Scholarlink

<p>Denoising diffusion models set the state of the art for image super-resolution but remain limited by the memory and compute cost of full-resolution sampling and by their reliance on large paired low- and high-resolution datasets that are scarce in clinical practice. This study introduces a semi-supervised generative framework, a scale-coherent mean-teacher consistency for hierarchical patch diffusion. The denoiser operates on small, sparsely sampled patches across a discrete scale hierarchy and is conditioned on a deep context fusion vector that propagates coarse-to-fine structure, decoupling per-step cost from the target resolution. The teacher is an exponential moving average of the student, and the consistency objective is evaluated in the diffusion noise space at every timestep and level using a shared patch operator with nested offsets, so labeled and unlabeled gradients act on co-located receptive fields. A ramped consistency weight and a stop-gradient teacher pass shield the context-fusion pathway from early pseudo-label noise, yielding a per-level variational bound on a mixed conditional distribution. With only 10% labels, the framework attains 29.27/0.8913 (PSNR/SSIM) on CT Lung and 30.89/0.9002 on COVID-19 X-Ray, surpassing the diffusion baseline ResShift (27.53/0.8491 and 28.68/0.8712), while reducing per-image inference to 8.7&#xa0;s and peak memory to 6.3&#xa0;GB versus 184.3&#xa0;s and 18.7&#xa0;GB for SR3. A fully supervised ablation on Set5, B100, and Urban100 (e.g., 28.09/0.8542 on Urban100 vs. 27.45/0.8371 for ESRGAN) further confirms that the hierarchical patch architecture is competitive in its own right. <a href="https://github.com/devesh-deepak/Scale-Coherent_Mean-Teacher_Hierarchical_Patch_Diffusion">https://github.com/devesh-deepak/Scale-Coherent_Mean-Teacher_Hierarchical_Patch_Diffusion</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

Scale-coherent mean-teacher hierarchical patch diffusion for label-efficient medical image super-resolution

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Devesh Kumar,
  • Geeta Sikka,
  • Samayveer Singh

摘要

Denoising diffusion models set the state of the art for image super-resolution but remain limited by the memory and compute cost of full-resolution sampling and by their reliance on large paired low- and high-resolution datasets that are scarce in clinical practice. This study introduces a semi-supervised generative framework, a scale-coherent mean-teacher consistency for hierarchical patch diffusion. The denoiser operates on small, sparsely sampled patches across a discrete scale hierarchy and is conditioned on a deep context fusion vector that propagates coarse-to-fine structure, decoupling per-step cost from the target resolution. The teacher is an exponential moving average of the student, and the consistency objective is evaluated in the diffusion noise space at every timestep and level using a shared patch operator with nested offsets, so labeled and unlabeled gradients act on co-located receptive fields. A ramped consistency weight and a stop-gradient teacher pass shield the context-fusion pathway from early pseudo-label noise, yielding a per-level variational bound on a mixed conditional distribution. With only 10% labels, the framework attains 29.27/0.8913 (PSNR/SSIM) on CT Lung and 30.89/0.9002 on COVID-19 X-Ray, surpassing the diffusion baseline ResShift (27.53/0.8491 and 28.68/0.8712), while reducing per-image inference to 8.7 s and peak memory to 6.3 GB versus 184.3 s and 18.7 GB for SR3. A fully supervised ablation on Set5, B100, and Urban100 (e.g., 28.09/0.8542 on Urban100 vs. 27.45/0.8371 for ESRGAN) further confirms that the hierarchical patch architecture is competitive in its own right. https://github.com/devesh-deepak/Scale-Coherent_Mean-Teacher_Hierarchical_Patch_Diffusion