DESIRE: Deep learning Enabled System Integration for Retinopathy Evaluation

Abstract

Diabetic Retinopathy (DR) is one of the significant reasons for visual impedance in the middle age population. An automated approach for detection of DR has an extraordinary social effect on rural population due to lack of expert ophthalmologists. The pre-screening of DR is exceedingly viable in averting the irreversible vision loss however its detection is tedious and an impeded process. Several strategies have been proposed in the literature, the greater part of them depend on the morphological component extractions. Deep learning conversely enhances the results when trained on the annotated data. In this paper, an automated DR grading algorithm is exhibited that examines patients' fundus images having a distinctive field of view and light intensity. It then classifies the images to their particular severity grade using deep learning. The main focus is to diminish the false normal (Fnorm) recognition rate of the framework to maintain a strategic distance from the characterization of DR influenced eye as normal. The vast majority of the therapeutic recognition approaches center around expanding the precision as opposed to stifling the Fnorm rate which influences its appropriateness in a genuine. The paper initially compares different preprocessing techniques and their impact on enhancing the image quality followed by the medical data balancing approaches. The proposed deep learning based algorithms are trained over a publicly accessible EyePACS dataset of 35126 images and give a high accuracy of around 98.98% with a reasonable Fnorm value. DR identification approach can be utilized to supplant the manual screening procedures and help the ophthalmologist with precise primary screening.