Deep-Learning Based Multi-Joint Synchronous Tracking for Objective Quantification of Hindlimb Locomotor Kinematics in Rats

Summary

This protocol establishes a rat hindlimb kinematic evaluation pipeline using a markerless treadmill test with deep learning-driven multi-joint trajectory auto-labeling, which enables reproducible motion quantification.

Abstract

Accurate and objective assessment of rat gait is essential to neuroscience and kinesiology research fields. Conventional systems frequently depend on footprint imaging for indirect gait inference, which cannot fully capture hindlimb multi-joint kinematics. This study establishes a markerless treadmill-based gait analysis system for rodents, integrating custom deep learning algorithms with a programmable weight-support treadmill to enable real-time tracking of multiple lower-limb joints and multidimensional kinematic quantification. The system automatically extracts gait cycle parameters, joint trajectories, force distribution patterns, and movement smoothness under modulated speed (0-300 mm/s), incline (±30° gradient), and graded weight support (0-500 g) conditions, providing an objective assessment tool for neuromuscular behavior research. Using spinal cord injury (SCI) models, results demonstrate the system's sensitivity in detecting multidimensional differences in joint range of motion, trajectory continuity, propulsive force output, and movement smoothness between healthy and injured rodents, validating its disease discrimination and grading capabilities. Compared to traditional subjective scores or footprint methods, this platform eliminates personal bias and low-dimensional data limitations while combining deep learning-driven architecture and high-throughput data collection advantages. It is applicable to research involving central or peripheral nerve injury, neurodegenerative diseases, musculoskeletal disorders, and aging processes. Through synchronous integration of the neuroelectrophysiology modules, the system further enables temporal coupling of gait and neural signals, establishing a methodological framework for parsing central-peripheral control mechanisms and developing neuromodulatory strategies.

Introduction

In contemporary neuroscience and locomotor research, objective gait quantification is now imperative to dissecting neural-muscular functions, elucidating disease mechanisms, and evaluating rehabilitation strategies¹. Its high spatiotemporal resolution enables capture of subtle variations in synergies among neural commands, muscle activation profiles, and skeletal mechanics, thus establishing an empirical foundation for investigating the coordination of supraspinal pathways, peripheral feedback, and mechanical contexts. Reproducible kinematic metrics enhance detection sensitivity for lesions, degenerative changes, and therapeutic effects, accelerating the drug and device development pipeline².

Conventional assessment platforms such as CatWalk, DigiGait, or simplified footprint analyses indirectly infer joint kinematics based on plantar contact patterns. These approaches yield low-resolution and inconsistent metrics that fail to capture hindlimb multi-joint kinematic complexity^3,4. The overall goal of our method is to establish a high-precision, markerless, and weight-supported gait analysis system specifically designed for rodent models. Therefore, we establish a novel framework that: (i) maintains rodents in an upright posture under controlled axial loading (0-100% body weight) during continuous multi-joint imaging, (ii) directly localizes anatomical landmarks of the sagittal plane from lateral view analysis through deep learning to generate 2D spatial trajectories, (iii) Exports time-resolved joint angles and segmental kinematics for statistical modeling⁵. Fundamentally, this framework is grounded in the application of deep learning to computer vision tasks, which enables robust joint feature extraction from video data.

Compared to existing systems, this platform delivers four key advantages. Its markerless operation eliminates reflective or implanted tags while significantly reducing procedural complexity and motion artifacts⁶. The integrated force-feedback suspension system enables precise graded body-weight support (0-100%), facilitating comprehensive gait assessment across the full loading-to-unloading continuum, which is absent in commercial treadmills⁷. The deep learning architecture permits rapid retraining for novel strains, ages, or pathological conditions, outperforming manual tracking in scalability⁸. And high-accuracy multi-joint spatial coordinates enable fine-grained analysis of inter-joint kinematics coupled with quantitative profiling of aberrant gait patterns.

This protocol is widely applicable. It enables investigation of supraspinal control, proprioceptive feedback loops, and muscle synergies under modulated mechanical demands⁹. Pathophysiologically, it supplies sensitive locomotor biomarkers for spinal cord injury (SCI), stroke, neurodegeneration, peripheral neuropathies, and neuromuscular disorders^8,10. Therapeutically, it supports longitudinal evaluation of pharmacologic or device-based interventions with enhanced statistical power from high-resolution kinematics¹¹. The modular architecture allows seamless integration with complementary platforms, for example, electromyogram (EMG) or brain signals, facilitating mechanistic decoding of neural correlates in lower-limb motion.

Protocol

All animal procedures were approved by the Institutional Animal Care and Use Committee of the Beijing Institute for Brain Research and complied with the Guidelines for the Care and Use of Laboratory Animals. No procedure-related adverse events were observed.

1. Equipment preparation

1. Turn on the power supply of the treadmill system and select the Speed Control button on the touchscreen.
2. Let the belt run at a low speed for 30 s to observe whether the treadmill belt deviates.
3. Select the Incline Control button, then set the incline to 0°, click Zero, and confirm that the displayed value coincides with the level bubble.
4. Enter the Suspension interface, hang a 10 g weight, and check whether the tension reading is close to 0.10 N.
5. Turn on the Industrial Personal Computer (IPC) and enter the video acquisition and analysis software.
6. Create a directory for the day's experiment, including information such as the experiment name, animal ID, experimental group, speed, incline, and weight reduction percentage.
7. Fix the high-speed industrial camera (UC70, resolution: 1280X1024, frame rate: 60 fps), equipped with a manual focus lens (f = 12 mm), on an aluminum profile bracket. Position the camera lens perpendicular to the treadmill's longitudinal axis to capture a strictly lateral view, maintaining a horizontal working distance of 15 cm from the treadmill belt plane.
8. Place the lens horizontally using a bubble level and ensure that the rat's torso occupies the central 2/3 of the frame. Adjust the focus and aperture until the scale lines on the back ruler (1 cm grid) are clearly visible.
9. Click Preview to confirm that there is no trailing and the background brightness is uniform. If the image is too dark, use a light-emitting diode (LED) backlight to increase brightness; if overexposed, reduce the gain and narrow the aperture appropriately.
10. Click Start Recording to test the data recording quality. Click Stop Recording and set the data storage path. After completing the above equipment tests, begin animal testing.

2. Animal preparation and running test

1. Transfer the rat along with its original cage to the experimental room and let it rest for 10 min. During this period, allow free access to food and water.
2. Record the body weight using an electronic scale (accurate to 0.1 g). Include rats with matched body weights (±10 g) to minimize the confounding effects of body size.
3. If the fur color of the hind limbs does not contrast sufficiently with the background, use an electric shaver to gently remove a 1 cm wide strip of fur on the outer side of the thigh and the back of the lower leg to improve the accuracy of joint recognition. During the shaving procedure, the rat is manually restrained by an experienced handler without sedation to minimize animal distress. Monitor rats for signs of stress and exhaustion, such as reluctance to move and prolonged defecation.
4. Then use a soft tape to measure the straight-line distance from the hip to the ankle (mm), which will serve as the basis for subsequent stride normalization.
5. Put on gloves, grasp the rat, and loop the elastic chest strap around the front armpit, attaching it to the adjustable slide rail. Ensure that the strap is loose enough to insert one finger without affecting the animal's breathing.
6. Slowly rotate the suspension arm handle while observing the tension value on the screen, gradually increasing it to the target weight reduction ratio. Confirm that the rat's hind limbs still fully contact the ground.
7. Adjust the grip bar height. Loosen the top knob and position the front grip bar slightly below the rat's nose tip when standing upright, ensuring that the animal can grasp it easily without having to tilt its head back excessively or bend over.
8. On the touch control screen, set the speed to 150 mm/s and the incline to 0°. Conduct a 10-min acclimation session to adapt the animal to the specific body-weight support level. Monitor the rat for stress signs, e.g., freezing, defecation.
   1. If the gait remains irregular (defined as [>3 interruptions/min]), extend the duration by 2-min increments, up to a maximum of 20 min. If a rat exhibits persistent paw dragging or severe signs of distress, the acclimation session is immediately stopped. The animal is then returned to its home cage and allowed to rest for at least 24 hours before a subsequent attempt. A maximum of 3 acclimation attempts is permitted; animals that fail all attempts are excluded from the formal experiment.
9. The acclimation is considered successful when the rat maintains a continuous, uniform stride for at least 60 s without paw dragging, and the tail hangs naturally without stiffness. Exclude animals that fail to meet these criteria after the maximum acclimation period.
10. Enter the parameters required for the formal experiment. Sequentially input the desired speed (any value between 0-300 mm/s) and incline (−30° to +30°) on the touchscreen, and select the direction of the treadmill belt.
11. Conduct pre-training for five consecutive days before the formal experiment: walk for 10 min daily at 150 mm/s, 0° incline, and zero weight reduction, observing the gait.
12. After starting the treadmill, observe the rat's walking posture. If it deviates from the central 1/3 area, gently tap the right side wall once with the knuckle. Most animals will correct their position on their own. If two consecutive corrections are ineffective, reposition the rat and re-record the trial. Give the animal a 3–5 min break to rest in a neutral environment if repositioning and re-recording of the trial are necessary.
13. After stabilization, click Start Recording and continuously capture at least five complete gait cycles.
14. At the end of each trial, immediately reduce the speed to 0, unclip the chest strap, and return the rat to its corresponding cage.
15. At the end of the trial, the animal is placed back into its home cage, where water is provided ad libitum via a standard water bottle.
16. Record at least three trials under the same conditions. If an animal runs continuously for more than six trials, add an additional 2-min rest to prevent cumulative fatigue.

3. Gait data extraction

1. First, extract segments containing the target gait, usually around 10 s, which should include at least 10 analyzable stable gait cycles.
2. Import the .mov file into the analysis software and register the individual information of the rats, including ID number, group, and experimental conditions.
   NOTE: The system employs the YOLO-NAS Pose architecture, a state-of-the-art neural network optimized for high-performance pose estimation.
   Dataset Construction: The model was trained on a robust dataset of 6,000 video frames. These frames were manually annotated with the coordinates of five key hindlimb joints (iliac crest, hip, knee, ankle, and toe) using the DeepLabCut graphical user interface to ensure ground-truth precision.
   Training: The network underwent transfer learning on an NVIDIA RTX 3080 GPU to adapt the YOLO-NAS Pose weights to the specific features of the rat hindlimb.
   Initialization: In the software, select the pre-trained 'Rat_Hindlimb_YOLO_v1' model and set the Confidence Threshold to 0.8 to exclude low-certainty keypoints.
   During the analysis phase, the YOLO-NAS Pose model processes the video stream frame-by-frame. Unlike traditional computer vision methods that rely on background subtraction, this deep learning architecture directly localizes the rat and predicts joint coordinates within the bounding box.
   Post-Processing: The algorithm extracts x and y coordinates from the probability maps. To mitigate high-frequency jitter, a Median filter (window size: 3 frames) is applied to the raw trajectories.
   Output: The system successfully locks onto the five target joints within the video frame, generating time-resolved spatial coordinates for the entire gait sequence. These data are automatically exported as .csv files for kinematic calculation. Using the backboard ruler (1 cm grid) captured simultaneously, the system establishes a conversion relationship, automatically calculates the scaling factor, and batch-converts the pixel coordinates into millimeters.
3. Export the spatial position coordinates of each joint over time as a .csv file for subsequent index calculation.
   NOTE: The software automatically calculates stride length, step frequency, and foot height, and exports the changes in spatial position and angles of the five lower limb joints over time. Based on toe height, the software automatically divides each step into swing phase and stance phase, and calculates the proportion of swing phase and stance phase within each gait cycle.

4. Inter-group kinematic comparisons

1. Normalize each gait cycle within the group from 0% (foot contact) to 100% (same foot contact again) to unify the different cycle lengths, facilitating group averaging and comparison. The types of images that can be generated to illustrate gait dynamics are provided in Supplementary File 1.

5. Statistical analysis

1. Perform all statistical analyses using GraphPad Prism software.
2. Verify the normality of data distribution using the Shapiro-Wilk test.
3. For the kinematic validation and body mass index (BMI) demonstration comparisons, present data as Mean ± SEM.
4. Evaluate differences between the two experimental conditions using an unpaired Student's t-test.
5. Define statistical significance thresholds as *p < 0.05, **p < 0.01, and ***p < 0.005. Note that a total of n = 3 rats were included in this pilot validation study.

6. Embedding of the electrophysiological module

NOTE: To create the experimental model, animals were anesthetized using isoflurane (3% for induction, 1.5%–2% for maintenance). Buprenorphine (0.05 mg/kg) was administered to provide perioperative analgesia. Following standard skin preparation and sterile draping, target tissues were exposed to secure the recording electrodes based on the specific modality: the skull surface for EEG, the intervertebral foramen for spinal cord recordings, and the target hindlimb muscles for EMG. Animals underwent daily post-operative monitoring. Analgesic administration was continued for the first three days post-surgery to ensure adequate pain management. Under the identical anesthetic protocol described above, a standard T10 laminectomy was performed. A contusion injury was subsequently induced using a calibrated weight-drop device prior to wound closure. All procedures strictly adhered to institutional animal care guidelines.

1. After completing the animal modeling, implant the electrophysiological recording device.
2. For neural recording, place electrodes on the skull surface, epidural space, or cerebral cortex to record brain signals.
3. For spinal cord recording, insert recording electrodes into the epidural space of the intervertebral foramen.
4. For EMG recording, bury bipolar silver wires into muscles to record muscle electrical activity.
5. After the implantation, allow the animal to recover for 5-7 days, checking the wound and gait daily to ensure no signs of infection or pain before proceeding to the treadmill experiment. Allow the animal to undergo a full 5–7 day recovery period in its home cage to monitor for a return to normal baseline gait and ensure proper wound healing. No gait recording or treadmill testing is conducted on the device during this recovery period.
6. Complete the equipment and animal preparations as described above.
7. Synchronize the collection of electrophysiological parameters and movement videos, ensuring that the electrophysiological and video data share the same timestamp.
8. After data collection, align the neural signals with video frames during analysis to intuitively observe the electrophysiological patterns at different phases of the gait cycle, providing direct evidence for exploring the central-peripheral coupling mechanism.

Results

The schematic diagram and actual photo of the treadmill system structure are shown in Figure 1A,B, respectively. The device is capable of obtaining high-contrast lateral images without markers. The treadmill supports adjustments for speed, incline, direction, and weight reduction, and is equipped with a numerical control panel. The accompanying software interface includes a real-time monitoring and acquisition interface, as well as a video upload and gait extraction analysis...

Discussion

This protocol delineates a treadmill-based gait analysis system for rats and its operational procedures, providing a multidimensional and objective assessment tool for neuromuscular behavioral research. As demonstrated in the intergroup comparisons of gait parameters between SCI and healthy rats within the exemplar study, the system acquires raw spatial movement coordinates of subjects' lower limbs through continuous high-frame-rate gait videography and an AI-driven automated multi-joint annotation algorithm. This en...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Gait AI Annotation & Acquisition Software	Shanghai ClinAITrial Technology Co., Ltd.	MouseAI	MouseAI is the integrated analysis software for the MOVEMETRICS treadmill, unifying video recording, metadata logging, and automated gait analysis into a single workflow.
Mechanical impact	RWD Life Science, Shenzhen, China
Multi-channel electrophysio logical recording system	Kedou (suzhou) brain- computer Technology Co. ,	KedouBC RHD128A	KedouBC RHD128A is a compact, 128-channel in vivo electrophysiology acquisition system engineered for high-resolution neural recording in small to large laboratory animals. Integrating lightweight, plug-and-play headstages (≤ 1.3 g), 16-bit A/D conversion and programmable amplifiers, it captures EEG, ECoG, LFP, spikes, EMG and ECG signals from 0.1 Hz to 30 kHz with ≤ 2.4 µV system noise and > 10 GΩ input impedance. RHD128A delivers robust, user- scalable neural data to accelerate studies in brain-machine interfaces, neurodegeneration and rehabilitation.
Treadmill	Shanghai ClinAITrial Technology Co., Ltd.	MOVEMETRICS	MOVEMETRICS is a high-precision instrumented treadmill for advanced rodent gait analysis.