Scholarlink

<p>Retinal vessel segmentation is a crucial step in analyzing fundus images and plays a vital role in the early detection, diagnosis, and treatment of various diseases. To make the segmentation model more applicable to actual medical scenarios, A Lightweight Multi-scale Feature Refinement Network (LMFR-Net) based on dual-decoding structure is proposed for efficient retinal vessel segmentation. Using a dual-decoding structure to reduce information loss, an Improved Convolution Block (ICB) is proposed to enhance the ability to extract basic features. In addition, a Lightweight Multi-scale Attention Feature Fusion (LMAFF) module is designed to extract the multi-scale spatial structure features. A Feature Refinement Module (FRM) with dense connections is proposed to optimize detailed features and comprehensively improve network segmentation capability. Comparative experiments were conducted on the DRIVE, CHASEDB1, and STARE datasets to verify that LMFR-Net achieved the highest F1-score and Recall of 82.91% and 86.85%, respectively, with only 366kb of parameters. More refined segmentation results have also been achieved in the visualization comparison of segmented images, and the overall segmentation effect is well. This indicates that LMFR-Net achieves efficient retinal vessel segmentation with a significantly reduced computational complexity, making it well-suited for practical medical applications. The code is available at <a href="https://github.com/MCloud31/LMFR-Net">https://github.com/MCloud31/LMFR-Net</a>.</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

LMFR-Net: lightweight multi-scale feature refinement network for retinal vessel segmentation

  • 期刊论文

  • 发表日期：

    2025年2月13日

查看全文
  • WenHao Zhang,
  • ShaoJun Qu,
  • YueWen Feng

摘要

Retinal vessel segmentation is a crucial step in analyzing fundus images and plays a vital role in the early detection, diagnosis, and treatment of various diseases. To make the segmentation model more applicable to actual medical scenarios, A Lightweight Multi-scale Feature Refinement Network (LMFR-Net) based on dual-decoding structure is proposed for efficient retinal vessel segmentation. Using a dual-decoding structure to reduce information loss, an Improved Convolution Block (ICB) is proposed to enhance the ability to extract basic features. In addition, a Lightweight Multi-scale Attention Feature Fusion (LMAFF) module is designed to extract the multi-scale spatial structure features. A Feature Refinement Module (FRM) with dense connections is proposed to optimize detailed features and comprehensively improve network segmentation capability. Comparative experiments were conducted on the DRIVE, CHASEDB1, and STARE datasets to verify that LMFR-Net achieved the highest F1-score and Recall of 82.91% and 86.85%, respectively, with only 366kb of parameters. More refined segmentation results have also been achieved in the visualization comparison of segmented images, and the overall segmentation effect is well. This indicates that LMFR-Net achieves efficient retinal vessel segmentation with a significantly reduced computational complexity, making it well-suited for practical medical applications. The code is available at https://github.com/MCloud31/LMFR-Net.