Scholarlink

<p>This study introduces a novel diffusion probabilistic model, called Residual Attention Medical Diffusion Model (ResAtMedDiff), designed to enhance the segmentation of medical images. Our primary contribution is a novel multi-scale conditional fusion mechanism. This is achieved through a dual-network architecture comprising a dynamic encoder, Multi-scale Enhanced&#xa0;U-Net Encoder (MEUNet), which extracts hierarchical anatomical features from the source image, and a robust codec, Residual-attention U-Net (RsAtUNet). These multi-scale features are then integrated into the diffusion model’s denoising process to provide fine-grained guidance, significantly improving feature representation and segmentation precision over existing methods. ResAtMedDiff demonstrates significant improvements in segmentation accuracy, achieving a 6.5% increase in the Dice similarity coefficient and an 11.9% improvement in the Intersection over Union metric compared to state-of-the-art methods. The application of this model is validated for invasive breast cancer segmentation in magnetic resonance imaging (MRI). The results highlight the potential of this model to enhance diagnostic accuracy, offering practical advancements for clinical applications. The code is available for open source at <a href="https://github.com/xiaopeng113/ResAtMedDiff">https://github.com/xiaopeng113/ResAtMedDiff</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

A diffusion probabilistic model with multi-scale conditional fusion for enhanced medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年11月9日

查看全文
  • Peng Yang,
  • Chaorong Li,
  • Xudong Ling,
  • Fengqing Qin,
  • Yong Zheng,
  • Lihua Qiu,
  • Libin Cui

摘要

This study introduces a novel diffusion probabilistic model, called Residual Attention Medical Diffusion Model (ResAtMedDiff), designed to enhance the segmentation of medical images. Our primary contribution is a novel multi-scale conditional fusion mechanism. This is achieved through a dual-network architecture comprising a dynamic encoder, Multi-scale Enhanced U-Net Encoder (MEUNet), which extracts hierarchical anatomical features from the source image, and a robust codec, Residual-attention U-Net (RsAtUNet). These multi-scale features are then integrated into the diffusion model’s denoising process to provide fine-grained guidance, significantly improving feature representation and segmentation precision over existing methods. ResAtMedDiff demonstrates significant improvements in segmentation accuracy, achieving a 6.5% increase in the Dice similarity coefficient and an 11.9% improvement in the Intersection over Union metric compared to state-of-the-art methods. The application of this model is validated for invasive breast cancer segmentation in magnetic resonance imaging (MRI). The results highlight the potential of this model to enhance diagnostic accuracy, offering practical advancements for clinical applications. The code is available for open source at https://github.com/xiaopeng113/ResAtMedDiff.