Deep learning ensemble framework for multiclass diabetic retinopathy classification

Abstract

Diabetic retinopathy (DR) is the leading cause of blindness among adults and has no visible symptoms. Early detection is the key to prevent vision loss. Computer-aided deep learning using convolutional neural networks (CNN) have recently gained momentum for DR diagnosis as the cost can be significantly reduced while making the diagnosis more accessible. In this work, we present a fully automated framework DR network (DRNET) that fuses both image texture features and deep learning features to train the CNN model. The framework aggregates predictions from three CNN models using ensemble learning for more precise and accurate DR diagnosis when compared to standalone CNN. To strengthen the confidence of medical practitioners in acceptance of automated DR diagnosis, we extend the DRNET framework by producing model uncertainty scores and explainability maps along with the classification results.