Scholarlink

<p>Medical imaging data (e.g., CT and MRI) often originate from heterogeneous sources, leading to appearance discrepancies and consequently limited generalization when training data are scarce. Prior unsupervised domain adaptation (UDA) and multi-source domain generalization (MSDG) methods mitigate distribution shifts but require access to multiple domains during training. Single-source domain generalization (SSDG) offers a practical alternative. However, existing SSDG methods typically depend on global random image augmentations or implicit feature decoupling, which often result in residual style leakage and unstable cross-domain performance. To address this challenge, we propose GLASSD (Global-Local Augmentation with Style-Structure Decoupling), a novel SSDG framework that explicitly couples learnable data-level augmentation with feature-level decoupling. Specifically, a B-spline-based augmentation module samples knot vectors to parameterize deformation fields, enabling more diverse geometric variations than conventional affine or elastic transforms and thereby improving robustness to domain shifts. Meanwhile, a style-structure decoupling module disentangles style information from structural features, mitigating domain-specific style biases in downstream segmentation. We evaluate GLASSD on two challenging multimodal datasets and benchmark its performance against state-of-the-art domain generalization methods. Extensive experiments demonstrate that GLASSD achieves superior cross-domain performance. Our code is publicly available at <a href="https://github.com/bri-bing/GLASSD-main">https://github.com/bri-bing/GLASSD-main</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

Improve domain generalization through curve enhancement and style structure decoupling for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Ruibing Fu,
  • Haoran Zhou,
  • Ziyuan Zhang,
  • Runze Lu,
  • Guangyao Li

摘要

Medical imaging data (e.g., CT and MRI) often originate from heterogeneous sources, leading to appearance discrepancies and consequently limited generalization when training data are scarce. Prior unsupervised domain adaptation (UDA) and multi-source domain generalization (MSDG) methods mitigate distribution shifts but require access to multiple domains during training. Single-source domain generalization (SSDG) offers a practical alternative. However, existing SSDG methods typically depend on global random image augmentations or implicit feature decoupling, which often result in residual style leakage and unstable cross-domain performance. To address this challenge, we propose GLASSD (Global-Local Augmentation with Style-Structure Decoupling), a novel SSDG framework that explicitly couples learnable data-level augmentation with feature-level decoupling. Specifically, a B-spline-based augmentation module samples knot vectors to parameterize deformation fields, enabling more diverse geometric variations than conventional affine or elastic transforms and thereby improving robustness to domain shifts. Meanwhile, a style-structure decoupling module disentangles style information from structural features, mitigating domain-specific style biases in downstream segmentation. We evaluate GLASSD on two challenging multimodal datasets and benchmark its performance against state-of-the-art domain generalization methods. Extensive experiments demonstrate that GLASSD achieves superior cross-domain performance. Our code is publicly available at https://github.com/bri-bing/GLASSD-main.