Diabetic retinopathy is the leading cause of blindness among working-age adults, and tens of millions are affected around the world by this disease; many undiagnosed. In developing countries, access to the usual expensive tools to diagnose diabetic retinopathy is limited or nonexistent. On the other hand, smartphone technology is cheap and available nearly everywhere. In this work, we developed specialized smartphone software and hardware to screen for diabetic retinopathy through the development of an image-processing algorithm, smartphone application, and attachment. We used Residual Networks coupled with the cyclic pooling technique for data refinement to train on 35,126 retinal images. Following 161 epochs of training, we were able to diagnose diabetic retinopathy with an accuracy of 78.1% with a resulting area of 0.752 under the ROC curve. We coded an app to facilitate the communications between user and algorithm, housed in a server and added photo functionality. Finally, we were able to design and 3D print an attachment that enables a smartphone to take retinal images. These results demonstrate that the Eyeagnosis system is capable of assisting doctors to diagnose diabetic retinopathy in the field. .
1 Introduction .
Diabetes mellitus, a condition characterized by high blood sugar from the loss of functional insulin, is the eighth leading cause of death in the world [1] and affects 415 million people worldwide [1]. Also, as blindness is a major complications of diabetes, it is one of the leading causes of blindness around the world and the leading cause of blindness among working-age adults [1]. Despite these figures, around the world, as many as half of the cases of diabetes remain undiagnosed, most of whom reside in developing countries [1]. Thus, a tool that could identify those with diabetic retinopathy, or DR, would serve not only as a tool to improve eyesight through early detection and treatment but would also prompt previously undiagnosed diabetics to receive treatment.