Robust Yet Simple Deep Learning-Based Ensemble Approach for Assessing Diabetic Retinopathy in Fundus Images

Diabetes mellitus (DM) has epidemic proportions in incidence and prevalence worldwide. The people with DM are all susceptible to acquiring diabetic retinopathy (DR). Microaneurysms, haemorrhages, and exudates are the early symptoms of DR that manifest on the surface of the retina. This paper proposes a deep learning (DL) method based on non-clinical features for grading DR in retinal fundus images. The method is based on ensampling multiple DL networks. The multi-network architecture is well suited to represent the fundus image’s features properly. The proposed model consists of three well-known DL models, mainly VGG16, DenseNet121, ResNet50. For each network, a classifier network grades DR into five classes (i.e., No DR, Mild, Moderate, Severe, and Proliferative). A maximum voting algorithm is applied to merge the three probabilities resulting from the three classifiers to achieve the final DR grading. The proposed system is assessed using the public DDR dataset. The experimental results show that the proposed method based on voting with the DenseNet121 and VGG16, and ResNet50 outperforms the state-of-the-art.