Objectives Low back pain (LBP) is a leading cause of disability worldwide, yet current clinical assessments rely heavily on subjective reports and static imaging, providing limited objective quantification of spinal dynamic function. This study aims to develop and evaluate HumanMoveNet, a novel digital framework that reconstructs a temporally consistent 3D human model with precise spinal curvature from monocular visual data to enable objective LBP screening and rehabilitation assessment. Methods The proposed hybrid framework integrates static anatomical reconstruction, dynamic pose estimation, and temporal smoothing. From a human gait video, a 3D reconstruction network first generates a static human model with personalized spinal morphology. The gait video is then processed via optimized 2D pose estimation and parametric model regression to obtain frame-by-frame 3D human meshes. Graph convolutional and long short-term memory networks are employed to ensure temporal motion continuity. Finally, the static spine is fused with the dynamic pose sequence to create a “dynamic spine,” from which key biomechanical parameters—lumbar range of motion (ROM), pelvic tilt range, and spinal symmetry index—are extracted. Results Validation on 146 subjects demonstrated superior reconstruction performance, achieving a 5.6% improvement (18.85 vs 17.85) in Peak Signal-to-Noise Ratio (PSNR), a 28.4% reduction in Hausdorff Distance (2.1126 mm vs 2.9505 mm), and a 5.1% increase in Intersection over Union (IoU) (0.4122 vs 0.3921) compared with state-of-the-art methods. Analysis of spinal curvature variation showed <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\Delta \kappa \)</EquationSource> </InlineEquation> values of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(2.15 \pm 0.52^\circ \)</EquationSource> </InlineEquation> in females and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(1.98 \pm 0.61^\circ \)</EquationSource> </InlineEquation> in males, with no significant gender difference (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(p = 0.066\)</EquationSource> </InlineEquation>). Gender-specific analysis further revealed that females had greater pelvic mobility (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(6.5 \pm 2.1^\circ \)</EquationSource> </InlineEquation> vs <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(5.2 \pm 1.8^\circ \)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(p < 0.05\)</EquationSource> </InlineEquation>) and lumbar ROM (<InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(18.5 \pm 4.2^\circ \)</EquationSource> </InlineEquation> vs <InlineEquation ID="IEq9"> <EquationSource Format="TEX">\(16.8 \pm 4.7^\circ \)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq10"> <EquationSource Format="TEX">\(p = 0.024\)</EquationSource> </InlineEquation>). Conclusions HumanMoveNet provides a precise, label-free solution for assessing spinal dynamic function using conventional visual data. By combining high-fidelity spinal anatomy with dynamic motion analysis, it effectively captures LBP-related movement alterations, and shows strong potential for community-based screening, rehabilitation evaluation, and personalized care.

HumanMoveNet: a dynamic 3D spine reconstruction framework for low back pain screening and rehabilitation assessment

期刊论文
出版模式：

开放获取
发表日期：

2026年5月4日

查看全文

Tao Huang,
Zhiyuan Xia,
Jason P. Y. Cheung,
Yong Hai,
Xihe Kuang,
Teng Zhang

摘要

Objectives

Low back pain (LBP) is a leading cause of disability worldwide, yet current clinical assessments rely heavily on subjective reports and static imaging, providing limited objective quantification of spinal dynamic function. This study aims to develop and evaluate HumanMoveNet, a novel digital framework that reconstructs a temporally consistent 3D human model with precise spinal curvature from monocular visual data to enable objective LBP screening and rehabilitation assessment.

Methods

The proposed hybrid framework integrates static anatomical reconstruction, dynamic pose estimation, and temporal smoothing. From a human gait video, a 3D reconstruction network first generates a static human model with personalized spinal morphology. The gait video is then processed via optimized 2D pose estimation and parametric model regression to obtain frame-by-frame 3D human meshes. Graph convolutional and long short-term memory networks are employed to ensure temporal motion continuity. Finally, the static spine is fused with the dynamic pose sequence to create a “dynamic spine,” from which key biomechanical parameters—lumbar range of motion (ROM), pelvic tilt range, and spinal symmetry index—are extracted.

Results

Validation on 146 subjects demonstrated superior reconstruction performance, achieving a 5.6% improvement (18.85 vs 17.85) in Peak Signal-to-Noise Ratio (PSNR), a 28.4% reduction in Hausdorff Distance (2.1126 mm vs 2.9505 mm), and a 5.1% increase in Intersection over Union (IoU) (0.4122 vs 0.3921) compared with state-of-the-art methods. Analysis of spinal curvature variation showed \(\Delta \kappa \) values of \(2.15 \pm 0.52^\circ \) in females and \(1.98 \pm 0.61^\circ \) in males, with no significant gender difference ( \(p = 0.066\) ). Gender-specific analysis further revealed that females had greater pelvic mobility ( \(6.5 \pm 2.1^\circ \) vs \(5.2 \pm 1.8^\circ \) , \(p < 0.05\) ) and lumbar ROM ( \(18.5 \pm 4.2^\circ \) vs \(16.8 \pm 4.7^\circ \) , \(p = 0.024\) ).

Conclusions

HumanMoveNet provides a precise, label-free solution for assessing spinal dynamic function using conventional visual data. By combining high-fidelity spinal anatomy with dynamic motion analysis, it effectively captures LBP-related movement alterations, and shows strong potential for community-based screening, rehabilitation evaluation, and personalized care.