Diabetic Retinopathy Diagnosis Using a Two Stage Convolutional Neural Network

Manual detection of diabetic retinopathy is subject to diagnoses based on the knowledge and experience of an ophthalmologist. Implementing a system based on convolutional neural networks (CNN) helps in early recognition of the condition and its automatic severity’s degree classification. The dataset images are edited with pre-processing in order to clarify their characteristics. This paper describes a two-stage system, the first consists of a binary model that is responsible for identifying sick and healthy patients, while the second is a multi-class model that classifies the severity of the condition. The behavior of the model to classify new fundus images yields similar results to the public solutions presented in the APTOS 2019 Blindness Detection competition and in other projects. The main difference of the present proposal resulted in a model that is much lighter than the pre-trained neural networks and specialized by different stages.