Scholarlink

<p>The CT image segmentation task of non-small cell lung cancer (NSCLC) suffers from the challenges of small target and severe background noise interference, etc. Current research incorporates a variety of neural network architectures to improve the overall segmentation performance, but most of the frameworks still have difficulty in balancing local detail preservation and global background modeling, especially the lack of ability to perceive small nodules. Therefore, this paper proposes dynamic multi-scale vision Mamba UNet (DMS-Vmunet). HDVSS modules for enhancing the segmentation performance on small targets are proposed in this framework, including Multi-SS2D module and IDConv module. The Multi-SS2D module enhances the local sensing bias and perception of small targets through multi-scale parallel branching design, while the IDConv module focuses the convolution operation on the important regions of small targets through adaptive dynamic convolution. In addition, in order to ensure the effective fusion of local and global features in parallel branches, a dual-path feature fusion (DFC) module is introduced to fuse the multi-scale feature information from multiple branches. Experimental results show that the proposed framework can effectively improve the performance of CT image segmentation for non-small cell lung cancer, especially for small target segmentation.</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

DMS-Vmunet: dynamic multi-scale vision Mamba Unet for non-small cell lung cancer CT image segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Tao Wang,
  • Hui Zeng,
  • ZeFeng Zhang,
  • GuoLiang Zhang,
  • Rui Wang,
  • Feng Gao

摘要

The CT image segmentation task of non-small cell lung cancer (NSCLC) suffers from the challenges of small target and severe background noise interference, etc. Current research incorporates a variety of neural network architectures to improve the overall segmentation performance, but most of the frameworks still have difficulty in balancing local detail preservation and global background modeling, especially the lack of ability to perceive small nodules. Therefore, this paper proposes dynamic multi-scale vision Mamba UNet (DMS-Vmunet). HDVSS modules for enhancing the segmentation performance on small targets are proposed in this framework, including Multi-SS2D module and IDConv module. The Multi-SS2D module enhances the local sensing bias and perception of small targets through multi-scale parallel branching design, while the IDConv module focuses the convolution operation on the important regions of small targets through adaptive dynamic convolution. In addition, in order to ensure the effective fusion of local and global features in parallel branches, a dual-path feature fusion (DFC) module is introduced to fuse the multi-scale feature information from multiple branches. Experimental results show that the proposed framework can effectively improve the performance of CT image segmentation for non-small cell lung cancer, especially for small target segmentation.