Scholarlink

<p>Medical images produced by different acquisition instruments contain variations in spatial resolution, contrast characteristics, and frequency content, creating challenges for consistent lesion measurement and quantitative analysis. For jaw cyst assessment, the lack of a unified and large-scale benchmark further restricts the development and verification of reliable measurement algorithms. In addition, existing deep learning–based segmentation approaches primarily focus on spatial-domain feature extraction and often overlook the directional structural information and spectral characteristics of lesions, limiting their ability to deliver precise boundary measurement across complex imaging conditions. To address these limitations, we propose SSF-Net, a spatial-spectral Fourier convolution fusion network with a Taylor Expansion Attention Transformer for automatic jaw cyst segmentation. Specifically, a Taylor expansion attention-based transformer is introduced to achieve multi-directional and multi-scale feature modeling through high-order feature decomposition, enabling the extraction of global contextual information across spatial and channel domains. Meanwhile, a Fourier Fusion Convolution (FFC) module is designed within the spectral branch to integrate global and local frequency representations, enhancing feature expressiveness and promoting deep interaction between spatial and spectral features. Furthermore, we construct and publicly release the first large-scale jaw cyst segmentation dataset, establishing a unified benchmark for future research. Experimental results demonstrate that SSF-Net achieves Dice Similarity Coefficients (DSC) of 86.86\% and 86.35\%, outperforming existing state-of-the-art methods across multiple metrics. The results confirm that SSF-Net improves the accuracy and reliability of automated lesion measurement, offering a practical and reproducible approach for instrument-generated medical image analysis. The code will be available at: <a href="http://github.com/fangchj2002/SSF-Net">http://github.com/fangchj2002/SSF-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

Spatial-spectral Fourier convolution fusion network with Taylor expansion attention-based transformer for jaw cyst segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月18日

查看全文
  • Jiangxiong Fang,
  • Shikuan Qi,
  • Guoqiang Zhong,
  • Youyao Fu,
  • Shiqing Zhang,
  • Jie Jin,
  • Huaxiang Liu,
  • Linlin Chen,
  • Guoyu Wang

摘要

Medical images produced by different acquisition instruments contain variations in spatial resolution, contrast characteristics, and frequency content, creating challenges for consistent lesion measurement and quantitative analysis. For jaw cyst assessment, the lack of a unified and large-scale benchmark further restricts the development and verification of reliable measurement algorithms. In addition, existing deep learning–based segmentation approaches primarily focus on spatial-domain feature extraction and often overlook the directional structural information and spectral characteristics of lesions, limiting their ability to deliver precise boundary measurement across complex imaging conditions. To address these limitations, we propose SSF-Net, a spatial-spectral Fourier convolution fusion network with a Taylor Expansion Attention Transformer for automatic jaw cyst segmentation. Specifically, a Taylor expansion attention-based transformer is introduced to achieve multi-directional and multi-scale feature modeling through high-order feature decomposition, enabling the extraction of global contextual information across spatial and channel domains. Meanwhile, a Fourier Fusion Convolution (FFC) module is designed within the spectral branch to integrate global and local frequency representations, enhancing feature expressiveness and promoting deep interaction between spatial and spectral features. Furthermore, we construct and publicly release the first large-scale jaw cyst segmentation dataset, establishing a unified benchmark for future research. Experimental results demonstrate that SSF-Net achieves Dice Similarity Coefficients (DSC) of 86.86\% and 86.35\%, outperforming existing state-of-the-art methods across multiple metrics. The results confirm that SSF-Net improves the accuracy and reliability of automated lesion measurement, offering a practical and reproducible approach for instrument-generated medical image analysis. The code will be available at: http://github.com/fangchj2002/SSF-Net.