed them more data, and that the black-box problem was something people would learn to live with. The trick was convincing the world it was not a problem, and this would come through testing—proof that even if you could not see inside of them, they did what they were supposed to do. Hinton believed that machines, working alongside doctors, would eventually provide a hitherto impossible level of healthcare. In the near term, he argued, these algorithms would read X-rays, CAT scans, and MRIs. As time went on, they would also make pathological diagnoses, reading Pap smears, identifying heart murmurs, and predicting relapses in psychiatric conditions. "There's much more to learn here," Hinton told a reporter as he let out a small sigh. "Early and accurate diagnosis is not a trivial problem. We can do better. Why not let machines help us?" This was particularly important to him, he said, because his wife had been diagnosed with pancreatic cancer after it advanced beyond the stage where she could be cured.

Senthil Kumaran (Posts about Datasets)

Aravind Eye Hospital's Contribution to Deep Learning