Deep learning-based detection of diabetic retinopathy stages

In this master's thesis, a deep learning-based model was developed for the automatic classification of the five stages of diabetic retinopathy (DR). The study utilized retinal images from the APTOS 2019 dataset and employed eight different deep learning architectures—MobileNetV2, DenseNet169, NASNet, VGG19, ConvNeXt Large, EfficientNetB5, EfficientNetB6, and EfficientNetV2-S—trained using the transfer learning approach. Image processing and data augmentation techniques were applied to mitigate class imbalance issues and enhance model performance. Among the evaluated models, EfficientNetV2-S achieved the highest performance, which was further improved through fine-tuning. During the testing phase, the proposed model achieved 92.24% accuracy, 96.77% specificity, 92.28% sensitivity, 92.07% F1-score and 98.99% AUC by employing test-time augmentation. The findings indicate that the developed system can serve as an effective computer-aided diagnostic tool for the early detection of diabetic retinopathy.