Scholarlink

<p>Diffusion probabilistic model, as a generative model, has gained wide attention in image processing tasks, and its powerful image generation capability has made it flourish in the field of computer vision. In recent years, it has achieved excellent performance in the field of medical image segmentation, and nowadays denoising diffusion models supported by network architectures such as U-Net and Transformer are all successfully applied in medical image segmentation tasks. In order to explore the potential of pyramid structure in diffusion modelling, we propose a denoising diffusion model, DiffPVT, which uses the PVT as the network architecture. In order to ensure the diffusion model’s compatibility with the pyramid vision transformer (PVT), we use a dual-encoder architecture to convey the spatial information to ensure the effectiveness of the diffusion model in the denoising stage, while proposing a novel decoder structure to further enrich the reconstruction process, in which we retain the information embedding of the time steps in the second layer encoding stage while applying its extension to the decoder part. We combine the denoising task of the diffusion model with the critical need of the edge detection task in the field of medical image segmentation, and propose dual-mode information filtering module (DMIFM) and use it as a holistic structure to process the requirements of multiple types of tasks in parallel, thus enhancing the denoising process of the diffusion model and enriching the edge information of the feature images. We conduct extensive experiments on four public datasets and confirm that DiffPVT has excellent segmentation level in the field of medical image segmentation. <a href="https://github.com/cn-xvkong/DiffPVT">https://github.com/cn-xvkong/DiffPVT</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

Diffpvt:information filtering based diffusion model with PVT for medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年1月4日

查看全文
  • Chengming Wang,
  • Genji Yuan,
  • Mengjun Li,
  • Jinjiang Li

摘要

Diffusion probabilistic model, as a generative model, has gained wide attention in image processing tasks, and its powerful image generation capability has made it flourish in the field of computer vision. In recent years, it has achieved excellent performance in the field of medical image segmentation, and nowadays denoising diffusion models supported by network architectures such as U-Net and Transformer are all successfully applied in medical image segmentation tasks. In order to explore the potential of pyramid structure in diffusion modelling, we propose a denoising diffusion model, DiffPVT, which uses the PVT as the network architecture. In order to ensure the diffusion model’s compatibility with the pyramid vision transformer (PVT), we use a dual-encoder architecture to convey the spatial information to ensure the effectiveness of the diffusion model in the denoising stage, while proposing a novel decoder structure to further enrich the reconstruction process, in which we retain the information embedding of the time steps in the second layer encoding stage while applying its extension to the decoder part. We combine the denoising task of the diffusion model with the critical need of the edge detection task in the field of medical image segmentation, and propose dual-mode information filtering module (DMIFM) and use it as a holistic structure to process the requirements of multiple types of tasks in parallel, thus enhancing the denoising process of the diffusion model and enriching the edge information of the feature images. We conduct extensive experiments on four public datasets and confirm that DiffPVT has excellent segmentation level in the field of medical image segmentation. https://github.com/cn-xvkong/DiffPVT.