ENHANCED LEARNING HYBRID MODELS ON AUTOMATED DIABETIC RETINOPATHY DETECTION IN FUNDUS IMAGES.

Abstract

Diabetic retinopathy is a direct complication of diabetes mellitus and is commonly characterized by retinal lesion formation leading to impaired vision. If not detected early enough, it leads to total blindness. Diabetic Retinopathy is rarely reversible, and treatment for diabetic retinopathy only delays the onset of blindness. Therefore, the earlier the diagnosis of diabetic retinopathy and the management of the disease, the lower the likelihood of vision loss. Two separate datasets were compared in this study, both of which consist of five different classifications of diabetic retinopathy image types (mild, moderate, no diabetic retinopathy, proliferative diabetic retinopathy, and severe). In this study, 80% of the images were used for training, and 20% of the images were used for testing. The performance of an ensemble of Ensemble classification techniques, Random Forest, K-Nearest Neighbor and Logistic regression using Stacking, Voting and Averaging techniques will be evaluated using the accuracy, precision, recall, F1-score and ROC metrics in this study. These metrics are based on the confusion matrix produced by the classification method employed. As a result, it has been shown that the average of the classifiers produces the best results.