Scholarlink

<p>SvANet is an ultra-lightweight segmentation architecture specifically designed for small-object medical imaging tasks in resource-constrained environments. The network first employs the Dual Path Fusion (DPF) module to integrate convolutional detail preservation with Transformer-style global reasoning, aligning local textures with long-range dependencies in a unified representation space. Second, the Boundary Aware Module (BAM), combined with coordinate-guided optimization, enhances positional awareness and boundary continuity. Third, the Efficient Context Encoder (ECE) models global dependencies and aggregates information with linear complexity <i>O</i>(<i>n</i>), significantly reducing computational cost. This design effectively mitigates the loss of fine-grained region information, improves boundary localization accuracy, and strengthens long-range context modeling while maintaining model compactness (0.31M parameters, 0.15 GFLOPs). On the ISIC skin lesion segmentation task, SvANet achieves an average improvement of 3.2% in mIoU over lightweight baselines. In cross-domain evaluations, it attains 76.5% and 75.3% mIoU on the CVC-ClinicDB and Kvasir polyp datasets, respectively, demonstrating strong generalization capability. Experimental results indicate that SvANet exhibits promising potential for clinical applications under limited-resource conditions. The implementation code is available at: <a href="https://github.com/1502GaoYi/SvANet">https://github.com/1502GaoYi/SvANet</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

Svanet: synergistic vision and adaptation network for lightweight medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Yi Gao,
  • Haicheng Qu,
  • Chang Lin,
  • Ying Wang,
  • Yao Wei

摘要

SvANet is an ultra-lightweight segmentation architecture specifically designed for small-object medical imaging tasks in resource-constrained environments. The network first employs the Dual Path Fusion (DPF) module to integrate convolutional detail preservation with Transformer-style global reasoning, aligning local textures with long-range dependencies in a unified representation space. Second, the Boundary Aware Module (BAM), combined with coordinate-guided optimization, enhances positional awareness and boundary continuity. Third, the Efficient Context Encoder (ECE) models global dependencies and aggregates information with linear complexity O(n), significantly reducing computational cost. This design effectively mitigates the loss of fine-grained region information, improves boundary localization accuracy, and strengthens long-range context modeling while maintaining model compactness (0.31M parameters, 0.15 GFLOPs). On the ISIC skin lesion segmentation task, SvANet achieves an average improvement of 3.2% in mIoU over lightweight baselines. In cross-domain evaluations, it attains 76.5% and 75.3% mIoU on the CVC-ClinicDB and Kvasir polyp datasets, respectively, demonstrating strong generalization capability. Experimental results indicate that SvANet exhibits promising potential for clinical applications under limited-resource conditions. The implementation code is available at: https://github.com/1502GaoYi/SvANet.