Classification of Diabetic Retinopathy using Capsules

Deep learning models have performed exceptionally well in detection of diabetic retinopathy. Most of these existing works either use Multi Layered Perceptron (MLP) or Convolutional Neural Network (CNN) based models. A significant drawback of these models is their inability to retain spatial dependencies as we go deeper into the network. Because of such issue, these models focus only on extraction of feature maps which help in classification or detection. In recent years' transformers have shown enormous promise in both natural language processing and computer vision, due to which there have been significant work on transformer based techniques for image classification and object detection. One important model used in this research work is the Capsule Network which uses Set Transformers to perform vision tasks with high accuracy and has the capability to maintain spatial dependencies throughout the process. In this work, we leverage the power of capsules to the diabetic retinopathy classification and detection problem and trained capsules for classification of retinal fundus images are considered which were categorize into five different classes depending on the severity of retinopathy. Further, we also propose a sliding window based detector which can pinpoint the exact position of a blood burst in the retina which will ease the job for ophthalmologists while studying retinal fundus. From our experiments we found that capsules provide better results than existing convolutional neural network and multi layered perceptron based approaches in standard retinopathy datasets.