Scholarlink

Four-dimensional CT (4D-CT) tracks tumor motion throughout the breathing cycle for radiation therapy planning, but dose reduction per phase introduces spatio-temporal noise compromising tumor delineation. Existing learning-based denoising methods are either clinically impractical (requiring paired data) or lack interpretability (black-box networks). We present Filter2Noise-4D (F2N-4D), a zero-shot interpretable framework employing content-adaptive bilateral filtering that exploits spatio-temporal information from neighboring slices. Self-supervised training uses interpolation of neighboring slices to construct training pairs. With only 1.8k parameters, F2N-4D achieves competitive performance while maintaining transparency.

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

Interpretable Framework for Zero-shot 4D Low-dose CT Denoising

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yipeng Sun,
  • Linda-Sophie Schneider,
  • Siyuan Mei,
  • Chengze Ye,
  • Mingxuan Gu,
  • Fabian Wagner,
  • Siming Bayer,
  • Andreas Maier

摘要

Four-dimensional CT (4D-CT) tracks tumor motion throughout the breathing cycle for radiation therapy planning, but dose reduction per phase introduces spatio-temporal noise compromising tumor delineation. Existing learning-based denoising methods are either clinically impractical (requiring paired data) or lack interpretability (black-box networks). We present Filter2Noise-4D (F2N-4D), a zero-shot interpretable framework employing content-adaptive bilateral filtering that exploits spatio-temporal information from neighboring slices. Self-supervised training uses interpolation of neighboring slices to construct training pairs. With only 1.8k parameters, F2N-4D achieves competitive performance while maintaining transparency.