Diabetic Retinopathy Prediction Based on a Hybrid Deep Learning Approach

Abstract

In this paper, we address a medical condition that is thought to be among the most prevalent diseases that cause blindness in the working age, which is known as diabetic retinopathy. This medical condition is associated with blood vessels leaking in the retina, especially in those suffering from diabetes. Besides, the most difficult thing about this medical condition is that it doesn’t show any symptoms during its initial stage or that it comes with mild symptoms, which are very difficult to detect using the manual method by ophthalmologists, especially since early detection is essential for treatment. This paper introduces a novel automatic technique for identifying the degree of severity of diabetic retinopathy, based on a new hybrid deep learning approach (DenseNet121, Xception, and EfficientNetB3) with a pre-processing step. After that, the retinal image quality is enhanced using the pre-processing step. We feed the generated image to the hybrid deep learning model that starts with the three-transfer learning models to generate the feature maps, and then we serve these feature maps to a classification submodel to make decisions and detect images that are surfing from DR. In order to validate our approach and show its performance, we tested it on the APTOS dataset and obtained an impressive accuracy score of 86%.