Scholarlink

<p>Medical image segmentation faces inherent challenges, including irregular lesion shapes, complex backgrounds, and noise interference. Recently, diffusion models have achieved remarkable success in various generative tasks due to their powerful ability to model complex data distributions. To address these challenges, we propose a novel diffusion model segmentation framework tailored for medical images, called DBCG-Med. Our framework introduces a Biaxial Attention Module (BAM), which combines the orientation information of lesions by decomposing and weighting horizontal and vertical components to enhance spatial awareness. Furthermore, since diffusion models usually require long training time, we integrate Do-Conv into the feature encoder responsible for extracting features from the original image. This integration enhances feature extraction capabilities while improving convergence speed, training stability, and generalization capabilities. To provide consistent and stable features in the denoising stage, we design a dual context guided module (DCGM) to alleviate the encoder-decoder semantic gap. Experiments on four public medical image datasets show that our method achieves superior segmentation performance compared with mainstream methods, highlighting its great potential for practical application in clinical scenarios. <a href="https://github.com/yfz-yfz1/DBCG.">https://github.com/yfz-yfz1/DBCG.</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

Dbcg-med: diffusion-based bidirectional calibration and context guidance for medical image segmentation

  • 期刊论文

  • 发表日期：

    2025年8月6日

查看全文
  • Yunfei Zhu,
  • Peng Duan,
  • Zhen Hua,
  • Jinjiang Li

摘要

Medical image segmentation faces inherent challenges, including irregular lesion shapes, complex backgrounds, and noise interference. Recently, diffusion models have achieved remarkable success in various generative tasks due to their powerful ability to model complex data distributions. To address these challenges, we propose a novel diffusion model segmentation framework tailored for medical images, called DBCG-Med. Our framework introduces a Biaxial Attention Module (BAM), which combines the orientation information of lesions by decomposing and weighting horizontal and vertical components to enhance spatial awareness. Furthermore, since diffusion models usually require long training time, we integrate Do-Conv into the feature encoder responsible for extracting features from the original image. This integration enhances feature extraction capabilities while improving convergence speed, training stability, and generalization capabilities. To provide consistent and stable features in the denoising stage, we design a dual context guided module (DCGM) to alleviate the encoder-decoder semantic gap. Experiments on four public medical image datasets show that our method achieves superior segmentation performance compared with mainstream methods, highlighting its great potential for practical application in clinical scenarios. https://github.com/yfz-yfz1/DBCG.