DCHA-Net: Dynamic convolution and hybrid attention network for retinal vessel segmentation

Abstract

Changes in retinal vasculature morphology can lead to ophthalmic and cardiovascular issues, making accurate vessel segmentation crucial for diagnosing related diseases. However, this task remains challenging due to the wide range of variations in retinal vessels and the low contrast between vessels and the background. Convolutional neural networks (CNNs) have performed excellently in this task. Still, standard convolutions’ limited receptive field struggles with retinal vessel variations, and inadequate use of local context at skip connections affects accuracy. To address the challenges, We propose DCHA-Net, a retinal vessel segmentation network using dynamic convolution and attention mechanisms. It employs a convolutional fusion module to focus on vessel topology and a hybrid attention module to enhance high-level feature details. Additionally, a multi-scale supervision fusion module aggregates features at different decoder scales, reducing information loss during upsampling. Experiments on DRIVE, CHASE_DB1, and STARE datasets show DCHA-Net achieves segmentation accuracy of 96.86 $\mathrm{\%}$ , 97.45 $\mathrm{\%}$ , and 97.49 $\mathrm{\%}$ , with specificity of 98.24 $\mathrm{\%}$ , 98.55 $\mathrm{\%}$ , and 99.27 $\mathrm{\%}$ . These results demonstrate that our method outperforms existing ones in performance and segmentation accuracy.