Diabetic Retinopathy Classification using Vision Transformer

Abstract

An adult's death from diabetes ranks among the top 10 global causes of death. Eye conditions like diabetic retinopathy (DR) are more common in people with diabetes. Loss of eyesight may arise from DR's damage to the retina's blood vessels. Grading the severity of DR is a crucial step to aid in early identification and treatment and to halt the disease's progression to vision impairment. Most currently available solutions are built using conventional image processing and machine learning methods. This paper applies the emerging vision transformer (ViT) model to the DR dataset with different optimizers. The dataset is available publicly and is highly imbalanced. The optimizers such as Adam (Adaptive moment estimation), Nadam (Nesterov Adam), and Follow the Regularized Leader (FTRL) are used for minimizing the loss function. A convolutional neural network (CNN) model is also implemented with different optimizers, and the results are compared with the ViT model. The Adam optimizer with the ViT model shows a better F1-score (0.732) than the CNN model.