Comparative Analysis of Deep Learning Approaches for the Diagnosis of Diabetic Retinopathy

High levels of blood sugar causes diabetes. Diabetes on the other hand leads to various other health issues like heart diseases, kidney damage, nerves damage and eye damage. Diabetic retinopathy is one such disease that is being caused by diabetes and without early treatment or diagnosis it might lead to vision loss as well. Various Computer aided systems have been used and developed for the diagnostics of diabetic retinopathy in the past where it uses the traditional techniques to get low level features such as shape, colors or textures which are handcrafted features. With the advancement in AI and deep learning methodologies especially in the field of medical image analysis, efficient and accurate results are produced as it performs automatic extraction of features. These features can include high level features or some of the low-level features as well. Most commonly used deep learning method for image detection is convolutional neural network. In this paper, three different CNN architectures such as ResNet50, DenseNet121 and VGG16 are used as transfer learning models to carry out a comparative analysis of deep learning approaches for the diagnosis of Diabetic Retinopathy (DR). These models are used to classify diabetic retinopathy into multi-class (5 class based) classification on the basis of severity level and further into binary (2 class based) classification as 0-No DR and 1-DR. The accuracy so obtained on validation set for multi-class classification for the models ResNet50, DenseNet121 and VGG16 is 94.11%,94.64% and 91.52% respectively and that on test set as 94%, 92% and 86% respectively. Whereas the accuracy for binary classification on validation set for the models ResNet50, DenseNet121 and VGG16 is 98.35%, 98.48% and 95.86% respectively.