Scholarlink

<p>Scarcity and high cost of annotated medical images have made semi-supervised learning a research hotspot in medical image segmentation. Existing methods often overlook fluctuations in pseudo-label quality, allowing low-quality pseudo-labels to interfere with model learning and degrade segmentation performance. This paper proposes a semi-supervised medical image segmentation method based on pseudo-label enhanced consistency learning. This method uses unlabeled data to generate pseudo-labels and improves the model robustness and accuracy by enhancing the consistency of the pseudo-labels. Specifically, it introduces a new data augmentation strategy in the model training process, which increases data diversity by introducing masks to ensure that the model outputs consistent segmentation results when processing the same unlabeled image that has undergone different transformations. In addition, voting-based pseudo-label refinement mechanism can effectively improve the label quality, thereby further improving the performance of the model. Experimental results on three public medical image segmentation datasets demonstrate that the proposed method achieves competitive and stable performance under different labeled-data settings. For example, with 20% labeled data, our method achieves Dice coefficients of 90.42% on the 2D ACDC dataset and 91.25% and 82.21% on the 3D LA and Pancreas datasets, respectively, while maintaining competitive boundary-related metrics. Code is available at: <a href="https://github.com/xm8587/PECL">https://github.com/xm8587/PECL</a><Emphasis Type="Underline">.</Emphasis></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

Pseudo-label enhanced consistency learning for semi-supervised medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Mei Xiao,
  • Qingshan She,
  • Songkai Sun,
  • Xiaofei Zhou,
  • Yingchun Zhang

摘要

Scarcity and high cost of annotated medical images have made semi-supervised learning a research hotspot in medical image segmentation. Existing methods often overlook fluctuations in pseudo-label quality, allowing low-quality pseudo-labels to interfere with model learning and degrade segmentation performance. This paper proposes a semi-supervised medical image segmentation method based on pseudo-label enhanced consistency learning. This method uses unlabeled data to generate pseudo-labels and improves the model robustness and accuracy by enhancing the consistency of the pseudo-labels. Specifically, it introduces a new data augmentation strategy in the model training process, which increases data diversity by introducing masks to ensure that the model outputs consistent segmentation results when processing the same unlabeled image that has undergone different transformations. In addition, voting-based pseudo-label refinement mechanism can effectively improve the label quality, thereby further improving the performance of the model. Experimental results on three public medical image segmentation datasets demonstrate that the proposed method achieves competitive and stable performance under different labeled-data settings. For example, with 20% labeled data, our method achieves Dice coefficients of 90.42% on the 2D ACDC dataset and 91.25% and 82.21% on the 3D LA and Pancreas datasets, respectively, while maintaining competitive boundary-related metrics. Code is available at: https://github.com/xm8587/PECL.