Abstract
One common consequence of diabetes mellitus is DR, which destroys the retina and can cause permanent blindness if not caught in time. Treatment cannot undo DR, but catching it early and taking steps to prevent further vision loss can help keep eyesight intact. The necessity for trustworthy automated methods is highlighted by the fact that conventional diagnosis, which involves ophthalmologists utilising retinal fundus pictures, is time-consuming, expensive, and susceptible to human mistake. In this research, this survey current deep learning-based techniques for diabetic retinopathy detection and present DualAttTrans, a full-stack network that uses hybrid attention mechanisms to efficiently capture and analyse retinal image features on different scales and domains, without using additional priors or supervision. In order to get representative features for classification, the methodology begins with preprocessing to clean the data. Then, features are extracted. This is built a large-scale benchmark dataset with various retinal pictures at different DR severity levels to validate performance. The experimental results show that DualAttTrans outperforms the state-of-the-art methods with an impressive accuracy of 94.32%. The results show that DualAttTrans is an effective, efficient, and scalable technique for automated diabetic retinopathy detection; it can aid in early screening and protect at-risk patients' eyes from damage. https://doi.org/10.65470/james.2