Automated Diabetic Retinopathy Grading using Resnet
Doaa K. Elswah, Ahmed Elnakib, Hossam El-Din Moustafa
Abstract
This paper presents a deep learning framework for the classification of diabetic retinopathy (DR) grades from fundus images. The proposed framework is composed of three stages. First, the fundus image is preprocessed using intensity normalization and augmentation. Second, the pre-processed image is input to a ResNet Convolutional Neural Network (CNN) model in order to extract a compact feature vector for grading. Finally, a classification step is used to detect DR and determine its grade (e.g., mild, moderate, severe, or Proliferative Diabetic Retinopathy (PDR)). The proposed framework is trained using the challenging ISBI'2018 Indian Diabetic Retinopathy Image Dataset (IDRiD). To remove the training bias, the data is balanced to ensure that each DR grade is represented with the same number of images during the training process. The proposed system shows an improved performance with respect to the related techniques using the same data, evidenced by the highest overall classification accuracy of 86.67%.