Hybrid CNN and Vision Transformer Model for Diabetic Retinopathy Detection

Diabetic Retinopathy is one of the most common complications of Diabetes Mellitus, leading to visual impairment in millions, and permanent blindness if not detected early. Compared to computer-aided diagnosis systems, manual detection of DR is a time-consuming process, cumbersome, and expensive. Deep learning is recently emerging as a powerful tool for automated diagnosis, with CNNs proving very effective in the medical domain. Nonetheless, the ViTs have been studied for their ability to inject global image context, potentially yielding an advantage in DR detection. This paper proposes a new hybrid model combining CNNs and ViTs for classification of DR severity from color fundus images. The performance of this hybrid model, achieving accuracy of 85.52 percent, recall of 85.52 percent, precision of 85.20 percent, and F1-score of 84.92 percent is assessed through exhaustive experimentation and benchmarked against standalone CNN and ViT models.