Scholarlink

<p>Semi-supervised image segmentation has shown significant potential in medical image analysis. However, current mean teacher framework-based methods often suffer from confirmation bias, which limits their effectiveness. To address this issue, we propose AF-MT, an adaptive fusion method based on mean teacher with alternating loss update. AF-MT consists of a student model and two teacher models. The parameters of the teacher models are updated via exponential moving average (EMA) of the parameters of the student model. It features two key modules: a alternating loss update module (ALU) and a cross fusion module (CFM). The ALU module optimizes the under-performing teacher model using complementary data batches and labeled data loss, maintaining model diversity. The CFM module adaptively fuses the predictions from the teacher models using an entropy-based strategy. These two modules work in concert to enhance the ability of model to leverage both labeled and unlabeled data effectively. Experiments show AF-MT achieves Dice scores of 91.89% on the 3D LA dataset with 5% labeled data, 88.13% on the BraTS2019 dataset with 10% labeled data, and 91.96% on the 2D ACDC dataset with 5% labeled data, demonstrating its effectiveness in both 2D and 3D segmentation. These results highlight the potential of AF-MT to significantly improve segmentation accuracy in clinical applications.</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

AF-MT: adaptive fusion based on mean teacher with alternating loss update for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Quanzong Xu,
  • Kaizhi Chen,
  • Zuoyong Li,
  • Haiping Xu,
  • Guisheng Ding,
  • Yusheng Yao

摘要

Semi-supervised image segmentation has shown significant potential in medical image analysis. However, current mean teacher framework-based methods often suffer from confirmation bias, which limits their effectiveness. To address this issue, we propose AF-MT, an adaptive fusion method based on mean teacher with alternating loss update. AF-MT consists of a student model and two teacher models. The parameters of the teacher models are updated via exponential moving average (EMA) of the parameters of the student model. It features two key modules: a alternating loss update module (ALU) and a cross fusion module (CFM). The ALU module optimizes the under-performing teacher model using complementary data batches and labeled data loss, maintaining model diversity. The CFM module adaptively fuses the predictions from the teacher models using an entropy-based strategy. These two modules work in concert to enhance the ability of model to leverage both labeled and unlabeled data effectively. Experiments show AF-MT achieves Dice scores of 91.89% on the 3D LA dataset with 5% labeled data, 88.13% on the BraTS2019 dataset with 10% labeled data, and 91.96% on the 2D ACDC dataset with 5% labeled data, demonstrating its effectiveness in both 2D and 3D segmentation. These results highlight the potential of AF-MT to significantly improve segmentation accuracy in clinical applications.