Publication

Improved Automated Detection of Diabetic Retinopathy on a Publicly Available Dataset Through Integration of Deep Learning

Michael David Abràmoff; Yiyue Lou; Ali Erginay; Warren Clarida; Ryan Amelon; James C. Folk; Meindert Niemeijer

Automation is a prerequisite to improve health care efficiency affordability and accessibility.^1,2 In the management of the estimated 23 million Americans and 59 million Europeans with diabetes, automation of the retinal exam is sorely needed. The reason is that adherence to the regular eye examinations–necessary to diagnose diabetic retinopathy (DR) at an early stage, when it can be treated with the best prognosis and visual loss can be delayed or deferred^3–5–is frequently less than 60%.^4,6 This leaves millions of people with diabetes at risk for potentially preventable visual loss and blindness. Though OCT and widefield imaging have been proposed to improve screening performance, most present-day DR screening programs use 1 or 2 field retinal color fundus imaging, in order to reach cost-effectiveness.^7–9

Over the last two decades, the automated analysis of retinal color images for DR has been studied by many groups.^10–12 Though the lack of widely accepted, well characterized, and representative datasets makes comparison difficult, recent studies show that complete DR screening systems using such algorithms achieve adequate safety, as for example the Iowa Detection Program (IDP).^1,13–16 These algorithms are all based on classical expert designed image analysis, using carefully designed transformations including mathematical morphology and wavelet transformations.^17–20 More recently, we used data-driven machine learning to learn the lowest level wavelet transformations from training data, but this resulted in only marginal performance improvements.²¹