Scholarlink

<p>The cross-teaching strategy, as an effective semi-supervised learning approach, has been widely applied to medical image segmentation tasks. However, the training curves of different subnetworks exhibit variations, and their predictions are influenced by cognitive uncertainty, which poses challenges in transmitting reliable instructional information between subnetworks. We propose a novel framework with Dual Uncertainty-aware and Cross-Teaching(DU-CT). This framework is composed of a CNN subnetwork and a Transformer subnetwork, with a pyramid prediction structure integrated into the decoder of each subnetwork. Building on this foundation, We introduce a dual uncertainty-aware strategy to assess the reliability of pseudo-labels from both unsupervised and supervised perspectives. For unlabeled data, we weight the pseudo supervision losses across different branches by calculating the sample-level uncertainty of the pseudo-labels, while also balancing the contributions of positive and negative samples. For labeled data, we employ a nonlinear approach to reduce the impact of pseudo-labels generated by underperforming networks, thereby alleviating the interference from training curve disparities across learning paradigms in model optimization. Finally, we impose consistency constraints on the multi-scale predictions of both subnetworks to enhance training stability. Our method, evaluated on lung nodule CT and brain glioma MRI datasets, demonstrates superior performance over state-of-the-art techniques. The code will be available at <a href="https://github.com/sunshine-cd-cell/du-ct">https://github.com/sunshine-cd-cell/du-ct</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-teaching with dual uncertainty awareness for semi-supervised medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年9月6日

查看全文
  • Qinglong Xu,
  • Haixing Zhu,
  • Yuan Wang,
  • Zhongjie Shi,
  • Weipeng Liu

摘要

The cross-teaching strategy, as an effective semi-supervised learning approach, has been widely applied to medical image segmentation tasks. However, the training curves of different subnetworks exhibit variations, and their predictions are influenced by cognitive uncertainty, which poses challenges in transmitting reliable instructional information between subnetworks. We propose a novel framework with Dual Uncertainty-aware and Cross-Teaching(DU-CT). This framework is composed of a CNN subnetwork and a Transformer subnetwork, with a pyramid prediction structure integrated into the decoder of each subnetwork. Building on this foundation, We introduce a dual uncertainty-aware strategy to assess the reliability of pseudo-labels from both unsupervised and supervised perspectives. For unlabeled data, we weight the pseudo supervision losses across different branches by calculating the sample-level uncertainty of the pseudo-labels, while also balancing the contributions of positive and negative samples. For labeled data, we employ a nonlinear approach to reduce the impact of pseudo-labels generated by underperforming networks, thereby alleviating the interference from training curve disparities across learning paradigms in model optimization. Finally, we impose consistency constraints on the multi-scale predictions of both subnetworks to enhance training stability. Our method, evaluated on lung nodule CT and brain glioma MRI datasets, demonstrates superior performance over state-of-the-art techniques. The code will be available at https://github.com/sunshine-cd-cell/du-ct.