Gait analysis in cats and other animals is generally performed with custom-made or commercially developed software to track reflective markers placed on bony landmarks. This often involves costly motion tracking systems. However, deep learning, and in particular DeepLabCutTM (DLC), allows motion tracking without requiring placing reflective markers or an expensive system. The purpose of this study was to validate the accuracy of DLC for gait analysis in the adult cat by comparing results obtained with DLC and a custom-made software (Expresso) that has been used in several cat studies. Four intact adult cats performed tied-belt (both belts at same speed) and split-belt (belts operating at different speeds) locomotion at different speeds and left-right speed differences on a split-belt treadmill. We calculated several kinematic variables, such as step/stride lengths and joint angles from the estimates made by the two software and assessed the agreement between the two measurements using intraclass correlation coefficient or Lin's concordance correlation coefficient as well as Pearson's correlation coefficients. The results showed that DLC is at least as precise as Expresso with good to excellent agreement for all variables. Indeed, all 12 variables showed an agreement above 0.75, considered good, while nine showed an agreement above 0.9, considered excellent. Therefore, deep learning, specifically DLC, is valid for measuring kinematic variables during locomotion in cats, without requiring reflective markers and using a relatively low-cost system.

中文翻译：

---

用于猫步态分析的监督深度学习验证。

猫和其他动物的步态分析通常使用定制或商业开发的软件进行，以跟踪放置在骨骼标志上的反射标记。这通常涉及昂贵的运动跟踪系统。然而，深度学习，尤其是 DeepLabCutTM (DLC)，允许在不需要放置反射标记或昂贵的系统的情况下进行运动跟踪。本研究的目的是通过比较 DLC 和已在几项猫研究中使用的定制软件 (Expresso) 获得的结果，验证 DLC 在成年猫步态分析中的准确性。四只完整的成年猫在分裂带跑步机上以不同的速度和左右速度差异进行系带（两条带以相同速度）和分带（以不同速度运行的带）运动。我们计算了几个运动学变量，例如步/步长和关节角度来自两个软件的估计，并使用类内相关系数或林的一致性相关系数以及皮尔逊的相关系数评估两个测量值之间的一致性。结果表明，DLC 至少与 Expresso 一样精确，所有变量都具有良好到极好的一致性。事实上，所有 12 个变量的一致性都高于 0.75，被认为是好的，而 9 个变量的一致性高于 0.9，被认为是优秀的。因此，深度学习，特别是 DLC，对于测量猫运动过程中的运动学变量是有效的，不需要反射标记并使用相对低成本的系统。