Deep Hybrid Vision System for Automated Clinical Grading of Diabetic Retinopathy

DR is one of the most significant problems of diabetes and the largest causes of blindness in the world, so the correct assessment of its stages, Mild, Moderate, Severe and Proliferative, is critical in terms of ensuring timely treatment. The word manual grading of fundus images is labor intensive and subject to variability, which provides the need to rely on good automated systems. The proposed project will utilize a hybrid deep learning architecture that is based on a Vision Transformer (ViT) more than a ResNet to identify DR stages. ResNet takes the local features like the microaneurysms and hemorrhases whereas ViT uses the global retinal structure and allows the model to examine both the micro-level and the macro-level of the image. The system was also trained and tested on more than 500 images in one stage and preprocessing and data augmentation were used to improve the performance. The model was trained based on cross-entropy loss and Adam optimizer and had an accuracy of about 100 percent, indicating that there was little misclassification between stages. All these findings suggest that the proposed method can be an efficient automated DR severity grading tool that would assist ophthalmologists in performing early diagnosis, disease progression, and treatment plans.