Diabetic retinopathy (DR) is a major complication of diabetes, and the leading cause of blindness among working adult people worldwide. About 600 million will have diabetes by 2040, with a third having DR and 10% with severe vision-threatening DR (VTDR). DR screening along with timely referral and treatment, is a universally accepted strategy for the prevention of visual impairment. Currently, DR screening by fundus photography, usually within a tele-ophthalmology framework, with assessment of the fundus photographs by human assessors (e.g., ophthalmologists, general physicians, technicians) is the most commonly used method for DR screening. However, this type of DR screening program is limited by availability and training of human assessors, and long-term financial sustainability. The need for low cost, sustainable DR screening programs is substantial

Artificial intelligence (AI) using deep learning (DL) technology has substantial potential for DR screening. Previous technology for automated DR screening using traditional “pattern recognition” techniques to detect specific DR lesions (e.g., microaneurysms) have been promising but has not achieved the sensitivity and specificity to break the “translational gap” from research to clinical adoption. DL uses much larger datasets and a convoluted neural network (“black box”) approach to mine, extract and learn patterns and/or features to determine a disease state or condition. Recently, researchers from Google Brain using DL learning technology have reported high sensitivity and specificity (>90%) in detecting referable DR from retinal photographs. Our group has achieved similar results using based on multi-ethnic Asian population. For translational impact, AI and DL technology should be validated in “real-world” DR screening programs where fundus images have varying qualities (e.g. cataract, poor pupil dilation, poor contrast/focus), in patient samples of different ethnicity (i.e. different fundi pigmentation) and systemic control (poor and good control). Furthermore, in any screening programs for DR, the detection of incidental but common vision-threatening conditions such as glaucoma and age-related macular degeneration should be incorporated, as missing such cases may not be acceptable to clinicians. Finally, there needs to be careful management of patient and physician acceptance of AI’s “black box” approach. Only then will deep learning technology be applicable in large scale screening programs for DR.



Professor & Medical Director, Singapore National Eye Center
Chair of Ophthalmology & Vice-Dean, Duke-NUS Medical School
National University of Singapore
Email : wong.tien.yin@singhealth.com.sg
Cell Phone: +6598565354
Work Phone: +6563228333