Scholarlink

<p>Medical image segmentation is crucial for accurate disease diagnosis, treatment planning, and outcome monitoring. Traditional fully supervised segmentation methods require extensive labeled data, which is often limited and costly to acquire in medical applications. To address this, we propose a novel semi-supervised medical image segmentation approach that effectively utilizes a small amount of labeled data to segment a large volume of unlabeled data. Our method leverages label-driven spatial consistency category swapping, which preserves the relative positions of foreground and background while enriching their information. This technique exposes the model to various backgrounds of the same foreground object, mitigating over-reliance on specific background features. Additionally, we incorporate feature perturbation consistency training into the network structure to enhance the model’s robustness. Experimental results on four publicly available datasets demonstrate that our method outperforms existing semi-supervised segmentation approaches, achieving superior segmentation performance. Code is publicly available at <a href="https://github.com/Yanzj621/SCCS">https://github.com/Yanzj621/SCCS</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

Semi-supervised medical image segmentation through label-driven space structure augmentation

  • 期刊论文

  • 发表日期：

    2025年5月29日

查看全文
  • Zhijun Yan,
  • Yonghong Hou,
  • Pengyu Zhao

摘要

Medical image segmentation is crucial for accurate disease diagnosis, treatment planning, and outcome monitoring. Traditional fully supervised segmentation methods require extensive labeled data, which is often limited and costly to acquire in medical applications. To address this, we propose a novel semi-supervised medical image segmentation approach that effectively utilizes a small amount of labeled data to segment a large volume of unlabeled data. Our method leverages label-driven spatial consistency category swapping, which preserves the relative positions of foreground and background while enriching their information. This technique exposes the model to various backgrounds of the same foreground object, mitigating over-reliance on specific background features. Additionally, we incorporate feature perturbation consistency training into the network structure to enhance the model’s robustness. Experimental results on four publicly available datasets demonstrate that our method outperforms existing semi-supervised segmentation approaches, achieving superior segmentation performance. Code is publicly available at https://github.com/Yanzj621/SCCS.