Diabetic Retinopathy Detection and Classification Using Pre-Trained CNN Architectures and Vision Transformers

Diabetic retinopathy (DR) is a significant complication of diabetes and a leading cause of vision impairment globally. Early detection and treatment are crucial for preventing severe consequences. However, manual screening is time-consuming and resource-intensive. Automated solutions employing deep learning have demonstrated potential in mitigating these challenges by facilitating accurate, efficient, and scalable DR classification. This study presents a robust ensemble framework that combines pre-trained deep learning models, such as EfficientNet-B1, VGG19, ResNet-50, Inception v3, DenseNet-121, and ViT-B/16, to classify retinal fundus images into multiple stages of DR. The dataset is preprocessed using resizing, normalization, and augmentation techniques to ensure model generalization and address class imbalance. This approach underscores the potential for deploying ensemble-based deep learning systems in practical DR screening applications, offering a scalable and impactful solution for healthcare.