Deep learning based classification for diabetic retinopathy diagnosis

Diabetic Retinopathy (DR) is a complication that can cause vision loss and blindness in people with diabetes, affecting the blood vessels of the retina, the light-sensitive tissue layer at the back of the eye. In the early stages of DR, the walls of the blood vessels in the retina weaken and protrude. As a result, the tissue swells or leaks blood. This condition is detected by imaging devices and medical examination. In this thesis, we present an approach to achieve the highest accuracy in the diagnosis of DR using convolutional neural network (CNN) architectures. For the detection of DR disease with ESA models, the APTOS 2019 dataset with five classes, which is widely used and up to date in the literature, was preferred. To achieve high classification accuracy, the original dataset was transformed into three different datasets with datasets created using Grayscale and Gaussian filters. The three different datasets were trained with the same optimal parameters with the image-based ESA architectures ResNET152V2, ResNet101V2, MobileNet, MobileNetV2, DenseNet169, and NASNetLarge, and the test results were evaluated. The models were applied to five different classifications No DR, Mild, Moderate, Severe, and Proliferative DR, and accuracy, precision, sensitivity, and f1 score metrics were used to evaluate the classification performance. When the results obtained are evaluated, the NasNetLarge model achieved the highest accuracy values with 98.96% in training on the original APTOS 2019 dataset, the MobileNet V2 model with 99.26% in the Grayscale filtered dataset and MobileNet V2 with 99.54% in the Gaussian filtered model. It is seen that the filters applied to the original dataset contributed to the increase in classification accuracy. As a result, it is seen that the use of the MobileNetV2 model in the Gaussian filtered dataset for the diagnosis of DR disease provides very high accuracy in disease diagnosis.