Pre-Trained Convolutional Neural Network for Automated Grading of Diabetic Retinopathy

Diabetic retinopathy is an irreversible eye disease that is a prominent cause of blindness globally. For proper diagnosis and vision protection, it is essential to detect DR early and accurately. In the current scenario it is challenging to detect DR in an automated manner using fundus images. This work investigates the performance of deep learning based CNN models in an effective DR classification. We explore three popular models (Xception, InceptionV3, ResNet50) for the automated categorization of this retinal disease. A comparative study of three pre-trained models are analysed for multi-class DR grading. The simulation results show that the Xception model performs well with 89.8 % classification accuracy. To assess the efficacy of the suggested method, the accuracy, mis-classification rate, quality index, F-score, sensitivity, and specificity of each model are compared.