A Deep Learning Framework for the Identification of Distinct Stages in Diabetic Retinopathy through Retinal Image Analysis

Diabetic retinopathy is the leading cause of blindness among diabetic patients. It occurs when the light-sensitive tissue in the retina is damaged. According to WHO statistics, the current global prevalence of diabetic retinopathy is approximately 103 million and is expected to be 161 million by 2045. The existing diagnosis methods are enriched with advanced retinal image processing and feature extraction techniques are frequently used to detect the presence of diabetic retinopathy in patients rather than focusing on early-stage detection. Therefore, detecting diabetic retinopathy at an early stage is crucial to prevent severe complications. The current study has explored the feasibility of identifying different stages of diabetic retinopathy using retinal images analyzed through deep learning algorithms. Labeled data corresponding to the five stages of diabetic retinopathy from the APTOS 2019 blindness detection dataset were preprocessed and a transfer learning approach was applied, utilizing pre-trained models from ImageNet for training. Among the transfer learning algorithms, the neural networks of ResNet_101, DenseNet_201 and EfficientNet_b0 were selected to train three DL models to select the optimal model based on test accuracy. The proposed method achieved a test accuracy of 91% using the fine-tuned EfficientNet_b0 model.