Scholarlink

<p>Accurate retinal vessel segmentation is crucial for the early diagnosis and screening of fundus diseases. However, many existing methods rely mainly on spatial-domain features, while frequency-domain information and multi-scale vascular structures are insufficiently explored. High-frequency components emphasize vessel edges and thin structures, while low-frequency suppression reduces background interference. Multi-scale representations are essential as retinal vessels range from major vessels (8–12 pixels) to fine capillaries (2–4 pixels). To address these challenges, we propose a Multi-scale Frequency–Spatial Hybrid Progressive Fusion Network (MFSH-Net). The encoder employs a Frequency-Spatial Selective Hybrid module for adaptive frequency filtering, a Saliency-Aware Topology Attention Routing module for long-range vessel dependencies, and a Hierarchical Scale-Aware Attention module that replaces simple skip connection concatenation with attention-weighted fusion. Experiments on DRIVE, CHASE_DB1, and STARE demonstrate that MFSH-Net achieves superior performance in noise suppression and fine vessel detection.</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

Multi-scale frequency-spatial hybrid progressive fusion network for retinal vessel segmentation

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Lingling Kan,
  • Yuan Zhang,
  • Hai Zhao,
  • Jingzhe Yan,
  • Cong Li

摘要

Accurate retinal vessel segmentation is crucial for the early diagnosis and screening of fundus diseases. However, many existing methods rely mainly on spatial-domain features, while frequency-domain information and multi-scale vascular structures are insufficiently explored. High-frequency components emphasize vessel edges and thin structures, while low-frequency suppression reduces background interference. Multi-scale representations are essential as retinal vessels range from major vessels (8–12 pixels) to fine capillaries (2–4 pixels). To address these challenges, we propose a Multi-scale Frequency–Spatial Hybrid Progressive Fusion Network (MFSH-Net). The encoder employs a Frequency-Spatial Selective Hybrid module for adaptive frequency filtering, a Saliency-Aware Topology Attention Routing module for long-range vessel dependencies, and a Hierarchical Scale-Aware Attention module that replaces simple skip connection concatenation with attention-weighted fusion. Experiments on DRIVE, CHASE_DB1, and STARE demonstrate that MFSH-Net achieves superior performance in noise suppression and fine vessel detection.