Diabetic Retinopathy Detection with Vision Transformer based Deep Learning Model

Diabetic Retinopathy (DR), is one of the most common causes of blindness. It is required to have early and precise recognition for optimum cure. Convolutional Neural Networks (CNNs), often struggle with capturing fine-grained retinal abnormalities and handling noise in medical images. To deal with these limitations, this paper suggests innovative DR recognizing framework that integrates Vision Transformers (ViTs), Capsule Networks, BM3D filtering, and CLAHE enhancement. This approach enhances retinal images using BM3D (Block-Matching and 3D Filtering) to remove noise while preserving crucial details. CLAHE (Contrast Limited Adaptive Histogram Equalization) further improves image contrast, making early-stage DR features more distinguishable. Unlike conventional CNNs, which may lose spatial information, Capsule Networks retinal hierarchical relationships between image features, reducing misclassification. Meanwhile, Vision Transformers (ViTs) provide a global perspective by using retinal pictures to preserve long rangerelationships, improving this model's ability to differentiate between DR severity levels. This model evaluates on benchmark diabetic retinopathy datasets, achieving 98% training accuracy and 84% test accuracy. While the high training accuracy demonstrates the model's ability to learn complex patterns, test performance indicates a need for further improvements in generalization. These findings pave the way for future optimizations, including larger datasets and additional regularization techniques, to further refine the model's generalization and clinical applicability.