Scholarlink

<p>Retinal vessel segmentation holds important clinical value in early screening and auxiliary diagnosis of fundus diseases. However, fundus images often suffer from a severe class imbalance problem, where tiny vessels occupy only a small proportion of pixels and are highly sensitive to noise interference and edge blurring. It hinders existing methods from accurately segmenting large vessels and microvessels simultaneously. To this end, a novel Multi-level Attention Feature Interaction Network (MAFI-Net) is proposed. To alleviate the imbalance at the training sample level, this paper proposes an automated coarse and fine vessel patch classification algorithm. In addition, a Multi-level Attention Guided Fusion (MAGF) module is designed to unify deep semantic cues with shallow spatial details, effectively mitigating the degradation of structure details during feature fusion. Meanwhile, a Multi-Constraint Guided Loss (MCGL) is proposed, which introduces constraints related to vessel skeleton and boundary information on top of traditional loss factors, thereby achieving more robust structural optimization. Experiments on DRIVE, STARE, and CHASE_DB1 show that MAFI-Net achieves competitive segmentation performance.</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

MAFI-Net: A Multi-level Attention Feature Interaction Network for Retinal Vessel Segmentation

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Xuelin Xu,
  • Jianwei Chen,
  • Bingfeng Ruan,
  • Huabin He

摘要

Retinal vessel segmentation holds important clinical value in early screening and auxiliary diagnosis of fundus diseases. However, fundus images often suffer from a severe class imbalance problem, where tiny vessels occupy only a small proportion of pixels and are highly sensitive to noise interference and edge blurring. It hinders existing methods from accurately segmenting large vessels and microvessels simultaneously. To this end, a novel Multi-level Attention Feature Interaction Network (MAFI-Net) is proposed. To alleviate the imbalance at the training sample level, this paper proposes an automated coarse and fine vessel patch classification algorithm. In addition, a Multi-level Attention Guided Fusion (MAGF) module is designed to unify deep semantic cues with shallow spatial details, effectively mitigating the degradation of structure details during feature fusion. Meanwhile, a Multi-Constraint Guided Loss (MCGL) is proposed, which introduces constraints related to vessel skeleton and boundary information on top of traditional loss factors, thereby achieving more robust structural optimization. Experiments on DRIVE, STARE, and CHASE_DB1 show that MAFI-Net achieves competitive segmentation performance.