Scholarlink

<p>Accurate segmentation of vertebral bodies and intervertebral discs in spinal magnetic resonance imaging (MRI) is fundamental for diagnosis and treatment planning. This task is challenged by pronounced anatomical homogeneity, pathological deformations, edge blurring, detail loss, and the computational burden of volumetric data. We propose FE-UNet, a hybrid CNN-Transformer framework that couples frequency-domain enhancement with multi-scale feature fusion. The network integrates a spectral preprocessing module that performs Fourier decomposition with adaptive separation of high/low-frequency content and dual-stage denoising to strengthen edge discrimination; an edge-attentive upsampling mechanism that fuses hierarchical features via dynamically generated spatial attention to mitigate information loss during resolution recovery; and a compact morphological refinement module that applies area-aware morphological denoising while preserving structural integrity by updating only changed pixels. On the public MRSpineSeg2021 benchmark, FE-UNet achieves 84.65% DSC, 75.91% IoU, 85.83% PPV, and 86.41% TPR, outperforming contemporary methods and demonstrating robustness under challenging imaging conditions and weak-boundary cases.</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

Frequency-domain enhanced and feature fusion network for spinal MR image segmentation

  • 期刊论文

  • 发表日期：

    2025年10月29日

查看全文
  • Deting Zeng,
  • Mingwei Wang,
  • Fei Mao,
  • Weizhuo Wang,
  • Tao Lei

摘要

Accurate segmentation of vertebral bodies and intervertebral discs in spinal magnetic resonance imaging (MRI) is fundamental for diagnosis and treatment planning. This task is challenged by pronounced anatomical homogeneity, pathological deformations, edge blurring, detail loss, and the computational burden of volumetric data. We propose FE-UNet, a hybrid CNN-Transformer framework that couples frequency-domain enhancement with multi-scale feature fusion. The network integrates a spectral preprocessing module that performs Fourier decomposition with adaptive separation of high/low-frequency content and dual-stage denoising to strengthen edge discrimination; an edge-attentive upsampling mechanism that fuses hierarchical features via dynamically generated spatial attention to mitigate information loss during resolution recovery; and a compact morphological refinement module that applies area-aware morphological denoising while preserving structural integrity by updating only changed pixels. On the public MRSpineSeg2021 benchmark, FE-UNet achieves 84.65% DSC, 75.91% IoU, 85.83% PPV, and 86.41% TPR, outperforming contemporary methods and demonstrating robustness under challenging imaging conditions and weak-boundary cases.