The Implementation of Ensemble Voting in Deep Neural Networks for the Automated Classification of Diabetic Retinopathy

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Yıldız Teknik Üniversitesi, Fen Bilimleri Enstitüsü, --------------------, Türkiye

Tezin Onay Tarihi: 2024

Tezin Dili: İngilizce

Öğrenci: MOTHNA MEZHER ALRUBAYE

Danışman: Hamza Osman Ilhan

Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu

Özet:

Diabetic retinopathy is one of the most dangerous symptoms resulting from human infection with diabetes, as it develops in most people with this disease, so it is necessary to detect it early to treat it, or to slow the disease to prevent loss of vision. This thesis aims to investigate the automatic methods that contribute to the detection of (DR) disease. According to the images taken of the fundus of the patient's eye, the classification was made into five categories (no DR, moderate DR, mild DR, proliferative DR, severe DR). The difficulty of dealing with this image made the efficiency of deep learning techniques one of the best solutions to facilitate the classification process. In this thesis, three deep learning models namely (ResNet50, Densenet201, and InceptionV3) were used in the image classification of the APTOS 2019 dataset. Two approaches have been proposed in the classification process. In the first approach the individual experiments of the models, using transfer learning with fine tuning and fusion ideas were implemented. The highest classification performance was obtained with the idea of soft voting, where the accuracy of the original image data set was obtained by 85%. In the second approach, to increase the classification performance, we used a balancing technique based on oversampling and augmentation operations performed on the original APTOS 2019 dataset. The highest classification performance was obtained with the implementation of the idea of soft voting over the models increased by 90%.