Scholarlink

<p>Magnetic resonance (MR) imaging is a cornerstone of modern clinical diagnosis, yet its quality is often limited by acquisition constraints, potentially obscuring subtle pathological details and hindering diagnostic accuracy. To address this, we propose a novel multi-head attention network (MHAN) for MR image super-resolution, specifically engineered to capture and reconstruct high-dimensional medical features. The core of MHAN features two key innovations: a 3D multi-head attention module (3D-MHead) and a depth–width feature fusion module (DW-FFM). The 3D-MHead employs multi-scale 3D convolutions to dynamically recalibrate feature weights, enhancing the model’s focus on critical high-frequency details such as lesion boundaries. Concurrently, the DW-FFM effectively integrates shallow spatial textures with deep semantic information, ensuring both global structural coherence and local feature fidelity in the final output. Our extensive experiments on public MR datasets, such as IXI, demonstrate that MHAN significantly outperforms state-of-the-art methods in both quantitative metrics (PSNR/SSIM) and qualitative visual assessment. For instance, on the IXI-PD dataset at a <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times 4\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>×</mo> <mn>4</mn> </mrow> </math></EquationSource> </InlineEquation> scale, MHAN achieves a PSNR of 33.16 dB and an SSIM of 0.9439, showcasing superior detail recovery. The model’s strong generalization capabilities are further validated on standard natural image benchmarks. This work presents an effective and efficient solution for enhancing MR image quality, holding considerable promise for advancing diagnostic precision in clinical practice. The source code is available at <a href="https://github.com/Etleventt/MHAN">https://github.com/Etleventt/MHAN</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

Enhancing MR image super-resolution with a multi-head attention network: capturing high-dimensional medical features

  • 期刊论文

  • 发表日期：

    2025年12月16日

查看全文
  • Xiaobin Pei,
  • Feihong Zhu,
  • Zhen Wang,
  • Jiaqi Huang,
  • Kehua Guo,
  • Wei Zheng

摘要

Magnetic resonance (MR) imaging is a cornerstone of modern clinical diagnosis, yet its quality is often limited by acquisition constraints, potentially obscuring subtle pathological details and hindering diagnostic accuracy. To address this, we propose a novel multi-head attention network (MHAN) for MR image super-resolution, specifically engineered to capture and reconstruct high-dimensional medical features. The core of MHAN features two key innovations: a 3D multi-head attention module (3D-MHead) and a depth–width feature fusion module (DW-FFM). The 3D-MHead employs multi-scale 3D convolutions to dynamically recalibrate feature weights, enhancing the model’s focus on critical high-frequency details such as lesion boundaries. Concurrently, the DW-FFM effectively integrates shallow spatial textures with deep semantic information, ensuring both global structural coherence and local feature fidelity in the final output. Our extensive experiments on public MR datasets, such as IXI, demonstrate that MHAN significantly outperforms state-of-the-art methods in both quantitative metrics (PSNR/SSIM) and qualitative visual assessment. For instance, on the IXI-PD dataset at a \(\times 4\) × 4 scale, MHAN achieves a PSNR of 33.16 dB and an SSIM of 0.9439, showcasing superior detail recovery. The model’s strong generalization capabilities are further validated on standard natural image benchmarks. This work presents an effective and efficient solution for enhancing MR image quality, holding considerable promise for advancing diagnostic precision in clinical practice. The source code is available at https://github.com/Etleventt/MHAN.