Classification of Diabetic Retinopathy Disease with Deep Learning Methods

Diabetes is defined as a chronic disease that occurs as a result of an increase in blood sugar level (hyperglycemia), in which the organism cannot make sufficient use of carbohydrates, fats and proteins due to the inability of the pancreas to produce enough insulin hormone or the inability of this hormone to function. According to the chronic diseases report published by the World Health Organisation, diabetes ranks first in terms of intensity. One of the complications of type 1 diabetes is that it causes diabetic retinopathy. Diabetic retinopathy is defined as an eye condition caused by damage to the blood vessels in the light-sensitive tissue (retina) located at the back of the eye due to diabetes complications. According to the International Diabetes Federation (2021) Diabetes Atlas 10th Edition, diabetes is among the top three diseases that cause blindness. Blindness caused by diabetes is mostly caused by the destruction of small vessels in the retina due to long-term hyperglycemia. Approximately 25% of diabetic patients worldwide have diabetic retinopathy at any level. There are approximately 2 million diabetic patients in our country and 25% of these patients have diabetic retinopathy. There are 5 classes of diabetic retinopathy. These are non-proliferative diabetic retinopathy (npdr), mild non-proliferative retinopathy, moderate non-proliferative retinopathy, severe non-proliferative retinopathy, proliferative diabetic retinopathy (pdr) from the lowest to the most severe. In this study, using the APTOS2019 dataset, a computer-aided diagnosis system is created to help doctors make early diagnosis with convolution-based deep learning models. Two- and five-class classification was performed using state of the art models VGG16, InceptionResNetV2, ResNet152V2, EfficientNetB0, MobileNetV2, which are frequently preferred in the classification of medical images in the literature. Since the amount of data in the five-class classification in diabetic retinopathy disease images was not equal, the data were equalised by using data augmentation techniques using the albumentations library in the training dataset. Among the state of the art models used in the two-class classification, VGG16 was the best model since its accuracy, precision, sensitivity and f1-score metric values were 0.97. Among the models used in five-class classification, VGG16 was the best model due to its accuracy, precision, sensitivity and f1-score metric values of 0.78 and precision 0.79.