Diabetic retinopathy detection models: Transition from conventional to deep learning approaches and a hybrid model

Diabetes is a chronic and common disease all over the world that has many side effects. One of the main side effects of this disease is diabetic retinopathy (DR), which if not diagnosed and treated in time, causes low vision and eventually blindness. Early diagnosis of this disease prevents the occurrence of complications of it. As artificial intelligence and deep learning technologies progress, various models for diagnosing medical images like fundus images in diabetic retinopathy became available. Due to the increase in the prevalence of DR, we propose a hybrid method with VGG-16 deep learning model and a hand-crafted feature extraction technique, namely Principal Component Analysis, for the classification of DR. To achieve this goal, two well-known datasets, namely APTOS 2019 and Messidor-2, are considered. On each of the datasets, two hand-crafted models, namely Principal Component Analysis (PCA) and Local Binary Patterns (LBP), and various deep learning models are implemented. In order to evaluate the models and compare them with each other and the state-of-the-art in this field, standard metrics such as accuracy are calculated. The results show that deep learning models are significantly superior to manual models regarding performance in diagnosing DR and its degrees. Moreover, the proposed hybrid method shows notable improvement in classification performance. Furthermore, the results obtained in this study have shown better performance than some other recent studies in the literature.