Scholarlink

<p>Retinal vessel segmentation remains challenging due to the presence of fine vessel structures, complex backgrounds, and vascular continuity preservation. Although deep learning has achieved promising performance, existing methods still struggle to effectively integrate local detail with global contextual information, which may lead to missed fine vessels and fragmented predictions. So we propose CR-Net, a dual-branch network for retinal vessel segmentation. CR-Net consists of two complementary branches for local detail recovery and global context modeling, respectively, and introduces a cross-branch interaction strategy to enhance the joint segmentation of fine capillaries and major vascular structures. By strengthening local-global feature collaboration, the proposed network improves the delineation of thin vessels and suppresses interference from complex backgrounds. Experiments on three public datasets, DRIVE, STARE, and CHASEDB1, demonstrate that CR-Net achieves competitive performance compared with existing methods. CR-Net achieves ACC/SE/AUC/Dice of 0.9820/0.7949/0.9791/0.8191 on DRIVE, ACC/SE/AUC/Dice of 0.9899/0.8351/0.9897/0.8747 on STARE, and ACC/SE/AUC/Dice of 0.9936/0.7188/0.9758/0.6592 on CHASEDB1. The results indicate that CR-Net is effective for retinal vessel segmentation, particularly in capturing fine vessel structures while maintaining structural consistency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A dual-branch network with cross-interaction for retinal vessel segmentation

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • Shuai Huang,
  • Jihui Mao,
  • Minshan Jiang

摘要

Retinal vessel segmentation remains challenging due to the presence of fine vessel structures, complex backgrounds, and vascular continuity preservation. Although deep learning has achieved promising performance, existing methods still struggle to effectively integrate local detail with global contextual information, which may lead to missed fine vessels and fragmented predictions. So we propose CR-Net, a dual-branch network for retinal vessel segmentation. CR-Net consists of two complementary branches for local detail recovery and global context modeling, respectively, and introduces a cross-branch interaction strategy to enhance the joint segmentation of fine capillaries and major vascular structures. By strengthening local-global feature collaboration, the proposed network improves the delineation of thin vessels and suppresses interference from complex backgrounds. Experiments on three public datasets, DRIVE, STARE, and CHASEDB1, demonstrate that CR-Net achieves competitive performance compared with existing methods. CR-Net achieves ACC/SE/AUC/Dice of 0.9820/0.7949/0.9791/0.8191 on DRIVE, ACC/SE/AUC/Dice of 0.9899/0.8351/0.9897/0.8747 on STARE, and ACC/SE/AUC/Dice of 0.9936/0.7188/0.9758/0.6592 on CHASEDB1. The results indicate that CR-Net is effective for retinal vessel segmentation, particularly in capturing fine vessel structures while maintaining structural consistency.