Though most of vision loss from DR is avoidable through early detection and effective treatment strategies, DR screening programs are not widely implemented in many areas. This is mainly because inadequate availability of trained eye care personnel and financing, which is particularly important in the developing countries.

 

In the most recent 20 years, many studies tried to apply artificial intelligence (AI) aided method in DR screening with relatively small datasets (<1000 images). The machine-learning method used was including support vector machine, random forest, k-nearest, etc. Such studies had areas under the curve of the receiver operating characteristic curves (AUC) of up to 0.970. However, the algorithms build on such small training set usually were over-fitting and could not be generalized to external datasets.

 

In the wake of deep learning technique introduced by Prof. Hinton, several studies reported novel data on using AI for the detection of DR or referable DR. Gargeya et al. (2017) and Abramoff et al. (2016) validated their models using externally public datasets and reported high diagnostic accuracy for the classification of any DR with AUC of 0.94-0.98.

 

Google also developed a tool using AI to detect referable DR and reported excellent accuracy (AUC 0.99) when validated on external retinal images (almost 10000 fundus photographs). In addition, Ting et al. developed an AI model used 76,370 retinal photographs and reported robust accuracy for referable DR and vision-threatening DR against 10 multi-ethnic retinal image datasets. We also trained an AI using almost 100 000 fundus images and further validated in Malay, Caucasian Australians and Indigenous Australians with high accuracy (a sensitivity of 92.5% and a specificity of 98.5%).

 

Recently, the FDA of the U.S. approved a software called IDx-DR as the first AI medical device to classify fundus images for DR when used a specific retinal camera named Topcon NW400.

 

Overall, AI for detecting DR had achieved excellent performance. However, there are still many problems need to be addressed before application AI in to real-world setting.

 

Prof. Mingguang He is currently Professor of Ophthalmic Epidemiology in University of Melbourne and Centre for Eye Research Australia, Director of WHO Collaborating Centre for Prevention of Blindness (Australia). He is the former Associate Director and Professor of Ophthalmology in the Zhongshan Ophthalmic Center (ZOC), Sun Yat-sen University in Guangzhou, China. His research interests include clinical and epidemiological research, randomized clinical trials, twin studies, imaging technology and big data research. He ran the first glaucoma survey and twin registry in China. He has published more than 290 scientific articles in international peer-reviewed journals, including the Journal of the American Medical Association (JAMA), the Lancet  as well as important book chapters, with more than 11900 citations. He has given more than 90 invited lectures at regional Asia and international conferences. He serves as editorial board member for several important journals, including Ophthalmology, the top ranked ophthalmology journal. He has received several awards including a distinguished young scholar award from the National Natural Science Foundation of China (2011) and the Holmes Lecture Award from the Asia-Pacific Academy of Ophthalmology (2015).

This is the ninth in a series of discussions introducing some of the programme sessions that will be featured at the IDF Congress 2019 in Busan, Korea, 2-6 December 2019.