Automated Diabetic Retinopathy Detection with FastDRNet on OCT Imaging

One of the main causes of vision loss in the world is diabetic retinopathy (DR), emphasizing the need for timely detection to enable effective management. This study introduces FastDRNet, a novel automated approach for DR classification using Optical Coherence Tomography (OCT) images. FastDRNet leverages transfer learning with the FastAI framework and utilizes a ResNet18 architecture pretrained on ImageNet. The model integrates advanced data augmentation techniques such as flipping, rotation, and zooming to enhance generalization. Employing cyclical learning rates and progressive fine-tuning, FastDRNet demonstrated exceptional performance on a balanced OCT dataset. Key macro-averaged metrics, including precision (0.9421), recall (0.9366), and F1-score (0.9393) underscore the model’s effectiveness. Additionally, the Matthews Correlation Coefficient (0.9368) and Cohen’s Kappa score (0.9367) further validate its reliability and robustness. These findings highlight FastDRNet’s potential as a reliable and scalable method for early diabetic retinopathy detection, which will help with prompt diagnosis and treatment planning.