Scholarlink

<p>Medical image segmentation is pivotal in computer-aided diagnosis systems, yet its effectiveness is often constrained by the scarcity of annotated data. Semi-supervised learning offers a promising solution by leveraging both labeled and unlabeled data, though existing methods often apply uniform data augmentation strategies that may hinder the model’s ability to learn progressively meaningful representations. To address these challenges, we introduce a novel progressive region exchange method (PREM) combined with a multi-dimensional collaborative pseudo-labeling strategy. PREM enhances data diversity by generating hybrid samples through bidirectional copy-paste operations and scale transformations, creating image-label pairs with varied semantic complexity. This facilitates progressive learning by exposing the model to a wider range of structural variations. Simultaneously, the multi-dimensional pseudo-labeling mechanism provides richer supervision signals to guide the learning process. Extensive experiments on three public benchmarks–ACDC, PROMISE12, and LA–demonstrate that our method achieves state-of-the-art performance, with up to 1.84% Dice score improvement over existing methods using only 5% labeled data on the ACDC dataset. Our code is publicly available at: <a href="https://github.com/FANR888/PREM.git">https://github.com/FANR888/PREM.git</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

Progressive region exchange: enhancing semi-supervised medical image segmentation through incremental complexity

  • 期刊论文

  • 发表日期：

    2025年12月24日

查看全文
  • Rongze Fan,
  • Dinghan Chen,
  • Cheng Wang,
  • Tao Wang,
  • Jin Song,
  • Yuanbin Chen,
  • Tong Tong,
  • Xinlin Zhang

摘要

Medical image segmentation is pivotal in computer-aided diagnosis systems, yet its effectiveness is often constrained by the scarcity of annotated data. Semi-supervised learning offers a promising solution by leveraging both labeled and unlabeled data, though existing methods often apply uniform data augmentation strategies that may hinder the model’s ability to learn progressively meaningful representations. To address these challenges, we introduce a novel progressive region exchange method (PREM) combined with a multi-dimensional collaborative pseudo-labeling strategy. PREM enhances data diversity by generating hybrid samples through bidirectional copy-paste operations and scale transformations, creating image-label pairs with varied semantic complexity. This facilitates progressive learning by exposing the model to a wider range of structural variations. Simultaneously, the multi-dimensional pseudo-labeling mechanism provides richer supervision signals to guide the learning process. Extensive experiments on three public benchmarks–ACDC, PROMISE12, and LA–demonstrate that our method achieves state-of-the-art performance, with up to 1.84% Dice score improvement over existing methods using only 5% labeled data on the ACDC dataset. Our code is publicly available at: https://github.com/FANR888/PREM.git.