Scholarlink

<p>Semantic segmentation is a fundamental task in medical image analysis, yet its performance heavily relies on large-scale annotated datasets, which are costly to obtain. Semi-supervised methods alleviate this issue but often suffer from insufficient utilization of unlabeled data and information loss during perturbation. To address these challenges, we propose OaCMatch, a novel weak-to-strong consistency framework that explicitly leverages raw unlabeled images. The key novelty of our method lies in two aspects: (1) an Unlabeled Information Constraint Perturbation (UICP) module that extracts attention from original unlabeled images to guide strong perturbations, improving pseudo-label reliability; and (2) a Joint Feature Perturbation (JFP) module that injects global feature information from raw images into the feature perturbation process, mitigating semantic loss and expanding the perturbation space. Extensive experiments on ACDC and LA datasets demonstrate that the proposed method consistently outperforms existing semi-supervised approaches across multiple evaluation metrics, highlighting the effectiveness of directly incorporating raw unlabeled image information into consistency learning.</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

OaCMatch: Semi-supervised medical image segmentation network based on unlabeled information constraints and joint feature perturbation

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Jianwu Long,
  • Wenlian Xie

摘要

Semantic segmentation is a fundamental task in medical image analysis, yet its performance heavily relies on large-scale annotated datasets, which are costly to obtain. Semi-supervised methods alleviate this issue but often suffer from insufficient utilization of unlabeled data and information loss during perturbation. To address these challenges, we propose OaCMatch, a novel weak-to-strong consistency framework that explicitly leverages raw unlabeled images. The key novelty of our method lies in two aspects: (1) an Unlabeled Information Constraint Perturbation (UICP) module that extracts attention from original unlabeled images to guide strong perturbations, improving pseudo-label reliability; and (2) a Joint Feature Perturbation (JFP) module that injects global feature information from raw images into the feature perturbation process, mitigating semantic loss and expanding the perturbation space. Extensive experiments on ACDC and LA datasets demonstrate that the proposed method consistently outperforms existing semi-supervised approaches across multiple evaluation metrics, highlighting the effectiveness of directly incorporating raw unlabeled image information into consistency learning.