Abstract
Diabetes is a chronic condition characterized by elevated high sugar levels in blood. One of the serious complications of the disease is Diabetic Retinopathy (DR), where the minute retinal blood vessels are blocked causing several symptoms ranging from blurred vision to vision loss. Early diagnosis and detection are crucial to control the symptoms and possibly delay the progression of the disease. Fundus photography is a cheap and accurate diagnosis modality used by ophthalmologists. However, fundus photography datasets suffer an extreme imbalance among different classes of DR. In this paper, we introduce a two-stage Deep Convolutional Neural Network (DCNN) architecture that is able to successfully categorize DR into one of 3 groups (i.e. controls, moderate DR, and severe DR) using a fine-tuned ResNet-50. Furthermore, a fine-tuned ResNet-18 was used to further classify moderate DR and a fine-tuned ResNet-50 was deployed to classify severe DR. The proposed architecture utilizes a preprocessing stage with image resizing and data augmentation included. Training and 5-fold cross validation were executed on the Kaggle APTOS 2019 dataset of 3,648 fundus images for 10 epochs. The proposed architecture achieved a 5-fold cross validation accuracy of 91% in the first stage, 90%, and 80% in the second sub-stages outperforming the-state-of-the-art architectures.