Early detection of diabetic retinopathy using deep learning techniques

Abstract

We, humans, are the bearer of diseases. While most of them have a thoroughly researched and contemplated solution set, some of them do not. Diabetes is one of those common diseases that do not have a clear solution but has ways to minimize its e ects. It is a globally prevalent condition that leads to several complications in- cluding those that are deadly. One of those intricate complexities includes Diabetic retinopathy (DR), a human eye disease that may a ect one or both eyes hamper- ing the functionality and leading to compromised vision and eventually, permanent blindness. Thus, detection of diabetic retinopathy in the primitive stages will help reduce the chances of getting visually impaired, following proper treatment and other necessary precautions. The prime objective of our paper is to take aid from the state-of-the-art models which are pretrained on di erent images and also to pro- pose a basic CNN model that will have comparative results. To be more precise, we have used transfer learning models like DenseNet121, Xception, Resnet50, VGG16, VGG19, and Inception to classify the data based on single-label and multi-label. In our approach, single-label classi cation using categorical cross-entropy and softmax function works better as we reached the best accuracy, precision, and recall values using the approach. In our case, Xception has reached an accuracy of 82% which is a state-of-the-art result for the used dataset. In addition, our proposed model reached an accuracy of 71% which worked better than some of the transfer learning models. Finally, most of our approaches classi ed the data correctly even though the dataset is very unevenly distributed.