Classification of diabetic retinopathy in fundus images with improved deep learning models

Diabetic retinopathy (DR) is one of the most dangerous complications of diabetes. Early diagnosis and treatment is extremely important to reduce costs. It is very difficult to make a definitive diagnosis in cases of diabetic retinopathy. It is possible to determine the stages of the disease by examining fundus images by a specialist physician. Simplifying the process of determining the stages of this disease, which is seen in millions of people, will accelerate the diagnosis and treatment processes. However, the creation of labelled data is a costly process and creates differences of opinion among specialist physicians. This situation also affects the performance of the methods used in the diagnosis process. Classification of diabetic retinopathy using evolutionary neural networks is one of the deep learning approaches and there have been many recent studies on this subject in the literature. In this study, DR stages in fundus images were detected using a learning-based approach based on evolutionary neural networks. In addition, the study also examined whether the categorical attention mechanism is added to the models used in the stage classification process and whether the classification results are improved. The success of the study was tested by training on different network architectures. It was observed that improvement was achieved in all architectures obtained by training.