Deep Effective Transfer Learning-Enabled Systems for Accurate Detection of DR (Diabetic Retinopathy) from Eye Images

One of the most common causes of blindness in humans is diabetic retinopathy. Reviewing DR pictures by hand is time-consuming and error-prone. Current approaches to illness detection frequently use manually-crafted characteristics, which severely restrict their versatility and classification accuracy. Adult-onset blindness is most commonly caused by diabetic retinopathy (DR), a serious microvascular consequence of diabetes mellitus. Traditional screening methods have limitations in scalability in resource-constrained situations due to their reliance on expert evaluation, time consumption, and manual nature. Therefore, early and precise identification of DR is crucial in preventing irreversible vision loss. Automated DR identification utilizing retinal fundus images is presented in this research using a system enabled by deep transfer learning. Using the publicly available Messidor-1 dataset—which includes 1200 annotated fundus images classified into four severity levels of DR—the suggested model makes use of the ResNet50 architecture. Improving data quality and balancing class distribution was achieved through the application of data augmentation techniques and a four-stage preprocessing pipeline that included picture cropping, denoising, histogram equalization, and scaling. Various classification jobs, including those with multiple classes, are supported by the framework. With a 99.37% accuracy rate, 98.62% precision, 99.75% recall, and an F1-score of 99.42%, the experimental findings show that the diagnostic performance is high. This research shows that deep learning could be a great tool for DR screening in hospitals and other healthcare facilities, both locally and remotely