Scholarlink

<p>Semi-supervised learning (SSL) has received more attention in medical image segmentation as it helps alleviate the heavy labeling burden. However, effectively exploiting the relationship between labeled and unlabeled data to generate reliable pseudo-labels for the unlabeled data is still challenging. While foundational models like the Segment Anything Model (SAM) have shown their strong generalization capabilities, the application of these strengths to SSL within the medical field requires further exploration. In this paper, we propose a novel Cross-sample Consistency Learning framework for semi-supervised medical image segmentation (termed <b>CCL-Seg</b>) that fully exploits useful sample-based correlation and conducts information conversion between labeled and unlabeled data to boost the segmentation performance. Specifically, we present a Sample-based Correlation Learning strategy that promotes mutual learning between labeled and unlabeled data, resulting in a more discriminative and compact feature representation for each class. Within the SCL framework, we develop a Global Consistency Preservation module to harness global correlations, ensuring that features from the same class are consistent throughout the dataset. Furthermore, we present a SAM-guided Pseudo-label Hardness Weighting (PHW) supervision strategy that generates high-quality pseudo-labels, with supervision intensified at the pixel level. Extensive results demonstrate the effectiveness of our model against other state-of-the-art semi-supervised segmentation methods. More significantly, our model demonstrates a distinct advantage in achieving superior segmentation performance even with extremely limited annotations. The implementation code is released at <a href="https://github.com/taozh2017/CCLSeg">https://github.com/taozh2017/CCLSeg</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

Cross-sample Consistency Learning for Semi-supervised Medical Image Segmentation

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Tao Zhou,
  • Yunqi Gu,
  • Kaiwen Huang,
  • Huazhu Fu,
  • Xiao-jun Wu,
  • Josef Kittler

摘要

Semi-supervised learning (SSL) has received more attention in medical image segmentation as it helps alleviate the heavy labeling burden. However, effectively exploiting the relationship between labeled and unlabeled data to generate reliable pseudo-labels for the unlabeled data is still challenging. While foundational models like the Segment Anything Model (SAM) have shown their strong generalization capabilities, the application of these strengths to SSL within the medical field requires further exploration. In this paper, we propose a novel Cross-sample Consistency Learning framework for semi-supervised medical image segmentation (termed CCL-Seg) that fully exploits useful sample-based correlation and conducts information conversion between labeled and unlabeled data to boost the segmentation performance. Specifically, we present a Sample-based Correlation Learning strategy that promotes mutual learning between labeled and unlabeled data, resulting in a more discriminative and compact feature representation for each class. Within the SCL framework, we develop a Global Consistency Preservation module to harness global correlations, ensuring that features from the same class are consistent throughout the dataset. Furthermore, we present a SAM-guided Pseudo-label Hardness Weighting (PHW) supervision strategy that generates high-quality pseudo-labels, with supervision intensified at the pixel level. Extensive results demonstrate the effectiveness of our model against other state-of-the-art semi-supervised segmentation methods. More significantly, our model demonstrates a distinct advantage in achieving superior segmentation performance even with extremely limited annotations. The implementation code is released at https://github.com/taozh2017/CCLSeg.