Comparative Analysis of Retinal Vessel Segmentation Utilising Convolutional and Transformer-Based Architectures

Abstract

Retinal vessel segmentation plays a major role in identifying and screening eye-related diseases such as glaucoma, diabetic retinopathy, and microaneurysm. Utilising the automatic feature extraction capabilities of deep learning models, there is a potential to reduce the reliance on human expertise for disease diagnosis, enabling physicians to make faster and more efficient diagnoses. This study focuses on evaluating the effect of uncomplicated data augmentation techniques on retinal vessel segmentation, offering a comparative analysis across eight diverse Convolutional-based and transformer-based architectures, including UNet, Attention-UNet, Res-UNet, UNet++, UCTransNet, TransUNet, UNeXt and SwinUNet. Each model is trained on four publicly available datasets with different resolutions, from low-res to high-res images, including $DRIVE$, $STARE$,$ CHASE-DB1$, and $HRF$. Subsequently, the study provides comparative results, offering insight into the efficiency of these models in retinal vessel segmentation. Obtained results indicate that even though the overall difference between models' performance is small, convolutional models perform better than transformers when trained on lower-resolution datasets (i.e. DRIVE dataset), while transformers are best on Higher-resolution datasets (i.e. HRF).