DG-SConv-LKRF: A Retinal Vessel Segmentation Method for Connectivity Modeling of Elongated Structures
Year: 2026
Type: conference-paper
Abstract: Retinal vessel segmentation is a fundamental task in ophthalmic computer-aided diagnosis. However, the elongated and tortuous morphology of vessels, together with large-scale variations and frequent discontinuities, poses significant challenges to connectivity modeling and fine-detail recovery. This paper proposes a lightweight segmentation framework, termed DG-SConv-LKRF, for connectivity-aware modeling of elongated structures. In the encoding stage, a Direction-Gated Dynamic Snake Convolution is introduced to enhance geometric continuity representations before downsampling via direction-adaptive structural responses. In the decoding stage, a Large-Kernel Skip Refinement Fusion module is designed to refine and repair skip-connection features by leveraging multi-scale large-kernel contextual information. Experimental results on the DRIVE dataset demonstrate that the proposed method outperforms U-Net, U-Net++, and Attention U-Net in terms of Dice and mIoU, and is able to better preserve the continuity and integrity of thin retinal vessels.
Authors: Jingjing Ma, Chaoying Tan, Zhiqiang Guo
Institution: Sichuan Vocational College of Finance and Economics
Language: English
Topic: Retinal Imaging and Analysis
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7212246723