Scholarlink

While Low-dose computed tomography (LDCT) reduces the risks associated with radiation exposure, it also introduces noise and artifacts. However, balancing global dependency capture and computational efficiency is difficult because of the quadratic complexity of self-attention concerning sequence length, particularly in high-resolution dense prediction tasks. In this paper, we propose a network based on the U-Net architecture enhanced with variable shifted window self-attention mechanisms, achieving a flexible balance between local feature extraction and global context modeling for effective LDCT denoising. Our approach enables self-attention over larger window ranges, thereby achieving superior denoising quality, without increasing computational complexity. We mitigate the limitation of window size in our module based on Swin Transformer, allowing our module to accommodate feature maps of varying resolutions with greater flexibility maintain a balance between local feature extraction capability and global context awareness. Our proposed model achieves significantly better results than other state-of-the-art methods on the Mayo-2016 dataset.

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

Unet-Like Pure Transformer Model with Variable Shifted Windows for Low-Dose CT Denoising

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yijie Pu,
  • Juan Chen,
  • Huaqiang Yuan,
  • Janfang Li,
  • Yakang Li

摘要

While Low-dose computed tomography (LDCT) reduces the risks associated with radiation exposure, it also introduces noise and artifacts. However, balancing global dependency capture and computational efficiency is difficult because of the quadratic complexity of self-attention concerning sequence length, particularly in high-resolution dense prediction tasks. In this paper, we propose a network based on the U-Net architecture enhanced with variable shifted window self-attention mechanisms, achieving a flexible balance between local feature extraction and global context modeling for effective LDCT denoising. Our approach enables self-attention over larger window ranges, thereby achieving superior denoising quality, without increasing computational complexity. We mitigate the limitation of window size in our module based on Swin Transformer, allowing our module to accommodate feature maps of varying resolutions with greater flexibility maintain a balance between local feature extraction capability and global context awareness. Our proposed model achieves significantly better results than other state-of-the-art methods on the Mayo-2016 dataset.