Researchers from NUS Computing, the Singapore National Eye Centre (SNEC) and the Singapore Eye Research Institute (SERI) have developed an artificial intelligence (AI) screening technology capable of identifying retinal images showing signs of diabetic retinopathy with high accuracy.
Diabetic retinopathy is a diabetes-related complication that affects one in three diabetic people. It is currently diagnosed using manual assessments of retinal photographs involving large teams of trained professionals.
The screening technology employs a deep learning system that uses representation-learning methods (in machine learning, representation learning is a set of techniques that allows a system to automatically discover the representations needed for feature detection or classification from raw data) to process large amounts of data, and recognise intricate structures and meaningful patterns that may not be visible to the human eye.
The researchers developed and trained the system to recognise and classify retinal images, and tested its performance against close to 500,000 images from multi-ethnic populations across the US, Australia, China, Hong Kong, Mexico and Singapore.According to the press release, this is by far the world’s largest dataset for evaluating the use of a deep learning system to screen for an eye condition.
(Researchers in different parts of the world have been trying to use AI for the diagnosis of diabetic retinopathy. In 2016, Google Brain collaborated with doctors both in India and the US, and created a development dataset of 128,000 images which were each evaluated by 3-7 ophthalmologists from a panel of 54 ophthalmologists. This dataset was used to train a deep neural network to detect referable diabetic retinopathy. In September 2017, researchers at the Commonwealth Scientific and Industrial Research Institute (CSIRO) announced the development of an AI-driven eye-screening technology for diabetic retinopathy.)
The same technology can also be employed to diagnose two other eye conditions – glaucoma suspect and age-related macular degeneration.
NUS Computing Professor Lee Mong Li Janice, who is part of the cross-institutional research team said, that the technology has sensitivity greater than 90 per cent and specificity greater than 85 per cent for the three eye conditions. (Sensitivity measures the proportion of positives that are correctly identified as such, while specificity refers to the proportion of negatives, that is people do not have the disease, who are correctly identified as such.)
There is a growing prevalence of diabetes in Singapore, with the disease affecting 400,000 Singapore residents, and one in three Singaporeans is at risk of developing it over his/her lifetime and the number of those with diabetes is projected to reach one million by 2050.
The researchers believe that their screening technology will provide a more efficient and sustainable way of screening patients for diabetic retinopathy.
Professor Wong Tien Yin, Medical Director of SNEC and Chairman of SERI, said that the technology will help increase efficiency and reduce cost. In countries with existing screening programmes, such as Singapore and the UK, it could replace a large proportion of the manual assessment required.
“In communities and countries without existing programmes and without sufficient ophthalmologists — such as developing countries, parts of China, India, South America — it can be used as a first line screening tool to accurately screen for cases that require referral to an ophthalmologist for treatment,” added Prof Wong, who is also Vice Dean (Clinical Sciences) at Duke-NUS Medical School.
The research team is now beta testing their AI screening technology in the Singapore Diabetic Retinopathy Screening Programme. They aim to increase their datasets to five million images from around the world over the next five years.
“We are also developing more complex algorithms for different severity levels of diabetic retinopathy, predictive algorithms for the incidence and progression of the eye condition and diabetes-related systematic complications such as stroke, coronary diseases and chronic kidney diseases,” added NUS Computing Professor Wynne Hsu.