Scholarlink

<p>Gliomas are among the most common and highly invasive tumors of the nervous system, and their accurate segmentation is essential for reliable clinical assessment and treatment planning. However, existing segmentation approaches face two major challenges: the difficulty in efficiently modeling long-range spatial dependencies and the insufficient integration of complementary information across multimodal MRI sequences, which often results in inaccurate delineation of structurally complex and heterogeneous tumor regions. To overcome these limitations, we propose a novel Cross-Modal Segmentation Network (CMNet). The proposed network incorporates three key components: a CNN-Vision Mamba Parallel Block (CVMPB), a Weak Region Enhanced Fusion Module (WREFM), and a Cross-Scale Feature Enhancement Module (CSFEM). The CVMPB simultaneously extracts local features and global contextual representations with linear computational complexity, effectively mitigating the computational burden inherent in Transformer-based architectures; the WREFM dynamically enhances the features of weaker modalities during skip connections, ensuring comprehensive utilization of cross-modal information; and the CSFEM integrates multi-scale and boundary information within the decoder to refine the segmentation boundaries and improve structural precision. The proposed CMNet architecture was rigorously validated through comprehensive evaluations on the BraTS 2020 and BraTS 2021 benchmark datasets. In the whole tumor area (WT), tumor core area (TC), and enhanced tumor area (ET), the Dice coefficients reached 90.24%, 91.35%, 84.37%, and 92.27%, 92.92%, 88.22%, respectively. Compared to recently developed analogous networks, its performance demonstrates a certain degree of competitiveness. Our code is publicly available at <a href="https://github.com/cwy1024/CMNet">https://github.com/cwy1024/CMNet</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

Fully utilizing cross modal features to achieve precise segmentation of brain gliomas

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Weiye Cao,
  • Kaiyan Zhu,
  • Jianhao Xu,
  • Tong Liu,
  • Yue Liu,
  • Weibo Song

摘要

Gliomas are among the most common and highly invasive tumors of the nervous system, and their accurate segmentation is essential for reliable clinical assessment and treatment planning. However, existing segmentation approaches face two major challenges: the difficulty in efficiently modeling long-range spatial dependencies and the insufficient integration of complementary information across multimodal MRI sequences, which often results in inaccurate delineation of structurally complex and heterogeneous tumor regions. To overcome these limitations, we propose a novel Cross-Modal Segmentation Network (CMNet). The proposed network incorporates three key components: a CNN-Vision Mamba Parallel Block (CVMPB), a Weak Region Enhanced Fusion Module (WREFM), and a Cross-Scale Feature Enhancement Module (CSFEM). The CVMPB simultaneously extracts local features and global contextual representations with linear computational complexity, effectively mitigating the computational burden inherent in Transformer-based architectures; the WREFM dynamically enhances the features of weaker modalities during skip connections, ensuring comprehensive utilization of cross-modal information; and the CSFEM integrates multi-scale and boundary information within the decoder to refine the segmentation boundaries and improve structural precision. The proposed CMNet architecture was rigorously validated through comprehensive evaluations on the BraTS 2020 and BraTS 2021 benchmark datasets. In the whole tumor area (WT), tumor core area (TC), and enhanced tumor area (ET), the Dice coefficients reached 90.24%, 91.35%, 84.37%, and 92.27%, 92.92%, 88.22%, respectively. Compared to recently developed analogous networks, its performance demonstrates a certain degree of competitiveness. Our code is publicly available at https://github.com/cwy1024/CMNet