Abstract:

Diabetic Retinopathy (DR) is a disease that occurs due to the damage caused by high blood sugar caused by diabetes in the permeability of blood vessels in the retina. In cases where the disease is not diagnosed and treated early, it causes severe eye disorders and loss of vision. Most of the complications can be prevented with blood sugar control and early treatment, but due to the complexity and variety of DR, it is difficult to detect with manual methods. Today, great success has been achieved with the Convolutional Neural Network (ESA), which is a deep learning model in the diagnosis of diseases that are difficult to detect by experts. In this study, a hybrid model is proposed in which ESA and machine learning are used together for the detection and grading of DR. In our model, the transfer learning model with ESA architecture acts as an automatic feature extractor and ResNET-50 is used. K-Nearest Neighbor Algorithm, Random Forest Algorithm and Extra Trees Algorithm are used separately for classification. In the study, the Aptos 2019 dataset, which includes 'Proliferative', 'Severe', 'Moderate', 'Mild' and 'No diabetic retinopathy' retinal images, was used according to the severity of the disease. Since there is a class imbalance in the data set, the imbalance between the classes was eliminated and data augmentation was achieved by using the synthetic data augmentation (SMOTE) technique. Classification was carried out by applying 10-fold cross-validation to our data set for training and testing. As a result of the classification process, an accuracy of 93% and an F1-score of 93% were obtained.

Keywords: