Scholarlink

In medical image analysis, accurate image segmentation is essential for precise disease diagnosis and treatment planning. However, medical images come from a wide range of sources and are of diverse types, making the generalization ability of segmentation models an important factor affecting diagnostic accuracy. In this study, the AttU-KAN model is proposed, which combines U-KAN with the Twin-Branch Feature Extraction (TBFE) and Hybrid Pooling Attention (HPA) modules, aiming to improve the model's generalization capabilities in medical image segmentation. To evaluate the performance of the model, the study uses a pneumothorax dataset composed of lung CT images and the GlaS public dataset for gland segmentation in pathological images. The experimental results show that on the pneumothorax dataset, the Dice of the AttU-KAN model reaches 81.90% and the IoU reaches 70.76%. On the GlaS public dataset, the Dice is 92.29% and the IoU is 85.71%, outperforming other segmentation networks in both indicators. The research shows that the AttU-KAN model can effectively handle medical images of different modalities, demonstrating good generalization ability in multiple disease diagnosis scenarios and providing reliable technical support for medical image segmentation and precise disease diagnosis.

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 U-net Architecture with Kolmogorov-Arnold Networks and Hybrid Pooling Attention for Medical Image Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xucheng Cai,
  • Zhuocheng Li,
  • Zhentao Yang,
  • Zhixian Tang,
  • Haojie Han,
  • Xuyang Zhang,
  • Xufeng Yao,
  • Ling Wei

摘要

In medical image analysis, accurate image segmentation is essential for precise disease diagnosis and treatment planning. However, medical images come from a wide range of sources and are of diverse types, making the generalization ability of segmentation models an important factor affecting diagnostic accuracy. In this study, the AttU-KAN model is proposed, which combines U-KAN with the Twin-Branch Feature Extraction (TBFE) and Hybrid Pooling Attention (HPA) modules, aiming to improve the model's generalization capabilities in medical image segmentation. To evaluate the performance of the model, the study uses a pneumothorax dataset composed of lung CT images and the GlaS public dataset for gland segmentation in pathological images. The experimental results show that on the pneumothorax dataset, the Dice of the AttU-KAN model reaches 81.90% and the IoU reaches 70.76%. On the GlaS public dataset, the Dice is 92.29% and the IoU is 85.71%, outperforming other segmentation networks in both indicators. The research shows that the AttU-KAN model can effectively handle medical images of different modalities, demonstrating good generalization ability in multiple disease diagnosis scenarios and providing reliable technical support for medical image segmentation and precise disease diagnosis.