The Combination Between Deep Learning and Ensemble Stacking for a Fast and Accurate Detection of Diabetic Retinopathy Using Fundus Images

Abstract

Diabetic retinopathy (DR) is the leading cause of human vision loss in the world. To slow the progression of the disease we need early detection and diagnosis. Hence, rapid detection and accurate classification of DR is crucial in patient care. The use of Deep learning (DL) alone turns out to be slow and expensive, also the use of a single Machine Learning (ML) algorithm for classification leaves a significant margin of error. The detection of the disease also proves to be more difficult, especially in its early stages, this is due to the characteristics of DR less clear and visible on images. Therefore, the combination of Deep Learning and ensemble stacking was a relevant solution that we chose to enhance the ability to detect the disease in its early stages. In our research, we introduce an innovative method utilizing an automated model to identify and classify the existence and stage of severity of DR directly from fundus images. We utilized both of Machine Learning algorithms and a Deep Learning model; to achieve an accurate and fast detection we used the transfer learning and ensemble stacking as techniques. The proposed model uses a dataset published on Kaggle; our solution achieves an accuracy of 99.50%.