(L-442) Evaluating Neural Network Approaches to Kinematic Measures of Rodent Gait Analysis

Introduction::

This study aimed to integrate DeepLabCut (DLC) with our existing technology to evaluate spatial, temporal, and dynamic characteristics of rodent gait. Prior work with Gait Analysis Instrumentation and Technology Optimized for Rodents (GAITOR) shows that it can be used to understand behavioral changes in multiple disease models. However, GAITOR cannot describe kinematic variables. Conversely, DLC is an effective tool for estimating posture and measuring kinematic angles in lab animals. Here, we integrate DLC and GAITOR and compare the gait variables achieved with both methods.

Materials and Methods:: Videos of naïve Sprague-Dawley rats (12 w.o. males, n=18) walking unprompted across an arena were collected. First, hand digitization of high-speed videos was conducted to estimate the location of the nose, back, tail base, forepaws, hind toe, midfoot, and heel. These data were used to train a neural network within DLC, allowing DLC to then automate the digitization of these same points in subsequent videos. Then, a MATLAB pipeline was developed to calculate average stride profiles for toe height, back angle, and midfoot angle from DLC coordinates. Videos were also processed through GAITOR. Finally, to benchmark DLC relative to GAITOR, velocity, stride length, step width, and duty factor were calculated for both methods.

Results, Conclusions, and Discussions:: The standard profiles of toe and heel height, foot flexion, and back angle were calculated for naïve animals and regressed against velocity to identify measurements that were velocity dependent. The maximum height of the hind toe and hind heel did not show a dependency on velocity (Figure 1 a and b, respectively). The minimum angle of the hind foot at the 5th metatarsal was inversely related to velocity (p< 0.005, Figure 1c). The maximum forepaw height increased with velocity (p< 0.0001, Figure 1d). The median back angle increased with velocity (p< 0.0001, Figure 1e), indicating that healthy animals had more spinal curvature at lower velocities. The range of the length from the hind paw toe to the tail base increased with velocity (p< 0.0002) as animals stretched their hindlimb slightly longer at higher velocities. All measures had high variability between trials; thus, large numbers of trials are generally needed to obtain a precise prediction.

The DLC approach had larger variability than AGATHA in measuring velocity, fraction stance time of the hind limb, stride length, and step width. However, integrating DLC with GAITOR provided kinematic measures of foot flexion, toe clearance, and back angle that cannot be assessed in GAITOR. When used in tandem, DLC can enhance the evaluation of gait using GAITOR. However, for overlapping variables, GAITOR provided lower variability. Decisions on whether to use GAITOR alone, DLC alone, or GAITOR-DLC in tandem are thus dependent on the number of videos collected. For overlapping variables, DLC is accurate, but imprecise, relative to GAITOR; thus, DLC will require more trials for mean estimates.

In summary, this study advances the field by 1) comparing DeepLabCut to an existing method of gait analysis (GAITOR) and 2) providing recommendations on using both technologies for future studies. Additionally, we will release the customizable pipeline for calculating relevant gait parameters from data generated using an included DLC network. This will ensure that kinematic variables can be easily added and implemented for any researcher currently using gait analysis with a lateral view, even if the videos have already been collected.

Acknowledgements (Optional): :

References (Optional): : [1] B.Y. Jacobs, E.H. Lakes, A.J. Reiter, S.P. Lake, T.R. Ham, N.D. Leipzig, S.L. Porvasnik, C.E. Schmidt, R.A. Wachs, and K.D. Allen. The Open Source GAITOR Suite for Rodent Gait Analysis. Scientific Reports, (2018).