Comparative analysis of diabetic retinopathy classification: methods & insights

Accurate and reliable diagnosis of diabetic retinopathy (DR) is still very difficult and problematic, manual image assessment is often hampered by variability between different observers and is very time intensive. Nevertheless, extensive image datasets can be analyzed with great precision and accuracy using Hybrid Vision Transformers (H-ViT). The deep learning architecture H-ViT is built with Convolutional Neural Networks (CNNs) which are prominent in vision tasks that include classification, detection, and segmentation, and Vision Transformers (ViTs), a new architecture inspired by significant breakthroughs in natural language processing. Moreover, ViTs have a unique method of breaking up images into patches and applying self-attention to perceive the entire image. This enables the model to capture and learn the intricate patterns found in medical images. The observed accuracy, precision and F1 score are 91.5%, 91.1% and 90.9%, respectively. The results showed that H-ViT holds great promise for future advancement in terms of automating DR diagnosis and screening.