A Deep Hybrid Approach for Automated Detection and Grading of Diabetic Retinopathy

Diabetic retinopathy (DR) is a serious global public health issue related to preventable blindness. Subsequently, timely and accurate evaluation of the condition is paramount. Herein, we propose a new deep learning architecture, HybridVisionNet, for the automated multi-class classification of diabetic retinopathy in fundus images. HybridVisionNet combines the representational capabilities of both ResNet50 and the Vision Transformer (ViT) architecture. It seamlessly integrates convolution and self-attention mechanisms to effectively harness the local and global feature representation powers. The model was trained using a balanced dataset achieved via random oversampling techniques and was subsequently fine-tuned for performance using sophisticated strategies to promote generalized performance, which included resizing and normalization, flipping, and color jittering. In the five-class diabetic retinopathy classification task, HybridVisionNet demonstrated consistent superiority across all classification tasks, achieving an average precision (0.9355), recall (0.9360), F1 score (0.9352), and accuracy (0.9360). Hy-bridVisionNet also attained high scores of 0.9201 and 0.9198, respectively, which, when coupled with the near 1.0 ROC-AUC scores across all stages of diabetic retinopathy, indicated the high, robust performance of the model. The findings indicate that HybridVisionNet is a dependable and time-saving approach for the early detection and staging of diabetic retinopathy, which may assist in clinical decision-making, along with improving patient care and prognosis.