A lightweight transfer learning based ensemble approach for diabetic retinopathy detection

Abstract

Diabetic retinopathy (DR) is a fatal and irreversible eye disease that affects millions of people worldwide. It occurs due to high blood sugar level in the body of a diabetic patient, so it requires immediate attention which goes beyond the clinical solutions. With the advancements in deep learning and computer vision there are maximum possibilities of predicting this disease at early stages. Based on the severity of disease, different labels have been assigned to different classes of this disease as follows: 4 for proliferative DR, 3 for severe DR, 2 for moderate DR, 1 for mild DR and 0 for No DR. In this paper we proposed a deep learning-based ensemble approach using pre-trained and customized bi-class (CNN) base-learners like MobileNet, InceptionV3and DenseNet121 which were identified during initial investigation. These deep learning models were used as the base learner because of their promising performance in ensembles compared to the other deep learning base learners. All the work in the literature has studied this as a single complex multi-class problem or a bi-class problem where earlier stages are grouped together (0 to 3) and treated as one class and 4 as separate another class. Our work breaks this multi-class problem into multiple simpler two class problems using OVO(One-Versus-One) approach. Several benchmark data sets such as APTOS 2019, IDRiD, Messidor-2 and DDR which are multi-class data sets were used for training and testing our models. Data augmentation techniques were also utilized. Performance metrics such as precision, recall, f1-score, and accuracy were used for evaluation. Our ensemble models showed a remarkable performance with precision, recall, f1-score, and accuracy for most of the datasets used in this study. In addition to this our ensemble models have minimum number of trainable parameters which makes them an ultimate choice.