<p>Precise segmentation of breast ultrasound (BUS) images is crucial for early cancer diagnosis but is often hindered by inherent speckle noise, low contrast, and morphological complexity. Existing methods often fail to balance noise suppression with detail preservation, resulting in fragmented or discontinuous segmentations. In this paper, we propose a Progressive Frequency-Spatial Feature Refinement Network (PFRU-Net) for precise breast lesion segmentation, to assist clinicians in identify boundaries and reducing misdiagnosis. Specifically, a Hybrid Down-sampling Strategy(HDS) is employed in the encoder to retain key frequency-domain and structural information. Subsequently, a Selective Dual-Convolution Block(SDC) is integrated to perform hierarchical feature screening and noise suppression, complemented by a Dual Context Enhanced Attention mechanism(DCEA) that fuses global semantics with local detail responses. Furthermore, a Gaussian-Guided Feature Refinement Module(GGFR) is introduced in the decoder to spatially smooth and rectify features, ensuring topological continuity. Extensive experiments on three public datasets compared with 15 mainstream segmentation networks demonstrate the superior accuracy of the PFRU-Net model. Additional robustness analyses and external validation confirm the model’s exceptional generalization capability in complex clinical scenarios. The code is available at <a href="https://github.com/ZhaoYXin/PFRU-Net">https://github.com/ZhaoYXin/PFRU-Net</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

PFRU-net: a progressive frequency-spatial feature refinement U-net for breast lesions segmentation in ultrasound images​

  • Qun Gu,
  • Yanxin Zhao

摘要

Precise segmentation of breast ultrasound (BUS) images is crucial for early cancer diagnosis but is often hindered by inherent speckle noise, low contrast, and morphological complexity. Existing methods often fail to balance noise suppression with detail preservation, resulting in fragmented or discontinuous segmentations. In this paper, we propose a Progressive Frequency-Spatial Feature Refinement Network (PFRU-Net) for precise breast lesion segmentation, to assist clinicians in identify boundaries and reducing misdiagnosis. Specifically, a Hybrid Down-sampling Strategy(HDS) is employed in the encoder to retain key frequency-domain and structural information. Subsequently, a Selective Dual-Convolution Block(SDC) is integrated to perform hierarchical feature screening and noise suppression, complemented by a Dual Context Enhanced Attention mechanism(DCEA) that fuses global semantics with local detail responses. Furthermore, a Gaussian-Guided Feature Refinement Module(GGFR) is introduced in the decoder to spatially smooth and rectify features, ensuring topological continuity. Extensive experiments on three public datasets compared with 15 mainstream segmentation networks demonstrate the superior accuracy of the PFRU-Net model. Additional robustness analyses and external validation confirm the model’s exceptional generalization capability in complex clinical scenarios. The code is available at https://github.com/ZhaoYXin/PFRU-Net.​