DATA ANALYSIS APPLIED TO DIABETIC RETINOPATHY SCREENING: PERFORMANCE EVALUATION

The number of people with diabetes mellitus (DM) has risen from 108 million in 1980 to 422 million in 2014. Diabetic retinopathy (DR) is one of the most common causes of blindness in the developed world. A pooled analysis of data from between 1980-2008 estimates that 93 million people around the world have DR. In this paper, we present a computer-aided automated image analysis system capable of handling images generated in real-life screening program. In this study, we analyzed 2932 color fundus images taken from 733 patients with DM, of which 454 (15%) images showed signs of DR validated by human graders. The system analyzed all images by detecting anatomical components such as the optic disc, macula and vascular system of the retina, then microaneurysms (MAs) and exudates as lesions. Once the presence/absence of the structures was determined, the combination of the results was subsequently used to provide a "DR/No DR" decision using a machine learning approach. The fundus images were graded by a trained and certified expert grader as well and the final diagnosis was compared to the outcome of the computer-based approach. The performances of the MA and exudate detectors used by the system were also evaluated. The area under the ROC curve (AUC) was 0.90 with the best performing setting of the algorithm. The evaluation of the proposed approach shows that it performs well against human graders and therefore might have the potential to be used in a clinical setting. There is a need for further evaluation on large scale, real-life clinical setting to explore its clinical utility.