Scholarlink

<p>Lesion segmentation plays a critical role in computer-aided clinical diagnosis but remains challenging due to scale variations, blurred boundaries, and complex morphological structures. Meanwhile, the increasing resolution of medical images and the growing complexity of deep neural network architectures impose substantial computational demands, highlighting the need for scalable and high-performance computing (HPC)-oriented model design. To address both segmentation accuracy and computational efficiency, this paper proposes BMF-SegNet, a unified frequency-aware and boundary-driven segmentation framework. The proposed approach integrates multi-scale spectral interaction, efficient state-space sequence modeling, and boundary-guided optimization to jointly capture global contextual information and fine-grained structural details. By leveraging channel-wise frequency transformation and Mamba-based modeling, the framework enables effective global–local feature interaction while maintaining scalable computational complexity suitable for parallel processing. A boundary-aware learning strategy further improves edge delineation without introducing significant computational overhead. Extensive experiments on breast lesion, skin lesion, and COVID-19 pneumonia datasets demonstrate that BMF-SegNet achieves competitive performance compared with representative state-of-the-art methods, particularly in challenging boundary regions. The computation-friendly design facilitates GPU acceleration and structured parallel execution, making the proposed framework well aligned with supercomputing-oriented medical image analysis. The code is available at <a href="https://github.com/hyb2840/BMF-SegNet">https://github.com/hyb2840/BMF-SegNet</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

BMF-SegNet: a boundary-aware mamba network with multi-scale frequency interaction for lesion segmentation

  • 期刊论文

  • 发表日期：

    2026年3月18日

查看全文
  • Yibing Huang,
  • Guanghua He,
  • Zhong Li,
  • Keli Hu,
  • Hancan Zhu

摘要

Lesion segmentation plays a critical role in computer-aided clinical diagnosis but remains challenging due to scale variations, blurred boundaries, and complex morphological structures. Meanwhile, the increasing resolution of medical images and the growing complexity of deep neural network architectures impose substantial computational demands, highlighting the need for scalable and high-performance computing (HPC)-oriented model design. To address both segmentation accuracy and computational efficiency, this paper proposes BMF-SegNet, a unified frequency-aware and boundary-driven segmentation framework. The proposed approach integrates multi-scale spectral interaction, efficient state-space sequence modeling, and boundary-guided optimization to jointly capture global contextual information and fine-grained structural details. By leveraging channel-wise frequency transformation and Mamba-based modeling, the framework enables effective global–local feature interaction while maintaining scalable computational complexity suitable for parallel processing. A boundary-aware learning strategy further improves edge delineation without introducing significant computational overhead. Extensive experiments on breast lesion, skin lesion, and COVID-19 pneumonia datasets demonstrate that BMF-SegNet achieves competitive performance compared with representative state-of-the-art methods, particularly in challenging boundary regions. The computation-friendly design facilitates GPU acceleration and structured parallel execution, making the proposed framework well aligned with supercomputing-oriented medical image analysis. The code is available at https://github.com/hyb2840/BMF-SegNet.