Diabetic Retinopathy Screening Using Computer Vision

Abstract

Diabetic Retinopathy (DR) is one of the main causes of blindness and visual impairment in developed countries, stemming solely from diabetes mellitus. Current screening methods using fundus images rely on the experience of the operator as they are manually examined. Automated methods based on neural networks and other approaches have not provided sensitivity or specificity above 85%. This work presents a computer vision based method that directly identifies hard exudates and dot haemorrhages (DH) from 100 digital fundus images from a graded database of images using standard computer vision techniques, and clinical observation and knowledge. Sensitivity and specificity in diagnosis are 95-100% in both cases. Positive and negative prediction values (PPV, NPV) were 95-100% for both cases. The overall method is general, computationally efficient and suitable for further clinical trials to test both accuracy and the ability to the track DR status over time.

Deep Learning Approach to Diabetic Retinopathy Detection

“…Thus, algorithms for DR detection began to appear. The first algorithms were based on different classical algorithms from computer vision and setting thresholds (Michael D. Abrmoff and Quellec, 2010;Christopher E.Hann, 2009;Nathan Silberman and Subramanian, 2010). Nevertheless, in the past few years, deep learning approaches have proved their superiority over other algorithms in tasks of classification and object detection (Harry Pratt, 2016).…”