Early Detection of Diabetic Retinopathy Using Dynamic Routing CapsNet with EfficientNet Feature Extraction

Abstract

Diabetes patients may develop diabetic retinopathy, an eye disorder that may result in blindness and vision loss. It is considered as the major cause of blindness in the world among the working-age people. It can result in blindness if it is not discovered early. Moreover, there is no cure for DR; treatment keeps the eyesight intact. Early diagnosis and treatment of DR can greatly lower the possibility of visual loss. This paper proposes a novel Dynamic Routing-CapsNet (DR-CN) algorithm by integrating Dynamic Routing algorithm and Capsule Networks (CapsNet). The Dynamic Routing algorithm is used to train the network and create relationships between the capsules. Then, the EfficientNet is used for feature extraction because of its high accuracy and scalability. Also, the Capsnet is used for employing the relationship between the features by enhancing the performance of the method to differentiate among the various stages of DR. This method was performed based on the dataset which achieves a result of 98% accuracy by using the Convolutional Neural Networks (CNN) employing classification accuracy. Moreover, CNN is very effective for the classification of the images because they can easily learn about the features from the input images. These findings demonstrate that Dynamic Routing-CapsNet (DR-CN) algorithm provides a solution for DR screening, efficiently helps in early detection, and is useful for the healthcare system by reducing its difficulty in detection. Dense U-Net demonstrated exceptional segmentation performance, achieving accuracy rates of 0.94 (training set variation) and 0.93 (K-Fold cross-validation). Additionally, DR-CN showcased outstanding diabetic retinopathy classification results with 98.6% accuracy, 94.4% sensitivity, 94.3% specificity, and 96.2% F-Measure.