International Research Journal of Education and Technology

One of the primary side effects of diabetes mellitus is diabetic retinopathy. Fundus examination is a frequent method used to obtain retinal pictures for the purpose of assessing the disease's progression. Nevertheless, those pictures might have issues with poor contrast, insufficient illumination, and an excessive amount of noise, among other things that could make medical analysis and action difficult. In this regard, the work seeks to provide low quality digital retinal pictures publicly available in the DDR, EyePACS/Kaggle, and IDRiD databases by applying the neural network VGG16 to classify the diabetic retinopathy in 5 categories and with an additional class (called class 5). Size adequacy pre-processing of retinal images, data cleaning (removing low-quality images from other classes and adding them to class 5), data augmentation and class balancing during the training phase, hyper parameter adjustment, and image classification using the VGG16 neural network are all included in the proposed methodology. This proposal has demonstrated the greatest performance for the DDR database in terms of accuracy, precision, sensitivity, specificity, and F1-score among the tests conducted on the three databases. This work adds performance measures sensitivity, specificity, accuracy, and F1-score, and improves the state-of-the-art results obtained in the DDR and IDRiD databases without DME. Keywords— Diabetic Retinopathy (DR), Vgg-16 Model, Retinopathy, Transfer Learning, Retinal Images