Optimal Detection of Diabetic Retinopathy Severity Levels Using Attention-Based CNN and Vision Transformers (ViT)

Diagnosing diabetic retinopathy (DR) from colour fundus images is challenging as it requires skilled clinicians to recognize and interpret multiple small features and utilize a complex grading system, making the process both difficult and time-consuming. Prior detection of this condition is important to prevent any further complications such as partial or complete blindness. This study presents an approach for detecting the severity levels of diabetic retinopathy using a combination of Attention-Based Convolutional Neural Networks (CNN) and Vision Transformers (ViT). The method addresses the need for accurate classification of retinal images across multiple severity levels of diabetic retinopathy. We made use of a unique combination of attention mechanisms in CNN and transformers, thereby enhancing the model's ability to focus on critical regions within the retinal images. Experimental results show improvements in classification accuracy of around $95 \%$, sensitivity, specificity, and other key performance metrics, making it a potentially better solution for automated diabetic retinopathy detection.