Scholarlink

<p>Medical image segmentation serves as a fundamental prerequisite for computer-aided diagnosis, treatment planning and disease progression monitoring. Traditional deep learning methods based on multi-layer perceptrons (MLPs) are limited by fixed activation functions and linear weight structures, making it difficult to accurately capture the nonlinear geometric features of lesions. Recently, U-KAN has been proposed to replace MLPs’ rigid transformations with learnable nonlinear spline basis functions. However, U-KAN suffers from inefficient downsampling and lacks attention mechanisms, limiting its feature extraction capability. In this paper, we proposed the ResUKAN method, which incorporates the pretrained ResNet18 into the U-KAN architecture to optimize the downsampling capability of U-KAN. Furthermore, we design a novel KAN-enhanced channel attention module (KECAM), which can model the high-order nonlinear dependence between channels through the data adaptive interpolation of spline basis function, and significantly improve the accuracy of feature selection. We validated our method on colorectal polyp images and breast cancer ultrasound images. Experimental results show that ResUKAN consistently outperforms existing methods across all evaluation metrics. Ablation studies confirm that KANs more effectively suppress background noise and improve segmentation accuracy for subtle lesions compared to traditional MLPs. This work provides a more lightweight and robust solution for medical image segmentation. The implementation code has been made publicly available at <a href="https://github.com/fengyu123-nb/ResUKAN">https://github.com/fengyu123-nb/ResUKAN</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

Resukan: an improved medical image segmentation method based on U-KAN architecture

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Yu Feng,
  • Haomeng Xie,
  • Chuanshen Zhao,
  • Haiyong Zhao

摘要

Medical image segmentation serves as a fundamental prerequisite for computer-aided diagnosis, treatment planning and disease progression monitoring. Traditional deep learning methods based on multi-layer perceptrons (MLPs) are limited by fixed activation functions and linear weight structures, making it difficult to accurately capture the nonlinear geometric features of lesions. Recently, U-KAN has been proposed to replace MLPs’ rigid transformations with learnable nonlinear spline basis functions. However, U-KAN suffers from inefficient downsampling and lacks attention mechanisms, limiting its feature extraction capability. In this paper, we proposed the ResUKAN method, which incorporates the pretrained ResNet18 into the U-KAN architecture to optimize the downsampling capability of U-KAN. Furthermore, we design a novel KAN-enhanced channel attention module (KECAM), which can model the high-order nonlinear dependence between channels through the data adaptive interpolation of spline basis function, and significantly improve the accuracy of feature selection. We validated our method on colorectal polyp images and breast cancer ultrasound images. Experimental results show that ResUKAN consistently outperforms existing methods across all evaluation metrics. Ablation studies confirm that KANs more effectively suppress background noise and improve segmentation accuracy for subtle lesions compared to traditional MLPs. This work provides a more lightweight and robust solution for medical image segmentation. The implementation code has been made publicly available at https://github.com/fengyu123-nb/ResUKAN.