Scholarlink

<p>Accurate automatic segmentation of gastric cancer in ultrasound images is crucial for early diagnosis and treatment. However, this task remains challenging due to severe speckle noise, tissue deformation, and the high computational cost of fully supervised models in real-time clinical workflows. While few-shot reference-guided methods offer an efficient alternative, they often struggle with cross-frame feature alignment under low-quality ultrasound signals. To overcome these limitations, we propose FS-DANet, a Frequency-Spatial Dual-domain Awareness and Dynamic Adaptation Network. FS-DANet first adopts a Dual-Frequency Disentangled Block (DFDB) to explicitly separate stochastic noise from anatomical structures in the frequency domain. A Mask-Guided Feature Calibration (MGFC) module is then introduced to suppress background artifacts and establish precise cross-frame correspondences using the support mask as an anchor. Furthermore, an Iterative Deformable Adaptation (IDA) decoder is employed to progressively reconstruct irregular lesion boundaries at the pixel level. Extensive experiments on a proprietary Gastric Cancer Ultrasound (GCUI) dataset and the public BUSI dataset demonstrate that FS-DANet consistently achieves superior performance over state-of-the-art reference-guided methods. In particular, it obtains Dice coefficients of 75.41% and 74.19% on the GCUI and BUSI datasets, respectively, even surpassing fully supervised baselines under a single-reference setting. These findings suggest that FS-DANet can substantially improve segmentation accuracy in complex ultrasound environments and holds strong potential for clinical application in personalized diagnosis.</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

FS-DANet: Dual-Domain Signal Enhancement and Dynamic Spatial Calibration for Gastric Ultrasound Artifact Mitigation

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Yuyi Bai,
  • Yanmin Luo,
  • Zhikui Chen,
  • Youjia Lin,
  • Mingling Zhuo,
  • Junli Ma,
  • Xin Lin,
  • Qiqi Xie,
  • Zihang Pan,
  • Yuxin Fang

摘要

Accurate automatic segmentation of gastric cancer in ultrasound images is crucial for early diagnosis and treatment. However, this task remains challenging due to severe speckle noise, tissue deformation, and the high computational cost of fully supervised models in real-time clinical workflows. While few-shot reference-guided methods offer an efficient alternative, they often struggle with cross-frame feature alignment under low-quality ultrasound signals. To overcome these limitations, we propose FS-DANet, a Frequency-Spatial Dual-domain Awareness and Dynamic Adaptation Network. FS-DANet first adopts a Dual-Frequency Disentangled Block (DFDB) to explicitly separate stochastic noise from anatomical structures in the frequency domain. A Mask-Guided Feature Calibration (MGFC) module is then introduced to suppress background artifacts and establish precise cross-frame correspondences using the support mask as an anchor. Furthermore, an Iterative Deformable Adaptation (IDA) decoder is employed to progressively reconstruct irregular lesion boundaries at the pixel level. Extensive experiments on a proprietary Gastric Cancer Ultrasound (GCUI) dataset and the public BUSI dataset demonstrate that FS-DANet consistently achieves superior performance over state-of-the-art reference-guided methods. In particular, it obtains Dice coefficients of 75.41% and 74.19% on the GCUI and BUSI datasets, respectively, even surpassing fully supervised baselines under a single-reference setting. These findings suggest that FS-DANet can substantially improve segmentation accuracy in complex ultrasound environments and holds strong potential for clinical application in personalized diagnosis.