Comparison of Deep Learning Architectures Used in the Classification of Diabetic Retinopathy Images

Abstract

Diabetes is a significant disease that adversely affects human health and can cause permanent problems in many patients. Today, detecting the disease using retinal images to prevent diabetes-related blindness is a popular research area. In this study, the performances of convolutional neural network-based (CNN) architectures EfficientNetB3, VGG16, and ResNet50, which use deep learning techniques based on artificial neural networks to detect diabetes-related blindness, were compared. The Asia Pacific Tele-Ophthalmology Society 2019 Blindness Detection (APTOS 2019 BD) dataset, which contains 3662 samples collected from many participants in rural India and organized by the Aravind Eye Hospital in India, was used to compare the performances of these architectures used in image classification. The dataset was created using the Fundus photography technique under different conditions and environments. In the study, which initially used pre-trained ImageNet weights, user-defined hyperparameters such as learning rate, momentum, and dropout were adjusted according to the architecture to improve their performance. In the study, where each class had 400 images, the highest accuracy performance was achieved with the EfficientNetB3 architecture with an accuracy rate of 0.8236. The findings obtained in this study will enable early diagnosis of diseases that may cause diabetes-related blindness and prevent the progression of the disease. © The Institution of Engineering & Technology 2024.