Scholarlink

<p>While U-shaped architectures perform well in medical image segmentation, they often struggle to efficiently capture both multi-scale anatomical features and global dependencies: spatial CNNs rely on fixed receptive fields, and frequency-domain processing is often computationally expensive due to repeated spatial-frequency conversions. We propose DAFS-Net, a lightweight U-shaped network that fuses spatial and frequency cues through two complementary modules. In shallow encoder layers, a Dynamic Spatial Attention (DSA) module employs adaptive multi-scale context fusion with learnable weights to dynamically tune the effective receptive field, improving responses to small lesions and indistinct boundaries. In deeper layers, a Frequency-domain Enhanced Shift (FS) module enhances frequency magnitudes while preserving phase information, enabling global contextual modeling with a single FFT/IFFT and avoiding frequent inter-domain conversions. The combination of shallow spatial adaptation and deep frequency enhancement yields accurate boundary reconstruction with low overhead. Experiments on the ISIC 2018 dermoscopy and BUSI breast ultrasound datasets, together with cross-dataset evaluations on PH2 and STU, show that DAFS-Net achieves a strong accuracy–efficiency trade-off and robust generalization across modalities and datasets.</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

DAFS-Net: Lightweight medical image segmentation via fusing spatial and frequency domains

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Jinfeng Xie,
  • Jia Hu,
  • Aiyuan Li,
  • Xiaohui Liu

摘要

While U-shaped architectures perform well in medical image segmentation, they often struggle to efficiently capture both multi-scale anatomical features and global dependencies: spatial CNNs rely on fixed receptive fields, and frequency-domain processing is often computationally expensive due to repeated spatial-frequency conversions. We propose DAFS-Net, a lightweight U-shaped network that fuses spatial and frequency cues through two complementary modules. In shallow encoder layers, a Dynamic Spatial Attention (DSA) module employs adaptive multi-scale context fusion with learnable weights to dynamically tune the effective receptive field, improving responses to small lesions and indistinct boundaries. In deeper layers, a Frequency-domain Enhanced Shift (FS) module enhances frequency magnitudes while preserving phase information, enabling global contextual modeling with a single FFT/IFFT and avoiding frequent inter-domain conversions. The combination of shallow spatial adaptation and deep frequency enhancement yields accurate boundary reconstruction with low overhead. Experiments on the ISIC 2018 dermoscopy and BUSI breast ultrasound datasets, together with cross-dataset evaluations on PH2 and STU, show that DAFS-Net achieves a strong accuracy–efficiency trade-off and robust generalization across modalities and datasets.