Scholarlink

For semi-supervised image segmentation models based on mean-teacher framework, Progressive weight coupling between sub-models remains a critical challenge. This coupling is particularly pronounced in medical image segmentation, as the lack of samples makes the model more prone to reaching performance bottleneck. To mitigate this, we propose HAP-MT, a dual-space alternating perturbation method based on Dual teacher model. Specifically, we design two independent augmentation branches employing data perturbation and feature perturbation strategies respectively. By alternately applying these augmentation branches with different teacher models, HAP-MT can effectively alleviate the weight coupling. Furthermore, we introduce a foreground copy-paste strategy between labeled and unlabeled data to reduce distribution disparity while preserving essential contextual information. To further improve model’s generalization capability, we develop a multi-level consistency regularization approach. Experimental results demonstrate that HAP-MT achieves outstanding segmentation performance. The relevant code can be found at https://github.com/sumfunpan/HAP-MT .

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

HAP-MT: Alternating Perturbation Strategies Across Data and Feature Levels in Semi-Supervised Medical Image Segmentation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jiawei Pan,
  • Yang Hua,
  • Jian Wang,
  • Xiaoning Song

摘要

For semi-supervised image segmentation models based on mean-teacher framework, Progressive weight coupling between sub-models remains a critical challenge. This coupling is particularly pronounced in medical image segmentation, as the lack of samples makes the model more prone to reaching performance bottleneck. To mitigate this, we propose HAP-MT, a dual-space alternating perturbation method based on Dual teacher model. Specifically, we design two independent augmentation branches employing data perturbation and feature perturbation strategies respectively. By alternately applying these augmentation branches with different teacher models, HAP-MT can effectively alleviate the weight coupling. Furthermore, we introduce a foreground copy-paste strategy between labeled and unlabeled data to reduce distribution disparity while preserving essential contextual information. To further improve model’s generalization capability, we develop a multi-level consistency regularization approach. Experimental results demonstrate that HAP-MT achieves outstanding segmentation performance. The relevant code can be found at https://github.com/sumfunpan/HAP-MT .