Proposal of a low­cost device to support remote diabetic retinopathy detecting based on fundus images.

Abstract

Diabetes causes several problems, including diabetic retinopathy, which when discovered belatedly can lead to total blindness. Brazil is also the 8th largest country in the world, with conurbation problems and an increase in diabetes diagnosis in the past 10 years. In this context, the present work aims to propose a low-cost prototype to support the diagnosis of diabetic retinopathy based on fundus examinations images so that physicians are able to perform early diagnosis in remote locations.This prototype should allow for early detection and treatment in loco, thus increasing the chances of a positive outcome for the patients. First we studied technical aspects relevant to the proposal such as physiological aspects of diabetic retinopathy, Artificial Neural Networks, Accelerated and Edge computing. Our methodology consisted in a comparison of embedded hardware with capabilities to perform complex computations, a survey of models for the classification of diabetic retinopathy and available databases, including research choices. Artificial Neural Networks to identify diabetic retinopathy were evaluated in our low-cost embedded system in terms of accuracy. The accuracy must be enough to determine the priority of the patient’s case for treatment. This work reached accuracy levels around 84% with a low cost system and less computational power, positioning itself well in the state of the art of systems within greater computational power. The results indicate that the platform is indeed low-cost and suitable for this application.