Scholarlink

<p>Medical image segmentation often faces the challenge of limited annotated data, particularly in scenarios requiring fine-grained anatomical precision. While recent semi-supervised methods mitigate this challenge by leveraging unlabeled data, they often struggle with under-explored structural semantics and insufficient cross-level feature consistency. Inspired by the observation that feature channels within a layer encode diverse yet complementary semantics, and that shallow and deep features exhibit inherent asymmetry, we propose StruBAR, a novel semi-supervised segmentation framework that contains three key modules: Structure-aware Channel Exchange (SCE) for task-driven perturbation, Bidirectional Deep-Shallow Alignment (BiDSA) for improved cross-layer consistency, and Soft-Boundary Guided Reconstruction (SBG-Rec) for enhanced semantic boundary recovery. Through extensive experiments on three public benchmark datasets, we demonstrate that StruBAR achieves significant improvements in both boundary and region-level segmentation metrics, outperforming most existing semi-supervised approaches. Our results validate the effectiveness of integrating structural semantics and feature alignment to improve segmentation performance under limited supervision. Our code will be available at <a href="https://github.com/UniverseHappiness/StruBAR.git">https://github.com/UniverseHappiness/StruBAR.git</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

StruBAR: a structure-aware consistency framework via bi-directional alignment and boundary reconstruction for semi-supervised medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • Xinghua Wu,
  • Qin Zhang,
  • Xiaodong Zhang

摘要

Medical image segmentation often faces the challenge of limited annotated data, particularly in scenarios requiring fine-grained anatomical precision. While recent semi-supervised methods mitigate this challenge by leveraging unlabeled data, they often struggle with under-explored structural semantics and insufficient cross-level feature consistency. Inspired by the observation that feature channels within a layer encode diverse yet complementary semantics, and that shallow and deep features exhibit inherent asymmetry, we propose StruBAR, a novel semi-supervised segmentation framework that contains three key modules: Structure-aware Channel Exchange (SCE) for task-driven perturbation, Bidirectional Deep-Shallow Alignment (BiDSA) for improved cross-layer consistency, and Soft-Boundary Guided Reconstruction (SBG-Rec) for enhanced semantic boundary recovery. Through extensive experiments on three public benchmark datasets, we demonstrate that StruBAR achieves significant improvements in both boundary and region-level segmentation metrics, outperforming most existing semi-supervised approaches. Our results validate the effectiveness of integrating structural semantics and feature alignment to improve segmentation performance under limited supervision. Our code will be available at https://github.com/UniverseHappiness/StruBAR.git.