Rodent models are commonly used to understand the underlying mechanisms of spinal cord injury (SCI). Kinematic analysis, an important technique to measure dysfunction of locomotion after SCI, is generally based on the capture of physical markers placed on bony landmarks. However, marker-based studies face significant experimental hurdles such as labor-intensive manual joint tracking, alteration of natural gait by markers, and skin error from soft tissue movement on the knee joint. Although the pose estimation strategy using deep neural networks can solve some of these issues, it remains unclear whether this method is adaptive to SCI mice with abnormal gait. In the present study, we developed a deep learning based markerless method of 2D kinematic analysis to automatically track joint positions. We found that a relatively small number (< 200) of manually labeled video frames was sufficient to train the network to extract trajectories. The mean test error was on average 3.43 pixels in intact mice and 3.95 pixels in SCI mice, which is comparable to the manual tracking error (3.15 pixels, less than 1 mm). Thereafter, we extracted 30 gait kinematic parameters and found that certain parameters such as step height and maximal hip joint amplitude distinguished intact and SCI locomotion.

啮齿动物模型通常用于了解脊髓损伤 (SCI) 的潜在机制。运动学分析是测量 SCI 后运动功能障碍的一项重要技术，通常基于捕获放置在骨骼标志上的物理标记。然而，基于标记的研究面临重大的实验障碍，例如劳动密集型手动关节跟踪、标记改变自然步态以及膝关节软组织运动引起的皮肤错误。尽管使用深度神经网络的姿态估计策略可以解决其中一些问题，但尚不清楚该方法是否适用于步态异常的 SCI 小鼠。在本研究中，我们开发了一种基于深度学习的无标记二维运动分析方法来自动跟踪关节位置。我们发现一个相对较小的数字（< 200) 手动标记的视频帧足以训练网络提取轨迹。完整小鼠的平均测试误差为 3.43 像素，SCI 小鼠为 3.95 像素，与手动跟踪误差（3.15 像素，小于 1 毫米）相当。此后，我们提取了 30 个步态运动学参数，发现某些参数，如步高和最大髋关节振幅可区分完整运动和 SCI 运动。