Scholarlink

<p>Mainstream methods in medical image segmentation, such as those based on convolutional neural network and Vision Transformer, have shown impressive performance but still exhibit notable limitations: The relationship between feature channels and spatial information has not been fully explored, and the inherent bias of convolutional operations limits the extraction of contextual information. Additionally, semantic discrepancies between the encoder and decoder are influenced by skip connections, which affect segmentation performance. To address these issues, we propose a novel deep dynamic self-adjusting UNet (DDS-UNet). The core innovations of the network include: (1) efficient multi-scale self-attention mechanism, which can simultaneously capture multi-scale and global feature information; (2) a novel semantic mitigation module, which effectively reduces the semantic gap between codec layers; (3) lightweight deformable residual module replaces the basic module of traditional UNet to enhance receptive field and improve contour feature extraction capability. Experiments on four publicly available medical image datasets show that DDS-UNet significantly outperforms existing methods, especially in segmentation tasks with complex structures. This study provides an efficient and universal solution for medical image segmentation, which has wide clinical application potential. Our code is available at <a href="https://github.com/ououyy/DDSUNet">https://github.com/ououyy/DDSUNet</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

Enhanced medical image segmentation via deep dynamic self-adjusting U-Net with multi-scale attention and semantic mitigation

  • 期刊论文

  • 发表日期：

    2025年4月17日

查看全文
  • Yanchi Ou,
  • Yufeng Chen,
  • Yuzhi Zhang,
  • Shukai Yang,
  • Xiaoqian Zhang,
  • Ying Zhou,
  • Siyu Chen,
  • Lifan Peng

摘要

Mainstream methods in medical image segmentation, such as those based on convolutional neural network and Vision Transformer, have shown impressive performance but still exhibit notable limitations: The relationship between feature channels and spatial information has not been fully explored, and the inherent bias of convolutional operations limits the extraction of contextual information. Additionally, semantic discrepancies between the encoder and decoder are influenced by skip connections, which affect segmentation performance. To address these issues, we propose a novel deep dynamic self-adjusting UNet (DDS-UNet). The core innovations of the network include: (1) efficient multi-scale self-attention mechanism, which can simultaneously capture multi-scale and global feature information; (2) a novel semantic mitigation module, which effectively reduces the semantic gap between codec layers; (3) lightweight deformable residual module replaces the basic module of traditional UNet to enhance receptive field and improve contour feature extraction capability. Experiments on four publicly available medical image datasets show that DDS-UNet significantly outperforms existing methods, especially in segmentation tasks with complex structures. This study provides an efficient and universal solution for medical image segmentation, which has wide clinical application potential. Our code is available at https://github.com/ououyy/DDSUNet.