基于PoseMesh框架的移动式三维步态分析在脊柱侧凸筛查中的应用研究

Research on the application of PoseMesh framework-based mobile three-dimensional gait analysis in scoliosis screening

中文关键词:  三维步态分析  脊柱侧凸  人体网格重建

英文关键词:3D gait analysis  Scoliosis  Human mesh reconstruction

中文摘要:

  【摘要】 目的：提出一种轻量化的AI驱动三维步态分析框架———PoseMesh，对人体姿态步态变换引发的脊柱变化相关参数在空间上进行精确估计，提升青少年特发性脊柱侧凸（adolescent idiopathic scoliosis，AIS）筛查的准确性与实用性。方法：步态数据来源于香港大学临床医学院数智健康实验室，经数据处理后保留138段高质量侧视角步态视频，均来源于香港大学医学院矫形及创伤外科学系收治的AIS患者，其中男性58例（42.0%），女性80例（58.0%）；检查时年龄为10～62岁（19.8±8.0岁），中位数为17.8（14.9～22.9）岁；按年龄段分布：10～14岁36例（26.1%），15～19岁50例（36.2%），20～24岁30例（21.7%），25～29岁13例（9.4%），≥30岁9例（6.5%）。患者身高165.2±9.9（125～185）cm，体重56.1±14.2（22～124）kg，体质指数（body mass index，BMI）为20.4±4.0（14.2～45.4）kg/m2。采用自研PoseMesh框架对这些RGB视频进行逐帧三维人体网格重建，并提取步态周期、步幅、步态不对称指数、骨盆倾角、躯干倾斜角及步行速度等6项关键指标。为进行对比，所有视频亦通过OpenPose进行二维关节点估计并提取对应参数。从参数精度、重复性及鲁棒性方面对两种方法进行系统评估比较，采用配对t检验与均方根误差（root mean squared error，RMSE）指标进行统计分析。结果：PoseMesh框架在所有视频中均成功完成三维重建。相较于二维分析，三维方法在各项参数的稳定性与重测一致性方面均显著提升：重复实验间RMSE平均降低42%；步幅误差由6.8cm降至3.2cm；骨盆倾角标准差由±3.6°降至±1.9°（提升47%）。三维方法在检测侧向运动偏差方面更具敏感性（不对称指数分布更集中），且估算的步行速度与实测结果误差更小（±3.5% vs ±8.4%）。近95%的序列无须人工干预即可完成全流程处理，体现出高鲁棒性。结论：本研究提出的PoseMesh框架实现了基于单目视频的高精度三维步态分析，在无须穿戴式设备的前提下，能够可靠地提取出与脊柱姿态相关的动态参数。相较传统二维方法，三维分析在关键指标上表现出更高的准确性与重复性，为功能性、生物力学驱动的AIS智能筛查提供了新方式。

英文摘要:

  【Abstract】 Objectives: To develop and validate a lightweight and AI-driven 3D gait analysis framework, PoseMesh, to estimate in spatial the key gait parameters relevant to spinal posture precisely thereby improving adolescent idiopathic scoliosis(AIS) screening accuracy. Methods: Gait data was sourced from the Digital Health Laboratory at the University of Hong Kong. After data processing, 138 high-quality lateral-view gait videos were retained, which were from the AIS patients admitted to the Department of Orthopedics and Traumatology, Faculty of Medicine, the University of Hong Kong. There were 58 males(42.0%) and 80 females(58.0%). The age at the time of examination ranged from 10 to 62 years(19.8±8.0years), with a median of 17.8(14.9-22.9) years. Distribution by age group: 36 cases(26.1%) were aged 10-14 years, 50 cases(36.2%) were aged 15-19 years, 30 cases(21.7%) were aged 20-24 years, 13 cases(9.4%) were aged 25-29 years, and 9 cases(6.5%) were aged 30 years or above. The height of the patients was 165.2±9.9(125-185)cm, the weight was 56.1±14.2(22-124)kg, and the body mass index(BMI) was 20.4±4.0(14.2-45.4)kg/m2. The proprietary PoseMesh framework was used for frame-by-frame 3D human mesh reconstruction from RGB videos. Six key gait parameters were extracted: gait cycle, stride length, step asymmetry index, pelvic tilt, trunk inclination, and walking velocity. For comparison, 2D joint estimation was performed using OpenPose. The two methods were systematically evaluated for parameter accuracy, repeatability, and robustness using paired t-tests and root mean square error(RMSE) analysis. Results: The PoseMesh framework successfully performed 3D reconstruction on all videos. Compared to 2D analysis, the 3D method demonstrated significantly improved stability and test-retest reliability across all parameters. The average RMSE between repeated experiments decreased by 42%; Stride length error was reduced from 6.8cm to 3.2cm; And the standard deviation of pelvic tilt improved from ±3.6° to ±1.9°(a 47% improvement). The 3D method was more sensitive in detecting lateral deviations, and its estimated walking velocity showed lower error against ground truth(±3.5% vs ±8.4%). Nearly 95% of sequences were processed end-to-end without manual intervention, indicating high robustness. Conclusions: The proposed PoseMesh framework enables high-precision, 3D gait analysis from monocular video without wearables or markers. It reliably extracts dynamic parameters related to spinal posture with superior accuracy and repeatability over traditional 2D methods， which provides a new paradigm for functional, biomechanics-driven intelligent AIS screening.