Weighted Voting Ensemble Learning of CNN Architectures for Diabetic Retinopathy Classification

Abstract

Diabetic Retinopathy (DR) is a diabetes disease that attacks the retina of the eye and can be recognized through retinal images. The process of assisting retinal images can be done by applying deep learning-based methods, one of which is the Convolutional Neural Network (CNN). CNN has many architectures that can perform image classification processes, namely ResNet-50, MobileNet, and EfficientNet. Weaknesses of each architecture can be overcome through ensemble learning methods that can add up the performance results of each classification method. The study applies the ensemble learning method to improve the performance of the ResNet-50, MobileNet, and EfficientNet architectures in paying for DR disease on the retina by weighted voting. The data used are the APTOS and EyePACS datasets. The method in this research is data collection, training, testing, and evaluation of each architecture and ensemble learning. The results of the superior ensemble learning performance in the value of accuracy, F1-Score, and Cohens Kappa were obtained respectively 93.3%, 93.42%, and 0.866. The best specificity value was obtained by Resnet-50 at 99.78% and the highest sensitivity value was obtained by EfficientNet at 96.2%. Based on the classification results of each architectural and ensemble learning, it can be interpreted that the proposed ensemble learning method is excellent to perform image classification for Diabetic Retinopathy.