Scholarlink

<p>Semi-supervised learning (SSL) is a promising high-performance computing solution for medical image segmentation with scarce clinical annotations, effectively easing the burden of acquiring labeled data. However, existing dual-network consistency and uncertainty-aware pseudo-labeling frameworks (e.g., CPS-style mutual teaching, mutual correction) still suffer from prediction bias accumulation and overconfident pseudo-label errors, limiting their ability to learn complex features reliably. In this paper, we propose a dual-rectification mutual learning framework with uncertainty measure (DRML-UM) for these challenges. Specifically, a conflict-aware mutual rectification module (CAMR) between multi-view networks is designed to fully exploit subnets’ capacity for diverse feature extraction and avoid potential prediction conflicts caused by accumulated learning biases. Additionally, a belief-entropy cross rectification (BECR) strategy based on uncertainty modeling and dynamic entropy comparison is proposed to mitigate overfitting in simple regions and generate more reliable pseudo-labels in complex regions. We evaluate DRML-UM on LA-MRI and NIH Pancreas-CT with 20% annotations, achieving consistent improvements (e.g., Dice: 89.93±1.23% vs. 88.96±1.45%, <i>p</i>&lt;0.05) over state-of-the-art SSL methods across four segmentation metrics.</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

A dual-rectification mutual learning framework with uncertainty measure for semi-supervised 3D medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年11月11日

查看全文
  • Yongkang Fu,
  • Xue Wang,
  • Taihui Liu,
  • Lingling Zhang,
  • Yongfa Zhu

摘要

Semi-supervised learning (SSL) is a promising high-performance computing solution for medical image segmentation with scarce clinical annotations, effectively easing the burden of acquiring labeled data. However, existing dual-network consistency and uncertainty-aware pseudo-labeling frameworks (e.g., CPS-style mutual teaching, mutual correction) still suffer from prediction bias accumulation and overconfident pseudo-label errors, limiting their ability to learn complex features reliably. In this paper, we propose a dual-rectification mutual learning framework with uncertainty measure (DRML-UM) for these challenges. Specifically, a conflict-aware mutual rectification module (CAMR) between multi-view networks is designed to fully exploit subnets’ capacity for diverse feature extraction and avoid potential prediction conflicts caused by accumulated learning biases. Additionally, a belief-entropy cross rectification (BECR) strategy based on uncertainty modeling and dynamic entropy comparison is proposed to mitigate overfitting in simple regions and generate more reliable pseudo-labels in complex regions. We evaluate DRML-UM on LA-MRI and NIH Pancreas-CT with 20% annotations, achieving consistent improvements (e.g., Dice: 89.93±1.23% vs. 88.96±1.45%, p<0.05) over state-of-the-art SSL methods across four segmentation metrics.