Scholarlink

<p>In medical image segmentation, CNNs and Vision Transformer often bring difficulties to in capturing lesion-sensitive details and nonlinear representations, often resulting in blurred boundaries and reduced accuracy. To overcome these limitations, we propose Cross-Scale Residual Attention Network (CSRANet), a symmetric encoder–decoder model incorporating a Transformer-style Cross-Scale Attention (CSA) block. The CSA block consists of a Cross-Scale Attention Gate (CSAG) to fuse multi-scale features with enhanced discrimination, and a Convolutional Kolmogorov–Arnold Networks (Conv KAN) module that combines static and dynamic convolutions with adaptive weighting for stable generalization. A pointwise convolution further refines features and integrates CSA outputs to improve segmentation. Experimental results, on BUSI, PH2, and DDTI datasets, show CSRANet achieves F1 scores of 80.17%, 81.06%, and 94.55%, with IoU scores of 72.12%, 70.96% and 89.96% respectively. These results prove that CSRANet effectively enhances lesion segmentation across modalities, offering improved accuracy, robustness, and generalization over existing approaches.</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

A Cross-Scale residual attention network with convolutional kolmogorov-arnold for medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年9月26日

查看全文
  • Shangwang Liu,
  • Yulin Cheng,
  • Yinghai Lin,
  • Xianglian Jin,
  • Hongwei Wang,
  • Yusen Wang

摘要

In medical image segmentation, CNNs and Vision Transformer often bring difficulties to in capturing lesion-sensitive details and nonlinear representations, often resulting in blurred boundaries and reduced accuracy. To overcome these limitations, we propose Cross-Scale Residual Attention Network (CSRANet), a symmetric encoder–decoder model incorporating a Transformer-style Cross-Scale Attention (CSA) block. The CSA block consists of a Cross-Scale Attention Gate (CSAG) to fuse multi-scale features with enhanced discrimination, and a Convolutional Kolmogorov–Arnold Networks (Conv KAN) module that combines static and dynamic convolutions with adaptive weighting for stable generalization. A pointwise convolution further refines features and integrates CSA outputs to improve segmentation. Experimental results, on BUSI, PH2, and DDTI datasets, show CSRANet achieves F1 scores of 80.17%, 81.06%, and 94.55%, with IoU scores of 72.12%, 70.96% and 89.96% respectively. These results prove that CSRANet effectively enhances lesion segmentation across modalities, offering improved accuracy, robustness, and generalization over existing approaches.