ADRS-Net: Automated Diabetic Retinopathy Screening Net

Diabetic retinopathy is a prevalent diabetic complication that can cause blindness. Early screening and diagnosis of the eye can help reduce the chances of visual impairment. In this paper, we propose a computer aided diagnosis method for retinopathy evaluation referred to as Automated Diabetic Retinopathy Screening Net (ADRS-Net). This network is able to identify healthy and diabetic retinopathy fundus images. In addition, it is capable of identifying the different severity levels of diabetic retinopathy, from non-proliferative to proliferative diabetic retinopathy. The residual non-local attention mechanism integrated into the ADRS-Net is able to attend to informative regions that help the decision-making process, improve computation time efficiency, and reduce model complexity. We evaluated the model on a public dataset. The proposed model is able to outperform existing methods with a 98.5% accuracy to diagnose diabetic retinopathy and a Quadratic Weighted Kappa score of 90.5% to screen for the five different diabetic retinopathy severity levels in fundus images.