Diabetes induced blindness: AI detection shows clinical promise

The researchers confirmed that the sensitivity and specificity of the screening tool is over 90%

January 06, 2018 05:41 pm | Updated 07:54 pm IST

 Diabetic retinopathy does not have to lead to blindness if the condition is detected early enough. (Photo used for illustrative purpose only.)

Diabetic retinopathy does not have to lead to blindness if the condition is detected early enough. (Photo used for illustrative purpose only.)

An artificial intelligence (AI)-based diagnostic tool developed by Google and researchers in India, for detecting diabetic retinopathy (DR), is showing clinical promise in the Indian setting, according to Rajiv Raman, a senior consultant at Chennai’s Sankara Nethralaya, one of the country’s leading eye hospitals.

Both Sankara Nethralaya and Aravind Eye Hospital in Madurai are working with 2,000 patients each to validate the AI diagnostic tool in a clinical setting. Dr Raman confirmed that the sensitivity and specificity of the screening are over 90%. The higher a test’s sensitivity, the higher the probability it will correctly identify a disease when it is present and the higher the test’s specificity, the higher the chance it will correctly identify the absence of the disease in individuals who do not have it.

Aravind Eye Hospital is now in the process of incorporating the deep learning system in their telescreening program, Dr R. Kim, Chief Medical Officer of the hospital told Th e Hindu.

AI-first world

AI has been the buzzword in tech for some time now, and tech giants want poll position in this space. We are moving from a mobile-first world to an AI-first world, according to Google CEO, Sundar Pichai, and Google, for one, seems to want to be the go to company for AI, as it has become in the search engine market.

Advances in the consumer space have found utility in medicine that can be applied readily, Lily Peng, a doctor of medicine turned product manager at Google, told a group of reporters at Google’s Tokyo offices a few weeks ago. Dr Peng and her colleagues have been working with ophthalmologists in India over the last few years to address DR, the fastest growing cause of blindness worldwide. Some 425 million people have diabetes, and another 352 million are at risk of developing it. In India alone, there were more than 70 million cases of diabetes in 2017. By 2045, India is projected to have 134 million cases, higher than China’s projected 120 million and making India the country with the most number of diabetics, according to the International Diabetes Federation.

Diabetic retinopathy does not have to lead to blindness if the condition is detected early enough. This is done by looking at the retina, where the presence of lesions — which can indicate fluid leakages and bleeding — is used to determine the condition’s severity. A 2016 scientific paper co-authored by Dr Peng, Dr Kim, Dr Raman and others announced that Google had developed an AI system to look at a retinal image and provide a diagnosis — referable DR (i.e., moderate and worse cases of DR), referable macular oedema or both. The tool is based on machine learning, a process by which a computer learns from patterns it identifies in examples fed to it, rather than via programmed rules. The examples in this case were retinal images captured by a fundus camera, i.e., a camera with a low powered microscope attached to it.

The program’s neural network (an enormous mathematical function) was trained by a process called deep learning, which involved comparing the program’s diagnosis for each image with that of the ophthalmological panel and adjusting the parameters of the function to reduce the error margin for each image. This process is repeated for each image until the program can make an accurate diagnosis based on the intensity of pixels in the retinal image.

The algorithm showed similar, in fact slightly better, levels of sensitivity and specificity as a panel of ophthalmologists.

The images for Google’s study were sourced from the U.S. as well as three hospitals in India: Sankara Nethralaya, Aravind Eye Hospital and Narayana Nethralaya in Bengaluru. Sankara Nethralaya was already working with an AI-based system to screen for DR according to Dr Raman. This test used a process called feature recognition, i.e., identifying specific features such as lesions, associated with the condition to predict DR. However, the specificity and sensitivity of these tests were in the 60-70% range, below the viability threshold of about 80%, Dr Raman told The Hindu. “[It] was a little bit primitive, deep learning is definitely the better way to go around [sic],” Dr Raman said. With sensitivity and specificity above 90%, both in the lab and in a clinical setting, the future looks bright.

Challenges remain

Nevertheless, challenges remain, especially in the Indian context. There is no screening programme for diabetes in India. Those who have diabetes are supposed to have an annual exam, yet, the screening has usually been opportunistic – i.e., if a diabetic comes to the hospital, their retina is screened.

The potential is enormous, as it will not only increase the amount of screening but also free up doctors to confirm diagnoses and provide treatment rather than do the first level of screening themselves. The AI system can also be used to diagnose other diseases.

At Google Tokyo, Dr Peng tells us that her team is working on providing ‘better ground-truth’ and enhancing the ‘explainability’ of the deep learning diagnostic tool. Better ground-truth refers to more relevant information that is fed to the system with each case; this can help improve the diagnosis (for example, more images of the retina, or a disagreement between the doctors on how to grade the case). ‘Explainability’ gets the system to provide a ‘why’ for its diagnosis, for instance, by producing a heatmap that shows the lesions that made it classify a specific case of DR a certain way.

Providing the AI system with better and often, more, ground truth can lead to interesting outputs. For instance, the system can detect the sex of the patient based on a retinal image. “Somehow it can detect gender, which we cannot detect with an eye examination. If I see a retinal image, I cannot say,” Dr Raman said.

While the eye has often been called a window to the soul, for Dr Raman, it is a window to the body. ““When it can tell gender, why not HbA1C [ glycated haemoglobin, a sugar test parameter]?” he says. The day may not be far when you will be able to take a picture of your eye with a smartphone and get a whole lot of information on your health.

(The writer was a guest of Google in Tokyo a few weeks ago.)

Top News Today

Sign in to unlock member-only benefits!
  • Access 10 free stories every month
  • Save stories to read later
  • Access to comment on every story
  • Sign-up/manage your newsletter subscriptions with a single click
  • Get notified by email for early access to discounts & offers on our products
Sign in


Comments have to be in English, and in full sentences. They cannot be abusive or personal. Please abide by our community guidelines for posting your comments.

We have migrated to a new commenting platform. If you are already a registered user of The Hindu and logged in, you may continue to engage with our articles. If you do not have an account please register and login to post comments. Users can access their older comments by logging into their accounts on Vuukle.