Hybrid Deformable and Basic Convolution Network for Fine Retinal Vessel Segmentation

Retinal vessel segmentation is an important task in computer-aided diagnosis for ophthalmic and cardiovascular diseases. However, conventional U-Net and its variants are limited by fixed-shape convolution kernels, leading to incomplete segmentation of tiny capillaries, discontinuous vessel structures, and failed detection of low-contrast thin vessels. To address these issues, we propose Hyde-net, an enhanced deformable encoder-decoder network for retinal vessel segmentation. Hydenet is built on the standard U-Net framework. A selective hybrid convolution scheme is adopted to rationally integrate deformable convolution (DCNv2) only at critical down-sampling and up-sampling steps. Experiments on the public DRIVE and CHASE_DB1 retinal vessel datasets demonstrate that Hyde-net achieves robust segmentation performance.