Scholarlink

<p>Multi-contrast magnetic resonance imaging (MRI) super-resolution (SR) methods effectively improve SR quality by using high-resolution reference images with similar anatomical structures. However, some medical devices have limited computational power to deploy multi-contrast MRI SR algorithms. Existing multi-contrast MRI SR methods mainly rely on computationally intensive large-scale models, making lightweight deployment of medical devices challenging. We propose a lightweight SR dynamic contrast-aware differential fusion network (LCDFNet) based on multi-contrast MRI. LCDFNet mainly consists of contrast-aware differential fusion modules (DCDFM). DCDFM includes two components: a lightweight dynamic contrast-aware channel attention block (DCAB) for MR images and a concise and efficient differential fusion block (DFB). The DCAB enhances edge feature extraction by dynamically adapting to different feature distributions, while the DFB performs weighted feature information fusion from the reference image, avoiding complex structures and redundant parameters. We tested LCDFNet on the IXI and BranTs2018 datasets, where it outperformed current state-of-the-art methods, reducing the number of parameters and computations by 90%. Compared with larger models such as SANet, LCDFNet is the first model to maintain performance while keeping the parameter count limited to 400KB and Floating-Point Operations per second (FLOPs) limited to 6G. By significantly reducing model size and computation, this study provides a promising direction toward lightweight deployment on medical devices.</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

Lightweight super-resolution dynamic contrast-aware differential fusion network for multi-contrast MRI

  • 期刊论文

  • 发表日期：

    2025年9月26日

查看全文
  • Wenjing Chen,
  • Longfeng Shen,
  • Jiacong Chen,
  • Jianyu Xiao,
  • Fangzhen Ge,
  • Li Shi

摘要

Multi-contrast magnetic resonance imaging (MRI) super-resolution (SR) methods effectively improve SR quality by using high-resolution reference images with similar anatomical structures. However, some medical devices have limited computational power to deploy multi-contrast MRI SR algorithms. Existing multi-contrast MRI SR methods mainly rely on computationally intensive large-scale models, making lightweight deployment of medical devices challenging. We propose a lightweight SR dynamic contrast-aware differential fusion network (LCDFNet) based on multi-contrast MRI. LCDFNet mainly consists of contrast-aware differential fusion modules (DCDFM). DCDFM includes two components: a lightweight dynamic contrast-aware channel attention block (DCAB) for MR images and a concise and efficient differential fusion block (DFB). The DCAB enhances edge feature extraction by dynamically adapting to different feature distributions, while the DFB performs weighted feature information fusion from the reference image, avoiding complex structures and redundant parameters. We tested LCDFNet on the IXI and BranTs2018 datasets, where it outperformed current state-of-the-art methods, reducing the number of parameters and computations by 90%. Compared with larger models such as SANet, LCDFNet is the first model to maintain performance while keeping the parameter count limited to 400KB and Floating-Point Operations per second (FLOPs) limited to 6G. By significantly reducing model size and computation, this study provides a promising direction toward lightweight deployment on medical devices.