Scholarlink

<p>Medical imaging is essential for clinical diagnosis and biomedical research. However, data scarcity arising from privacy and security constraints limits downstream imaging tasks. Existing synthesis methods often suffer from blurred structures or unrealistic textures, primarily due to their limited ability to preserve anatomical boundaries and integrate multi-scale features effectively. To address these challenges, we propose BGS-Diff, a diffusion-based network that improves global edge similarity through two novel components: a Boundary Guidance Module (BGM) and a Scale-aware Attention Fusion Module (SAFM). The BGM integrates differentiable Canny edge extraction as an end-to-end feature enhancement mechanism, combining multi-scale convolution, channel attention, and dense connections to adaptively calibrate anatomical boundaries prior to latent diffusion. In contrast to conventional fixed-weight fusion strategies, the SAFM employs content-adaptive dynamic weight allocation across multi-window self-attention branches to capture heterogeneous multi-scale medical features. We evaluate our method on three public medical image datasets: PH2 (skin lesions), BUSI (breast ultrasound), and COVID-19 (chest X-ray). Compared to the baseline, our approach achieves state-of-the-art performance, reducing FID by 105.8, 133.6, and 97.9, and LPIPS by 0.09, 0.08, and 0.05 on the three datasets, respectively. These results demonstrate its robustness and potential as a data augmentation tool for downstream medical image analysis under data-limited conditions.</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

BGS-Diff: Boundary-guided and scale-aware diffusion model for medical image synthesis

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Yanbin Zhu,
  • Liang Zheng,
  • Haowen Lu,
  • Hanling Zhang

摘要

Medical imaging is essential for clinical diagnosis and biomedical research. However, data scarcity arising from privacy and security constraints limits downstream imaging tasks. Existing synthesis methods often suffer from blurred structures or unrealistic textures, primarily due to their limited ability to preserve anatomical boundaries and integrate multi-scale features effectively. To address these challenges, we propose BGS-Diff, a diffusion-based network that improves global edge similarity through two novel components: a Boundary Guidance Module (BGM) and a Scale-aware Attention Fusion Module (SAFM). The BGM integrates differentiable Canny edge extraction as an end-to-end feature enhancement mechanism, combining multi-scale convolution, channel attention, and dense connections to adaptively calibrate anatomical boundaries prior to latent diffusion. In contrast to conventional fixed-weight fusion strategies, the SAFM employs content-adaptive dynamic weight allocation across multi-window self-attention branches to capture heterogeneous multi-scale medical features. We evaluate our method on three public medical image datasets: PH2 (skin lesions), BUSI (breast ultrasound), and COVID-19 (chest X-ray). Compared to the baseline, our approach achieves state-of-the-art performance, reducing FID by 105.8, 133.6, and 97.9, and LPIPS by 0.09, 0.08, and 0.05 on the three datasets, respectively. These results demonstrate its robustness and potential as a data augmentation tool for downstream medical image analysis under data-limited conditions.