Classification and Identification of DR from Fundus Images Based on Deep Convolutional Networks

Abstract

Abstract—A disease known as diabetic retinopathy (DR) can develop in people who have diabetes for an extended period of time.   Visual impairment can result from a postponed diagnosis. Diabetics are disproportionately likely to get DR due to their chronically elevated blood sugar levels. The retina's blood vessels are affected by this. This study demonstrates the use of the ResNet50 architecture in a deep learning-based method for the early detection and categorization of diabetic retinopathy (DR) using images of the retinal fundus. This research takes advantage of fundus photography, a non-invasive, high-resolution imaging technology, to detect retinal alterations even when no outward signs of diabetic retinopathy (DR) are present. Diabetes is on the rise around the world, and if not caught early, DR can lead to permanent visual loss, thus this is crucial.  The work guarantees strong training of the ResNet50 model by preprocessing images using normalization, augmentation, and scaling, and by controlling for class imbalances.  The APTOS dataset includes photos from all five severity levels of DR. The model demonstrated outstanding results in terms of recall, accuracy, precision, and F1-score during training, suggesting high reliability and promising clinical use.  Aiming to improve preventive diabetes treatment, particularly in places with limited resources, the research highlights the usefulness of AI in scalable, early-stage DR screening.