Retinal Vessel Segmentation Using an Attention-Enhanced U-Net Architecture

Abstract

Abstract The morphology and variations of retinal blood vessels are crucial indicators for diagnosing various ophthalmic diseases. Therefore, accurate segmentation of retinal vessels holds significant value in assisting clinical diagnosis. However, traditional image analysis methods often rely on manual annotations, which are susceptible to physicians’ experience and subjective judgment, thereby limiting segmentation efficiency and consistency. Developing high-performance retinal vessel segmentation techniques has thus become a key challenge. In this study, we propose RSA-UNet, a segmentation model based on the U-Net architecture that integrates ResNet50, a Spatial Channel Block Attention Module (SCBAM), and Atrous Spatial Pyramid Pooling (ASPP) to further enhance segmentation accuracy. The model was evaluated on the publicly available retinal fundus image dataset CHASEDB1. Experimental results demonstrate that RSA-UNet outperforms the traditional U-Net in both segmentation accuracy and stability, highlighting its potential for application in retinal vessel segmentation tasks.