Classification of Diabetic Retinopathy Using Improved ResNet-50 Deep Learning Model

Diabetic retinopathy (DR) is a disease that affects the vision of individuals with high blood sugar. DR needs to be diagnosed in initial stages, otherwise may lead to permanent blindness. Deep learning (DL) models have demonstrated remarkable results in the DR classification. In this work, improved pre-trained ResNet-50 model is utilized to develop deep model along with Adam optimizer is used to accelerate the convergence process. Also, Adam optimizer enhances training efficiency particularly with large datasets. Through experiments on DR images, performance of proposed model is evaluated on Kaggle APTOS and EyePac datasets. On the Kaggle APTOS dataset, the model achieved an accuracy of 74.18%, sensitivity of 85.08%, specificity of 87.50% and F1-score of 81.35% while for EyePac dataset these values are 82.37%, 91.04%, 88.28% and 87.35%, respectively. Findings of the presented model are also contrasted with those of other existing models which confirms the efficacy of proposed model.