Diabetic retinopathy (DR) is an eye complication associated with diabetes, resulting in blurred vision or blindness. The early diagnosis and treatment of DR can decrease the risk of vision loss dramatically.However, such diagnosis is a tedious and complicated task due to the variability of retinal changes across the stages of the diseases, and due to the high number of undiagnosed and untreated DR cases. In this paper, we develop a computationally efficient and scalable deep learning model using convolutional neural networks (CNN), for diagnosing DR automatically. Various preprocessing algorithms are utilized to improve accuracy, and a transfer learning strategy is adopted to speed up the process. Our experiment used the fundus image set available on online Kaggle datasets. As an ultimate conclusion of applicable performance metrics, our computational simulation achieved a relatively-high F1 score of 93.2% for stage-based DR classification. Povzetek: Opisana je metoda globokih nevronskih mrež za diagnozo težav vida zaradi sladkorne bolezni.

Diabetic retinopathy (DR) is a progressive eye disease associated with diabetes, resulting in blindness or blurred vision. The risk of vision loss was dramatically decreased with early diagnosis and treatment. Doctors diagnose DR by examining the fundus retinal images to develop lesions associated with the disease. However, this diagnosis is a tedious and challenging task due to growing undiagnosed and untreated DR cases and the variability of retinal changes across disease stages. Manually analyzing the images has become an expensive and time-consuming task, not to mention that training new specialists takes time and requires daily practice. Our work investigates deep learning methods, particularly convolutional neural network (CNN), for DR diagnosis in the disease's five stages. A pre-trained residual neural network (ResNet-34) was trained and tested for DR. Then, we develop computationally efficient and scalable methods after modifying a ResNet-34 with three additional residual units as a novel ResNet-n/DR. The Asia Pacific Tele-Ophthalmology Society (APTOS) 2019 dataset was used to evaluate the performance of models after applying multiple pre-processing steps to eliminate image noise and improve color contrast, thereby increasing efficiency. Our findings achieved state-of-the-art results compared to previous studies that used the same dataset. It had 90.7% sensitivity, 93.5% accuracy, 98.2% specificity, 89.5% precision, and 90.1% F1 score.