Scholarlink

<p>Lumbar disc herniation (LDH) is a predominant cause of low back pain, significantly affecting the quality of life of patients and closely linked to the condition of paraspinal muscles (PSM). However, existing two-dimensional (2D) measurement techniques that rely on muscle cross-sectional area exhibit limitations and fail to accurately represent the true state of these muscles. A series of previous studies conducted by our team has demonstrated that the semi-automatic reconstruction of paraspinal muscles using CT images yields muscle data that more comprehensively reflects the status of these muscles. This study innovatively constructs a lightweight 2D U-Net-based deep learning model to investigate the changing patterns of PSM in patients with LDH. This retrospective study included patients diagnosed with unilateral L4-5 lumbar disc herniation between January 2022 and August 2025. A lightweight 2D U-Net model was employed to achieve automatic segmentation of PSM. Muscle volume (MV), relative muscle volume (RMV), effective muscle volume (EMV), and fat infiltration rate (FI) were assessed through three-dimensional (3D) reconstruction, and comparative analyses were conducted between the involved and uninvolved sides. The study included a total of 144 patients. The lightweight 2D U-Net model developed in this study exhibited strong segmentation performance for PSM, achieving Dice coefficients greater than 0.80 and pixel accuracy (PA) exceeding 0.90. In Group A, the MV and EMV of the psoas (PS) at the L1-L4 levels were significantly greater on the uninvolved side compared to the involved side. In Group B, the FI on the involved side were significantly higher than those on the uninvolved side for the multifidus (MF) at the L5 level, as well as for the erector spinae (ES) at the L4 and L5 levels, and the psoas (PS) at the L4 and L5 levels. Significant differences were observed in the MV, EMV, and FI of the PSM on both sides of the spine in patients with LDH. The findings indicate that the PSM in LDH patients exhibit varying degrees of change at different time points in the lumbar spine.</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

Deep learning-based analysis of paraspinal muscle changes in lumbar disc herniation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月1日

查看全文
  • Sa Feng,
  • Zehua Jiang,
  • Wentao Liu,
  • Qing Lu,
  • Xiangyu Wang,
  • Mengmeng Zhou,
  • Xin Tan,
  • Shengbo Huang,
  • Haifan Zhao,
  • Ru Tao,
  • Ruhuan Liu,
  • Rusen Zhu,
  • Baoshan Xu

摘要

Lumbar disc herniation (LDH) is a predominant cause of low back pain, significantly affecting the quality of life of patients and closely linked to the condition of paraspinal muscles (PSM). However, existing two-dimensional (2D) measurement techniques that rely on muscle cross-sectional area exhibit limitations and fail to accurately represent the true state of these muscles. A series of previous studies conducted by our team has demonstrated that the semi-automatic reconstruction of paraspinal muscles using CT images yields muscle data that more comprehensively reflects the status of these muscles. This study innovatively constructs a lightweight 2D U-Net-based deep learning model to investigate the changing patterns of PSM in patients with LDH. This retrospective study included patients diagnosed with unilateral L4-5 lumbar disc herniation between January 2022 and August 2025. A lightweight 2D U-Net model was employed to achieve automatic segmentation of PSM. Muscle volume (MV), relative muscle volume (RMV), effective muscle volume (EMV), and fat infiltration rate (FI) were assessed through three-dimensional (3D) reconstruction, and comparative analyses were conducted between the involved and uninvolved sides. The study included a total of 144 patients. The lightweight 2D U-Net model developed in this study exhibited strong segmentation performance for PSM, achieving Dice coefficients greater than 0.80 and pixel accuracy (PA) exceeding 0.90. In Group A, the MV and EMV of the psoas (PS) at the L1-L4 levels were significantly greater on the uninvolved side compared to the involved side. In Group B, the FI on the involved side were significantly higher than those on the uninvolved side for the multifidus (MF) at the L5 level, as well as for the erector spinae (ES) at the L4 and L5 levels, and the psoas (PS) at the L4 and L5 levels. Significant differences were observed in the MV, EMV, and FI of the PSM on both sides of the spine in patients with LDH. The findings indicate that the PSM in LDH patients exhibit varying degrees of change at different time points in the lumbar spine.