CAPSULE NEURAL NETWORK FOR AUTOMATIC DETECTION OF DIABETIC RETINOPATHY
No Thumbnail Available
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
UMT, Lahore
Abstract
Diabetic Retinopathy (DR) is a disease which becomes the cause of blindness and visual impairment usually in middle-aged patients. It is estimated that over 93 million people are affected by DR. Detection of Diabetic Retinopathy manually is a very time consuming and expensive task, which requires trained ophthalmologists to examine and evaluate DR using digital fundus photographs of the retina. Many researchers are building CAD techniques from many decades for timely detection and reducing the burden of ophthalmologists.
Deep learning techniques have boosted the performance of fundus diabetic retinopathy (DR) image classification. More precisely, convolutional neural network (CNN) achieves superior performance to that of the conventional machine or deep learning algorithms. Recently, in November 2017, a novel type of neural network named capsule network (CapsNet) was introduced to overcome the shortcoming of traditional CNN models. In this thesis research, we present three CapsNet architectures with limited samples for five stage DR classification, which is inspired by the simplicity and comparability of the shallower deep learning models. The presented CapsNet architectures were trained using publicly available Kaggle dataset. The experimental results show that CapsNet shows convergence behavior and better accuracy for the complex dataset. For CapsNet by using the kaggle dataset, achieves 76% accuracy, 69% sensitivity, 83% specificity respectively. This can be considered as better performance for such a newly developed model which lack much important information till now. Moreover, we observed that training the CapsNet model requires significant computational resources and its performance falls below the average performance level of CNN. However, we argue that CapsNet seems to be a promising technique for overcoming the limitation of CNN. Further by using more robust computational resources and refine CapsNet architectures, the better performance can be achieved but still, our proposed system can be used to diagnose DR in early stages and assist in the grading of diabetic retinopathy.