Artificial intelligence for diabetic retinopathy screening

Abstract

AbstractDiabetes constitutes a significant problem for global health and wellbeing. The number of people living with diabetes raised from 108 million in 1980 to 422 million in 2014. By 2035, 592 million people will have diabetes, with the largest rise in low‐ and middle‐income regions. Diabetic retinopathy (DR) is one of the major complications of diabetes, estimated to be the leading cause of blindness among working‐age adults globally. The increasing worldwide population, coupled with increasing prevalence of diabetes and increased incidence of diabetic retinopathy all lead to increasing number of patients with ocular complications of diabetes. Only a few countries were able to successfully establish and continue diabetic retinopathy screening on national level, most prominently – United Kingdom and Singapore. The NHS Diabetic Eye Screening Programme, a continuation of an English screening programme set up in 2006, reported in 2014 that DR is no longer the leading cause of certifiable blindness in England and Wales for the first time in 50 years. The advent of deep‐learning based DR detection revealed a significant improvement in the accuracy of the newly developed or improved systems. It requires a careful balance between sensitivity and specificity, imaging modality, gradeability of the images, all of which will need to be weighed against the potential cost. Deep learning DR detection has been found to be cost‐effective in developed countries, like Singapore and United Kingdom. However, the feasibility of implementing AI DR screening in countries without a robust teleophthalmology screening programme setup beforehand and other resource‐limited settings should still be evidenced. IDx‐DR is the first autonomous diagnostic software and one of the very first AI‐based software's in medicine to receive Federal Drug and Administration (FDA) approval. Eyeart was approved by FDA for the similar purpose in 2020. In European Union, there are more than 10 different AI‐based devices for DR screening registered, however in different classes, I and IIa. There are also significant shortcomings in current studies of AI in DR, including that great majority of them are sponsored or dependent on the respective algorithm's' company. Independent studies, particularly comparisons or studies establishing objective criteria through which the respective algorithms could be compared are rare.