Diabetic retinopathy detection enhancement using deep learning algorithms on hybrid dataset

Background Diabetic Retinopathy (DR) is the leading cause of blindness in diabetics. It occurs when high blood sugar levels cause blood vessels to become blocked. There is a need for a hybrid deep learning framework that can detect DR early with high accuracy and generalization. The goal was to achieve high accuracy on a multi-class hybrid dataset using deep learning. Methods This study developed a novel hybrid model combining EfficientNetB0 with a Vision Transformer (ViT), preprocessed using CLAHE and Gaussian blur. Results The hybrid framework achieved an accuracy of 97.23%, and precision, recall f1 score and an AUC of more than 97%. Conclusion The interpretability techniques provided visual explanations of retinal characteristics for class predictions. Altogether, the hybrid framework outperformed previous studies and is highly suitable for application in real clinical settings.