Performances of CNN Architectures on Diabetic Retinopathy Detection Using Transfer Learning

Abstract

Deep Learning (DL) methods have become popular because they automatically extract and learn features without other feature extraction algorithms. Diabetic Retinopathy (DR) is a disease that is solely determined by blood vessel segmentation, but the diagnosis can differ from person to person. Thus, deep learning can simplify and improve the process of diagnosing DR. In this study, known Convolutional Neural Network (CNN) architectures were used as a transfer learning method to classify DRIVE dataset images of DR patients and healthy individuals to determine the most efficient architecture. In this study, CNN architectures were used both as classifiers and feature extraction methods. The most efficient CNN architecture was found to be ResNet18 with 100.0% accuracy when the classification was realized with ResNet18 itself and ANN which uses the activations of ResNet18 as features.