Establishment of a novel deep learning-based method for gait analysis after peripheral nerve injury in pigs

Abstract

Background

Peripheral nerve injury with deficits has poor functional prognosis, making motor function assessment during nerve regeneration crucial. Recently, pigs have been used as research animals for peripheral nerve regeneration from a translational perspective. However, the established methods for evaluating motor function remain insufficient, and the development of motor function assessment protocols for evaluating paralysis in medium-to large-sized animals is necessary.

Methods

Bilateral video recordings of healthy pigs (n = 4) walking straight ahead were obtained. A common peroneal nerve injury model was established on the left side of the animal (n = 3), and similar videos were recorded at 1 and 3 months post-operatively. Stride length, stance phase duration, swing phase duration, gait cycle duration, maximum heel height, joint angles, and foot velocity during one gait cycle were evaluated before and after surgery. Deep learning for posture estimation was employed to analyze joint angles and foot velocities.

Results

The post-operative stance phase duration on the intact side was prolonged and the maximum heel height on the impaired side was significantly higher after surgery. The accuracy of the posture estimated using deep learning was comparable to that estimated using manual human labeling. The ankle angle on the impaired side increased post-operatively and significant changes in foot velocity were observed at the end of the swing phase. Changes in walking patterns of the animals caused by nerve regeneration were also captured.

Conclusions

Deep learning-based gait analysis enabled the quantitative and objective identification of the characteristics of common peroneal nerve palsy and compensatory movements during one gait cycle. This analytical method is a potentially useful technique for studying recovery from paralysis associated with nerve regeneration.

Graphical abstract

Highlights

• •Deep learning (DL) gait analysis quantified motor deficits in pig nerve injury model.
• •Objective gait parameters identified common peroneal nerve palsy and compensation.
• •Posture estimation accuracy using DL was comparable to manual human labeling.
• •Ankle angle and foot velocity changes sensitively reflected motor impairment.
• •Our method enables translational assessment of nerve regeneration in large animals.

Abbreviations

AI

artificial intelligence

BBB

Basso, Beattie, and Bresnahan

CNNs

convolutional neural networks

CPN

common peroneal nerve

CT

computed tomography

DLC

DeepLabCut

MTP

metatarsophalangeal

Pre

before surgery

PO1M

1 month after surgery

PO3M

3 months after surgery

SFI

Sciatic Function Index

SN

sciatic nerve

TN

tibial nerve.

1. Introduction

Peripheral nerve injuries can impair both sensory and motor functions, consequently severely affecting patients’ quality of life. Globally, thousands of people suffer from peripheral injuries each year, with approximately 300,000 cases in Europe and 200,000 in the United States [1,2]. More than 50,000 patients with neurological disorders require surgical treatment annually [3].

To address the clinical challenges of peripheral nerve injury, various strategies have been explored for peripheral nerve regeneration. These include the direct application of methylcobalamin sheets to the site of nerve injury in animal models [4] and the development of peripheral nerve defect models mimicking traumatic nerve defects. In peripheral nerve defect models, conduits made of diverse biomaterials are used to bridge nerve defects by incorporating neurotrophic factors and cellular components to promote nerve regeneration [4].

Many studies have evaluated the therapeutic effects of nerve regeneration and provided important insights for future clinical applications. However, most nerve regeneration studies have been conducted using small animal models such as mice and rats. Although small animals are inexpensive and readily available, their nerve regeneration capacity is far greater than that of humans; therefore, they do not accurately reflect the clinical reality [5].

The ultimate goal of peripheral nerve regeneration research is clinical translation of these findings. Therefore, from the perspective of preclinical testing and safety, it is important to conduct experiments on large animals with physiological characteristics similar to those of humans. Although research using large animals requires careful ethical management, the number of experiments involving large animals has increased steadily in various fields in recent years.

Among large experimental animals, pigs have emerged as a particularly promising model for translational research. Pigs are abundant, accessible, anatomically and physiologically similar to humans, and can be maintained and transplanted under specific pathogen-free conditions. This has garnered attention in xenotransplantation research wherein pig hearts and kidneys are transplanted into humans [[6], [7], [8]]. Pigs are also used in peripheral nerve research because their neural structures are similar to those of humans [[9], [10], [11]].

Several constraints on the applications of basic research performed in pigs, particularly genetic heterogeneity and the complexity of genome editing, have been addressed in recent years [12,13]. Furthermore, there are currently no standardized methods for assessing motor function in pigs, which limits their utility in nerve regeneration research.

In contrast, there are well-established methods to evaluate motor function in small animals like mice and rats, including the Sciatic Function Index (SFI) and Basso, Beattie, and Bresnahan (BBB) locomotor rating scales. The SFI quantitatively evaluates sciatic nerve (SN) function by measuring multiple parameters, such as paw length and toe spread, from animal footprints [14,15]. The BBB locomotor rating scale evaluates motor function in spinal cord injury models by scoring lower-extremity movements, such as toe and ankle joint movements, weight bearing, and gait coordination. Both methods are widely used worldwide and recognized as common methods for evaluating motor function.

Unlike standardized approaches established for small animals, there is no globally accepted method for assessing motor function in large animals. Currently, evaluation in animals such as pigs relies on gait analysis in open fields, on treadmills, or using pressure sensor mats [16,17]. Thus, establishing a universal method for assessing motor function in large animals is crucial for nerve regeneration research [17].

Recent advances in artificial intelligence (AI), particularly deep learning, have revolutionized the field of gait analysis. Deep learning, particularly convolutional neural networks (CNNs), enables markerless tracking of anatomical landmarks from video data with high accuracy. This technology has enabled researchers to extract various kinematic parameters.

In neuroscience, for example, deep learning-based gait analysis has been applied to rodent spinal cord injury models, thereby allowing researchers to successfully detect subtle motor dysfunctions [18]. Additionally, deep learning-based gait analysis provides highly accurate analysis that is comparable to that of existing gait analysis equipment [19]. Therefore, we applied the currently popular deep learning-based gait analysis technology to pigs, developed a method to evaluate the motor function of medium-to large-sized animals using video data, and aimed to build a new evaluation system for peripheral nerve regeneration research.

2. Materials and methods

2.1. Animals

Four male micromini pigs (average age: 22.6 months, body weight: 20.6 kg at the time of surgery; Fuji Micra Inc., Shizuoka, Japan) were used in this study. The pigs were housed under a 12-h light/dark cycle, received feed once daily, and had free access to water. Three pigs were initially planned for analyses up to 3 months post-operatively as a peripheral nerve injury model. However, one pig died at 2 months post-operatively, and data for the three-month time point were therefore obtained from a different pig. All animal experiments were conducted in compliance with the protocol reviewed by the Institutional Animal Care and Use Committee, approved by the President of Niigata University (Permit Number: SA01681), and in accordance with the institutional ethical guidelines.

2.2. Kinematic recording and video creation

The walkway is shown in Fig. 1a and b. To record the pig's gait, a nonslip mat was placed along the walkway. Black lines were drawn on the mat to define a measurement area of 180 cm in length and 40 cm in width, with calibration markers placed at 16.25 cm intervals. The cages enclosed both ends of the walking path to allow the pigs to move smoothly.

Fig. 1.

Walkway setup and video recording for gait analysis. (a, b) Photograph (a) and schematic diagram (b) of the walking path used in this study. (c) Anatomical landmarks marked with black circles (Crest, Hip, Ankle, metatarsal phalangeal joint [MTP], and Toe). White arrows indicate corresponding body parts. (d) Video clip of one gait cycle: stance phase (top), swing phase (middle), and stance phase (bottom). Black arrowhead in the bottom image represents position of calibration lines placed on the nonslip mat on the floor. Scale bars: 10 cm.

Video recordings were captured using a digital camera (Tough TG-6; Olympus Inc., Tokyo, Japan; 240 frames per second; 1280 × 720 pixels). The camera was positioned 160 cm away from the walkway at a height of approximately 30 cm on a tripod, thereby ensuring that the entire 180 cm walkway was within the frame. Under these conditions, 1 pixel of the walking path corresponded to approximately 1.6 mm.

Before the first recording, the pigs were trained to walk straight along the walkway with a handler leading them. To optimize the tracking accuracy, the lower body of the pigs was shaved, and black markers were painted on key anatomical landmarks: the iliac crest (Crest), greater trochanter of the femur (Hip), lateral malleolus of the fibula (Ankle), fifth metatarsophalangeal (MTP) joint, and toe (Toe) (Fig. 1c). The knee position was not marked, as direct body surface marker tracking is prone to inaccuracies caused by skin movement artifacts [20].

The pig walking was recorded bilaterally (Fig. 1d) before surgery (Pre), 1 month after surgery (PO1M), and 3 months after surgery (PO3M). Microsoft Office Clipchamp (version 4.2.10020.0, Microsoft Corporation, Redmond, WA, USA) was used to adjust brightness and contrast of videos. The same software was used to standardize the gait direction by flipping videos of pigs walking to the right.

2.3. Creation of the common peroneal nerve injury (surgical procedure)

On the day of surgery, pigs were fasted. After Pre-anesthesia medication consisting of a mixture of medetomidine (0.05 mg/kg), midazolam (0.5 mg/kg), and butorphanol (0.2 mg/kg) was administered by intramuscular injection into the neck, general anesthesia was induced and maintained with inhaled isoflurane (induction at 4%, maintenance at 2%).

After pigs were positioned in right lateral recumbency, an intramuscular injection of antibiotics (cefazolin 20 mg/kg) was administered prophylactically, and a pulse oximeter was placed on the pig's ear during surgery to continuously monitor arterial oxygen saturation and pulse rate.

The surgical site was shaved and sterilized with povidone-iodine and 70% ethanol, and a skin incision was made from the left gluteus to the knee region. After careful dissection of the gluteal muscles, the SN, tibial nerve (TN), and common peroneal nerve (CPN) were exposed. The CPN was transected 10 mm distal to the bifurcation point where the SN divides into the TN and CPN, and a second transection was performed 20 mm distal to the first transection, thus creating a 20 mm nerve gap (Fig. 2a). The nerve gap was bridged using a sterilized silicone tube (Taiyo Kogyo Inc., Tokyo, Japan) filled with collagen (Nitta Gelatin Inc., Tokyo, Japan). A silicone tube with a length of 24 mm, inner diameter of 3 mm, and outer diameter of 5 mm was used. The proximal and distal nerve stumps were inserted 2 mm into each end of the tube and fixed with two horizontal mattress sutures using 8-0 nylon.

Fig. 2.

Pig peripheral nerve injury model and DeepLabCut workflow. (a) Surgical site showing a 20 mm common peroneal nerve (CPN) defect bridged by a collagen-filled tube. White arrow indicates the placement of the collagen-filled tube, which is located 10 mm distal to the bifurcation of sciatic nerve (SN) into the CPN (white arrowhead) and tibial nerve (TN). Scale bar: 10 mm. (b) Knee position estimation on reconstructed computed tomography (CT) using the intersection point of femoral and tibial arcs (white arrows), centered on the hip and ankle joints. Scale bar: 25 mm. (c) Anatomical landmarks labeled in DeepLabCut parts were color-labeled. (d) Definition of hip, knee, and ankle angles. (e) Deep learning workflow shown using DeepLabCut: (1) Frame extraction, (2) manual labeling, (3) network training (ResNet-50), (4) network evaluation, (5) video analysis. Scale bars: 10 cm.

The dissected muscles were sutured, and the skin was closed carefully with 3-0 nylon. Post-operatively, antibiotics (cefazolin, 20 mg/kg) and carprofen (3 mg/kg) were given for 3 days to prevent infection and control pain.

2.4. Evaluation of spatiotemporal gait parameters

Spatiotemporal gait parameters were measured from the videos. The parameters included stride length, stance duration, swing duration, gait cycle duration, maximum heel height, joint angle, and foot velocity relative to the hip. In each pig, 10 gait cycles were evaluated separately for the intact and impaired sides at Pre, PO1M, and PO3M. Gait cycles during starting or stopping motions were excluded from the analysis.

The stride length was defined as the distance traveled by the MTP joint from the initial contact of one foot to the next contact of the same foot. Stance duration was defined as the time from the toe touching the ground until the toe left the ground. Swing duration was defined from the time the toe left the ground to the next time the toe touched the ground. Gait cycle duration referred to the duration of one walking cycle (stance duration + swing duration). The maximum heel height was defined as the highest vertical distance between the ground and the heel during the swing phase.

Anatomical information of the pigs collected using computed tomography (CT; Alexion/2.0 M, Canon Medical Systems Corporation, Tochigi, Japan) before or after the gait video recording was used to support the gait parameter setting. Three-dimensional data was reconstructed using an image workstation (AZE Virtual Place; Canon Medical Systems Corporation). To minimize individual and growth-related differences, the stride length and maximum heel height were normalized by multiplying the raw values by the ratio of the mean tibial length to the tibial length of the individual animal. Stride length and maximum heel height were normalized to tibial length. Knee joint position was estimated by identifying the intersection of the arcs drawn from the greater trochanter and lateral ankle [21,22]. The nearly equal lengths of the femur and fibula also supported the identification of the knee position (Fig. 2b–d). All measurements, except for joint angles and relative foot velocity with respect to the hip joint, were obtained from videos using the Kinovea software (version 0.9.5; Joan Charmant & Contributors, Bordeaux, France).

Joint angles and foot velocity relative to the hip were measured using DeepLabCut (DLC, version 2.3.5), which is an open-source deep learning-based software for animal pose estimation [23].

2.5. Dataset creation and neural network training of DLC

Deep learning analysis (Fig. 2e) was performed using an NVIDIA RTX 4070 GPU (2023). A dataset was created by extracting a total of 1800 frames (20 frames per video) from each video. Anatomical landmarks (Crest, Hip, Ankle, MTP, Toe) corresponding to the black markers and knee locations were manually labeled on the extracted frames. For knee location, direct marker tracking is considered unreliable because of potential skin movement artifacts [20]. To improve the estimation of the knee joint, a CT scan of the pig hind limb was performed immediately before or after the gait video recording, and three-dimensional reconstruction was used to predict the knee position by identifying the intersection of the arcs drawn from the greater trochanter and lateral ankle [21,22]. The nearly equal lengths of the femur and fibula also supported the identification of the knee position. The labeled images were utilized to train a ResNet-50 CNN Pre-trained by ImageNet. Transfer learning was employed to enable the network to predict landmark positions in unlabeled frames based on the labeled dataset. Following training, the trajectory of each body part was generated as coordinate data on the x-y plane. After training, the accuracy of the neural network was evaluated by calculating the training and test errors, which were the average differences between the labeled and predicted points.

2.6. Evaluation of the neural network's accuracy

In the deep learning process, the number of iterations and training images required for proper learning were investigated. First, to optimize the training conditions, the correlation between the iterations and the loss of deep learning was analyzed. To determine the optimal conditions, a training dataset (total of 1800 frame images) was split into three groups: split 1 (10%; 180 images), split 2 (50%; 900 images), and split 3 (95%; 1710 images). The relationship between the iterations and loss was examined for each split. Under conditions of optimized iteration count where the loss converged, we evaluated the correlation between the number of training images (18, 36, 180, 900, and 1710 images) and the prediction accuracy.

2.7. Evaluation of human labeling variability

The reliability of the created neural network was assessed by comparing the network's test error with the human labeling variability. An experienced labeler labeled the same set of 200 frames twice with a substantial time interval between the first and second sessions. The human labeling variability, i.e., root mean square error values between the two labels for each anatomical body point were compared with test errors by the trained network.

2.8. Joint angle analysis

For each frame of the gait cycle, each joint angle was calculated from the x-y coordinates of each body point analyzed using DLC. The Hip angle was obtained from the three-point x-y coordinates of the Crest, Hip, and Knee, centered at the hip. Similarly, the Knee angle was obtained from the three-point x-y coordinates of the Hip, Knee, and Ankle centered at the Knee, and Ankle angle from the Knee, Ankle, and MTP centered at the ankle.

Since the gait cycle duration was different for each gait, MATLAB software (MATLAB R2020b, The MathWorks, Inc., Natick, MA, USA) was used to standardize the duration of each gait cycle using linear interpolation, such that the stance phase was 60% and the swing phase was 40%. After time normalization, the joint angles were resampled based on the standardized gait cycles. For each of the three pigs, 10 gait cycles were analyzed to calculate the mean, maximum, and minimum values. To reduce noise in the joint trajectories over the gait cycle, a 10-point moving average was applied because the raw trajectories appeared coarse. For each set of ten gait cycles, the mean, maximum, and minimum joint angles were extracted separately for the stance and swing phases. The mean, maximum, and minimum values for the ten gait cycles were calculated and graphed in relation to the gait cycle and each joint angle. The mean, maximum, and minimum standard deviations for each joint angle were plotted in the same manner.

2.9. Velocity of the foot

Velocity analyses were performed using Python (version 3.12.4; Anaconda distribution) with Pandas (version 2.2.2), NumPy (version 1.26.4), and SciPy (version 1.13.1). The velocity was calculated from the positions relative to the hip position obtained using DLC. First, the vertical length and time were normalized and standardized because they differed in each gait cycle. The vertical positions of the Ankle, MTP, and Toe were converted to their relative positions with respect to the hip position for each frame. Then, the relative positions were normalized to [−1, 1] for each gait cycle. Each gait cycle was normalized to 100% of the gait cycle duration, with 60% assigned to the stance phase (toes in contact with the ground) and 40% assigned to the swing phase (toes off the ground). The pre-swing phase to the initial swing phase was defined as 50–70% of the gait cycle, and the terminal swing phase was at 85–95%. The velocity was determined by calculating the difference in the relative positions of each point between consecutive frames. As the raw velocity data obtained by the above process contained noise, they were averaged using 10 data points in the temporal direction for data smoothing, followed by the normalization using the maximum absolute value for each gait cycle.

2.10. Statistical analysis

The Wilcoxon signed-rank test was used to compare stride length, stance duration, swing duration, gait cycle duration, maximum heel height, joint angles, and foot velocities between time points (Pre vs. PO1M, Pre vs. PO3M, and PO1M vs. PO3M) for both the intact and impaired sides. Statistical significance was set at p < 0.05. All the statistical analyses were performed using SPSS version 30.0.0 (IBM Corp., Armonk, NY, USA).

3. Results

3.1. Establishment of the pig peripheral nerve injury model

To develop a large animal model of peripheral nerve injury for gait analysis, a defect was created in the left CPN of pigs. The CPN was selected as the site of peripheral nerve injury because it was expected to exhibit drop foot, a characteristic of CPN paralysis, without significant impairment of gait function.

3.2. Validation of the machine-learning network performance for trajectory analysis of body landmarks

To understand the changes in joint angles (Hip, Knee, and Ankle angles) and foot trajectories caused by the CPN injury, key anatomical landmarks were labeled and analyzed using DLC. To explore the conditions under which the machine-learning network used in this study can achieve sufficient predictive accuracy, we explored the various conditions under which the loss was minimized.

The results of the relationship between the number of iterations and loss for each split are listed in Table 1. Under each condition (splits 1, 2, and 3), the loss decreased consistently as the number of training iterations increased and plateaued at approximately 0.0013 after approximately 500,000 iterations (Fig. 3a).

Table 1.

Numerical values of correlation between training iterations and loss in dataset using images of 10% (Split 1), 50% (Split 2), and 95% (split 3) of total frame.

Iteration	Split 1	Split 2	Split 3
1000	0.020 ± 1.4E-04	0.021 ± 2.1E-04	0.021 ± 3.9E-04
5000	0.0057 ± 2.3E-04	0.0058 ± 2.0E-04	0.0061 ± 2.4E-04
10,000	0.0041 ± 1.2E-04	0.0043 ± 1.7E-04	0.0046 ± 6.6E-05
50,000	0.0019 ± 6.4E-05	0.0021 ± 3.5E-05	0.0022 ± 1.5E-05
100,000	0.0016 ± 2.7E-05	0.0018 ± 2.5E-05	0.0019 ± 1.7E-05
200,000	0.0014 ± 1.5E-05	0.0016 ± 5.8E-05	0.0016 ± 3.2E-05
500,000	0.00092 ± 6.9E-04	0.0015 ± 1.2E-05	0.0015 ± 2.7E-05
700,000	0.0011 ± 5.77E-06	0.0013 ± 1.0E-05	0.0013 ± 2.1E-05
1,030,000	0.0011 ± 5.77E-06	0.0012 ± 2.7E-05	0.0013 ± 5.8E-06

Fig. 3.

Neural network accuracy and human labeling variability. (a) Learning curve showing correlation between loss and iterations in datasets using 10% (Split 1), 50% (Split 2), and 95% (split 3) of the total dataset. (b) Relationship between training data size (18, 36, 180, 900, and 1710 images) and model errors (test errors and training errors). (c) Spatial distribution of human labeling errors in x-y coordinates for each anatomical landmark. (d) Frequency distribution of human labeling errors for each anatomical landmark. (e) Comparison of human variability and trained neural network test error (2.56 pixels, dashed line) for each anatomical landmark and overall average (all).

Next, we examined the relationship between the number of training images and the training and test errors. The correlation between the number of training images and errors is presented in Table 2. Under the conditions analyzed in this study, the training errors remained between 2 and 2.5 pixels (Fig. 3b). However, the test errors decreased as the number of training images increased.

Table 2.

Numerical values of correlation between number of training images and errors.

Number of training images	18	36	180	900	1710
Training error (Pixels)	2.17	2.16	2.26	2.43	2.31
Test error (Pixels)	6.42	3.3	2.84	2.64	2.43

3.3. Errors of human labeling (human variability)

When the same labeler marked each frame twice, the displacement direction on the x-y coordinate plane between the first and second labels was generally identical for all body parts (Fig. 3c). The number of pixel discrepancies during the initial and second labeling for each joint and all body parts is shown in Table 3. In the evaluation of each body position, except for the knee, the magnitude of the pixel displacement rarely exceeded 3 pixels. The displacements at the Crest and Hip were mostly within 2 pixels. However, displacements exceeding 3 pixels were frequently observed at the knee (Fig. 3d). The mean and standard errors of the pixel count error between the initial and second labeling for each joint and all body parts are quantitatively demonstrated in Table 4. As the average discrepancy in labeling across all body parts, human labeling variability was 1.56 ± 0.24 pixels. The largest error was 2.88 ± 0.24 pixels for the unmarked Knee joint position calculated without markings (Fig. 3e). After sufficient deep learning processing, the test error was estimated to be approximately 2.5 pixels, and the difference between the test error and human labeling error was estimated to be approximately 1 pixel.

Table 3.

Numerical values of frequency distribution of human labeling errors for each anatomical landmark.

N (Pixels)	Crest	Hip	Knee	Ankle	MTP	Toe
0 ≤ N < 1	107	114	19	90	90	70
1 ≤ N < 2	73	65	31	60	61	48
2 ≤ N < 3	20	20	72	44	45	64
3 ≤ N < 4	0	1	33	4	3	11
4 ≤ N < 5	0	0	27	1	1	6
5 < N	0	0	18	1	0	1

Table 4.

Numerical values of human variability in body part labeling.

Body parts	Crest	Hip	Knee	Ankle	MTP	Toe	All
Difference (pixels)	1.09	1.04	2.88	1.33	1.33	1.72	1.56
Standard error	0.28	0.21	0.24	0.40	0.17	0.17	0.24

Based on these results, the test error for the dataset used to analyze the joint angles and foot trajectory was 2.56 pixels. This was obtained under training conditions with approximately 1000 labeled frames and 1,030,000 iterations, during which the cross-entropy loss plateaued at approximately 0.0013.

3.4. Quantitative analysis of gait parameters

Stride length, stance duration, swing duration, gait cycle duration, and maximum heel height were quantitatively analyzed on both the intact and impaired sides before and after surgery to assess the changes in motor function due to CPN injury (Fig. 4). On the intact side, stance duration was significantly longer at PO1M and PO3M than that at Pre (Pre vs. PO1M: p = 0.0068, Pre vs. PO3M: p = 0.0036), and maximum heel height was significantly greater at PO1M and PO3M than that at Pre (Pre vs. PO1M: p = 0.047, Pre vs. PO3M: p = 0.000894). There were no changes in the stride length, swing phase, or gait cycle duration. On the impaired side, stride length decreased significantly at PO1M and PO3M compared to that at Pre (Pre vs. PO1M: p = 0.0068, Pre vs. PO3M: p = 0.0036), and maximum heel height increased significantly at PO1M and PO3M than at Pre (Pre vs. PO1M: p = 0.0000020, Pre vs. PO3M: p = 0.0000021). In contrast, no significant changes were observed in stance duration and swing duration despite nerve injury.

Fig. 4.

Comparison of pre- and postoperative gait parameters. The blue bars represent Pre (before surgery), the red bars represent PO1M (1 month after surgery), and the purple bars represent PO3M (3 months after surgery). Left and right columns show intact and impaired sides, respectively. Data are presented as mean ± standard deviation (SD). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

3.5. Quantitative evaluation of joint angle changes

To confirm the presence of paralysis related to CPN injury, changes in the joint angles during one gait cycle, including the stance and swing phases, were evaluated on the intact and impaired sides before and after surgery (Fig. 5 and Table 5). On the intact side, the mean hip angle during the stance phase decreased significantly (p = 0.0021) and the minimum hip angle during the swing phase decreased significantly (p = 0.0032) at PO1M. The maximum knee angle during the swing phase decreased significantly (p = 0.0030), whereas the minimum ankle angle during the swing phase increased significantly (p = 0.018). On the impaired side, the mean, maximum, and minimum ankle angles throughout one gait cycle increased significantly at PO1M compared with that at Pre (p < 0.001) (Fig. 5). At PO3M, the mean and minimum ankle angles during the stance and swing phases remained significantly higher compared with that at Pre. However, compared with PO1M, ankle angles at PO3M decreased significantly.

Fig. 5.

Changes in joint angles of the intact and impaired sides over a single gait cycle. Normalized joint angle profiles of the hip, knee, ankle in one gait cycle were compared among Pre, PO1M, and PO3M for each side. The vertical axis represents the joint angle (°), and the horizontal axis represents the gait cycle (%). The dashed line at 60% of the gait cycle marks the transition between the stance phase (0–60%) and swing phase (60–100%). The thick solid lines represent the values of all individuals standardized to 100 points for each gait cycle, with error bars representing the SD at each point. The black arrow indicates significant increase in the ankle angle on the post-operative impaired side throughout the swing phase.

Table 5.

Descriptive statistics for joint angle changes.

Joint	Phase		Intact			Impaired
			Pre	PO1M	PO3M	Pre	PO1M	PO3M
Hip	Stance	Mean	102.8 ± 5.8	99.4 ± 8.3∗∗	111.0 ± 3.7†††###	102.8 ± 6.9	101.8 ± 5.6	104.1 ± 3.5
		Max	127.1 ± 5.0	124.7 ± 7.3	126.3 ± 5.5	124.2 ± 7.3	122.2 ± 4.2	125.5 ± 6.5#
		Min	82.1 ± 6.6	78.0 ± 7.4∗∗∗	92.0 ± 6.9†††###	85.5 ± 5.6	84.4 ± 7.5	88.6 ± 3.7
	Swing	Mean	95.1 ± 4.6	92.8 ± 6.0	97.2 ± 6.4	96.5 ± 6.9	94.0 ± 4.3	99.5 ± 4.5∗#
		Max	122.9 ± 5.6	120.6 ± 7.8	123.6 ± 6.3	122.1 ± 5.6	120.4 ± 3.7	123.4 ± 5.3#
		Min	79.3 ± 5.0	76.0 ± 5.2∗∗∗	82.3 ± 6.1†###	81.9 ± 6.3	79.2 ± 4.8	84.3 ± 3.0†###
Knee	Stance	Mean	116.4 ± 6.8	115.1 ± 8.2	108.1 ± 6.1	116.7 ± 7.6	118.9 ± 8.1	99.5 ± 2.37††###
		Max	127.5 ± 6.1	126.5 ± 6.2	126.6 ± 5.3	129.7 ± 7.4	130.2 ± 7.6	123.4 ± 2.44†††###
		Min	107.8 ± 7.8	106.7 ± 9.9	89.2 ± 7.1†#	107.6 ± 6.7	110.4 ± 8.6∗	84.3 ± 0.64††#
	Swing	Mean	105.6 ± 5.6	105.0 ± 6.1	105.2 ± 6.1†#	107.3 ± 5.5	107.7 ± 7.1	105.2 ± 3.2##
		Max	126.7 ± 4.6	124.2 ± 5.3∗∗	125.9 ± 5.3##	128.1 ± 7.4	129.5 ± 7.6∗	126.0 ± 3.1#
		Min	85.2 ± 2.0	85.8 ± 7.9	83.7 ± 7.1††#	88.2 ± 6.1	85.1 ± 7.0∗	83.7 ± 6.0
Ankle	Stance	Mean	132.7 ± 3.4	134.0 ± 3.1	133.8 ± 2.4	131.7 ± 3.0	139.8 ± 3.8∗∗∗	136.3 ± 2.1†††###
		Max	145.2 ± 4.1	147.3 ± 3.0	144.3 ± 3.5##	143.7 ± 3.9	150.4 ± 2.7∗∗∗	145.3 ± 1.9###
		Min	123.9 ± 4.2	126.4 ± 3.9∗∗	127.1 ± 3.2††	124.2 ± 3.3	132.9 ± 5.2∗∗∗	130.0 ± 3.3†††#
	Swing	Mean	117.3 ± 3.2	118.9 ± 5.1∗∗	118.6 ± 4.6	117.4 ± 4.8	130.8 ± 4.2∗∗∗	125.8 ± 11.2†††###
		Max	143.8 ± 4.6	145.1 ± 3.6	142.5 ± 6.7	142.4 ± 5.0	149.0 ± 3.5∗∗∗	143.8 ± 2.2###
		Min	94.1 ± 4.7	97.2 ± 8.2∗	98.1 ± 4.9††	96.0 ± 7.2	113.5 ± 5.7∗∗∗	107.2 ± 3.3†††###

Values are expressed as mean ± standard deviation. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001 between Pre and PO1M; †p < 0.05, ††p < 0.01, †††p < 0.001 between Pre and PO3M; #p < 0.05, ##p < 0.01, ###p < 0.001 between PO1M and PO3M.

3.6. Quantification of foot velocity relative to hip

Additional detailed analysis was conducted in terms of trajectory and velocity to analyze gait changes caused by peripheral nerve injury. Fig. 6 shows the changes in relative positions and velocity per gait cycle in the Ankle, MTP, and Toe. In the panels showing trajectories of relative positions on impaired side, PO1M showed small humped structures, which indicate vertically upward motion during the terminal swing phase (85–95% of the gait cycle), whereas Pre did not. This change indicates a motion in which the foot on the impaired side is raised before landing so that the toe does not touch the ground directly. In the panels showing changes in velocities on impaired side the upward humped structures were more prominently identified in the Ankle, MTP, and Toe (arrows). The prominence of the structure allowed for comparison of the amplitude of this motion. It was largest at the toe, followed by the MTP, and the ankle. The amplitude of this motion increased toward the toes, thereby indicating that it was a swinging motion.

Fig. 6.

Foot position and velocity changes relative to the Hip on the intact and impaired sides. The y-position and y-velocity of the Ankle, MTP, and Toe relative to the hip during a single gait cycle were compared among Pre, PO1M, and PO3M for each side. The dashed line at 60% of the gait cycle marks the transition between the stance phase (0–60%) and swing phase (60–100%). The black arrows indicate the significant increase in y-velocity in the impaired ankle, MTP, and toe during the terminal swing phase (gait cycle: 85–95%).

Table 6 shows the velocities of Ankle, MTP, and Toe in pre- and terminal-swing phases. On the intact side, no significant differences were observed between Pre and PO1M or PO3M. On the impaired side, significant differences between Pre and PO3M were observed for all parts in the terminal swing phase, whereas no significant differences were observed in the pre-swing phase. For the terminal-swing phase, the values of Ankle and Toe at PO3M were significantly different from Pre (p = 0.010 and p = 0.047, respectively), whereas the value of MTP was not significantly different (p = 0.055). Overall, the differences at PO3M were smaller than the highly significant differences observed between Pre and PO1M (p = 0.00042 in Ankle; p = 0.000006 in MTP; p = 0.0000051 in Toe).

Table 6.

Descriptive statistics for the velocity of Ankle, MTP, and Toe, relative to hip.

Body part		Intact			Impaired
	Phase	Pre	PO1M	PO3M	Pre	PO1M	PO3M
Ankle	Pre-swing	−0.39 ± 0.26	−0.41 ± 0.29	−0.44 ± 0.23	−0.40 ± 0.32	−0.43 ± 0.18	−0.50 ± 0.14
	Terminal-swing	0.18 ± 0.25	0.12 ± 0.20	0.16 ± 0.15	0.20 ± 0.19	0.04 ± 0.15∗∗∗	0.09 ± 0.19†
MTP	Pre-swing	−0.48 ± 0.32	−0.45 ± 0.32	−0.54 ± 0.27	−0.33 ± 0.35	−0.45 ± 0.26	−0.60 ± 0.18†††
	Terminal-swing	0.37 ± 0.30	0.30 ± 0.32	0.27 ± 0.21	0.37 ± 0.23	0.04 ± 0.25∗∗∗	0.29 ± 0.29###
Toe	Pre-swing	−0.60 ± 0.25	−0.56 ± 0.31	−0.73 ± 0.19#	−0.48 ± 0.34	−0.57 ± 0.22	−0.66 ± 0.21
	Terminal-swing	0.37 ± 0.26	0.24 ± 0.39	0.26 ± 0.23	0.32 ± 0.24	−0.08 ± 0.26∗∗∗	0.17 ± 0.31†###

Values are expressed as mean ± standard deviation. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001 between Pre and PO1M; †p < 0.05, ††p < 0.01, †††p < 0.001 between Pre and PO3M; #p < 0.05, ##p < 0.01, ###p < 0.001 between PO1M and PO3M.

4. Discussion

This study is the first to demonstrate that a deep learning-based gait analysis system can objectively quantify acute motor deficits in a pig CPN injury model using video data without specialized equipment. This deep learning-based gait analysis technology was realized through medical–engineering collaboration and enables the processing of large volumes of data, reduces manual labor, and allows rapid gait assessment [24,25]. This technology also enables objective and efficient quantification of motor function in medium-to large-sized animal models of peripheral nerve paralysis, an achievement previously unresolved in biomedical research, thereby laying the foundation for clinically applicable research. The greatest advantage of gait analysis using deep learning is that it can be performed anywhere, regardless of the species and size of the animals, without the need for specialized equipment by simply setting up an animal's walking path and camera to capture an appropriate gait video.

Open-field scoring, treadmill-based analysis, and pressure sensor mats have commonly been used in previous large animal gait studies [17]. However, open-field scoring has limited objectivity [26]. Analysis using a treadmill requires expensive equipment and is difficult to acclimate to the gait [27,28]. Pressure sensor mats are expensive, and the interpretation of the data obtained is complicated. In addition, pressure sensor mats can only evaluate vertical pressure on the soles of the feet [29].

In contrast, AI-based approaches do not require specialized equipment and enable a more objective, dynamic, and inexpensive gait assessment. In this study, circular enclosures placed at both ends of the walkway guided pigs to walk smoothly in a straight line after only a few days of acclimatization, thereby avoiding the prolonged training required for treadmill walking. Another advantage is that incorporating deep learning into gait analysis makes it possible to process massive quantities of data, thus enabling the continuous evaluation of gait cycles and dynamic assessments such as velocity.

Although AI-based gait analysis is highly advantageous, the results obtained depend primarily on the accuracy of the neural networks used in deep learning. In the deep learning analysis in this study, the test error of the trained model was approximately 2.5 pixels in the 1280 × 720 resolution videos (total 921,600 pixels) (Fig. 3b). Some previous studies using 1280 × 960 resolution videos (total 1,228,800 pixels) yielded a test error of approximately 3.5 pixels [18], whereas others analyzed 1920 × 900 resolution videos (total 1,728,000 pixels), yielding a test error of approximately 5.5 pixels [24]. The error normalized to the diagonal length of the images in this study was 0.17% compared to 0.219% and 0.259% in previous studies [18,24]. These results indicate that our neural network for pig gait analysis achieved a higher accuracy than previously reported models.

We also evaluated human variability to assess the accuracy of the test error compared with the error observed when human labeling was performed. The average interpersonal variability for all landmarks was 1.56 ± 0.24 pixels (Fig. 3e). This was approximately 1 pixel smaller than the value reported in a previous study [18]. The smaller interpersonal variability may be due to the use of lower resolution videos and the analysis of participants from a greater distance than that in previous studies. The 1-pixel (approximately 1.6 mm) difference between the test error and the interpersonal variability is unlikely to significantly affect joint angles or velocities, given that the length of a pig's lower limb is approximately 200 mm; therefore, this is considered an acceptable error.

Gait analysis of CPN-injured pigs revealed a decrease in stride length on the impaired side and a prolonged stance phase on the intact side (Fig. 4). The reduction in stride length on the impaired side was likely due to impaired ankle dorsiflexion, which resulted in decreased ankle push-off function and reduced forward propulsion. In addition, motor paralysis in the impaired limb may disrupt gait symmetry and balance leading to increased weight-bearing time in the intact limb. Post-operative increase in stance duration on the intact side likely represents a compensatory adaptation to reduce weight bearing on the impaired side. The significant increase in the maximum heel height on the impaired side at PO1M (Fig. 4) may represent a compensatory strategy to prevent toe dragging caused by impaired ankle dorsiflexion owing to CPN paralysis. In humans with CPN paralysis, a similar compensatory behavior results in a characteristic steppage gait wherein the foot is lifted by excessive hip and knee flexion [30]. In addition, in the current experimental model, maximum heel height decreased from PO1M to PO3M, tending to return to the value at Pre. This trend suggests a reduction in compensatory movements related to paralysis. In the CPN-injured pigs, no apparent post-operative changes in the hip or knee joint angles were observed between Pre and PO1M (Fig. 5 and Table 5), despite the increased heel height (Fig. 4). A possible explanation is that the pigs shifted their center of gravity and tilted their trunk toward the intact side, thereby facilitating foot clearance without joint flexion. This is consistent with the prolonged stance phase on the intact side. Although the hip abduction and external rotation may have contributed to the circular gait to avoid toe drag, the single-plane videography system in this study limited our ability to assess trunk tilt and gait asymmetry.

Joint angle analysis revealed a significant increase in the ankle angle on the impaired side throughout the gait cycle at PO1M (Fig. 5 and Table 5). The increase directly reflects the “drop foot” deformity characteristic of CPN paralysis, a finding that is consistent with previous studies [[31], [32], [33], [34]]. Ankle angles at PO3M decreased significantly compared with that at PO1M, indicating recovery from CPN paralysis. Previous studies generally focused on the minimum ankle angle during the swing phase or qualitative scoring of foot strike patterns [[31], [32], [33], [34]]. In contrast, our study quantitatively evaluated the ankle angles throughout the entire gait cycle, while encompassing both the stance and swing phases. This AI-based approach enables a more comprehensive assessment of the gait abnormalities caused by CPN injuries. Velocity analysis revealed an increase in the velocity of the Ankle, MTP, and Toe toward the hip joint immediately before the foot strike on the impaired side (Fig. 6 and Table 6). The increase in velocity was considered the result of compensatory movements to prevent the dorsum of the foot from contacting the ground. In a sheep model, it has been reported that in the absence of active control, the limb often compensates by contacting dorsum of the toes with the ground and “flicking” the foot forward before plantar contact [34]. The characteristic velocity pattern observed at PO1M, suggestive of a flicking movement, resolved at PO3M. This suggests a reduction in compensatory movements associated with CPN paralysis, likely reflecting recovery from the paralysis or adaptive changes to the condition. In this analysis, the flicking movement was quantitatively identified through changes in the distal foot velocity relative to the hip joint after surgery. To the best of our knowledge, this is the first study to quantitatively characterize this behavior. Capturing changes in velocity has been made possible by advances in deep learning, which allows the precise tracking of individual body parts in each video frame, thus enabling dynamic gait analysis.

In the present experimental model, we focused on CPN injury, which causes characteristic gait disturbances; however, this deep learning-based analysis method can be applied to the evaluation of paralysis resulting from a wide range of nerve injuries. Similar assessments can be made not only for peripheral nerve injury including sciatic and tibial nerve palsies, but also for the damage of the central nervous system including spinal cord injury and cerebral infarction. This technology has the potential to contribute not only to the field of orthopedics as in this case, but also to fields such as neurosurgery, neurology, and rehabilitation medicine. Additionally, while this study quantitatively evaluated post-injury changes and recovery by focusing on walking patterns, this technology may be applied to the behavioral evaluation of various regenerative therapy outcomes.

Finally, this study was limited to a two-dimensional analysis using a single camera, which analyzed trunk tilt and joint rotation insufficiently. Installing multiple angles for three-dimensional analysis would enable a more detailed analysis. Nevertheless, the present findings suggest that this deep learning-based evaluation system is capable of quantitative and objective assessment of functional changes associated with nerve injury and may serve as a useful tool for assessing both motor impairment and recovery processes.

5. Conclusions

This is the primary report of deep learning-based gait analysis with large experimental animal, enabling the quantitative and objective evaluation of the functional recovery after peripheral nerve injury. This novel technology is potentially useful for identifying outcome of regenerative therapy for future clinical applications.

Authors’ contributions

J.N.: conception and design, collection and assembly of data and interpretation, manuscript writing; K.O.: conception and design, collection and assembly of data, manuscript writing; R.U., Y.I., T.Y.: assembly of data and interpretation, manuscript writing; R.M., S.Y.: assembly of data and interpretation, manuscript writing; Y.U. and H.M.: provided technical instructions and support for operating the DeepLabCut software; E.N., J.K., T.K., K.K., A.F., T.Y., and H.K.: review and manuscript writing; S.S.: conception and design, administrative support, manuscript writing. All authors have read and approved the final version of the manuscript.

Ethics approval

Not applicable.

Availability of data and materials

The datasets are available from the corresponding author on reasonable request.

Funding

This research was supported by the Japan Agency for Medical Research and Development (AMED) (Grant Numbers: JP25ym0126168 to S.S., K.O., and J.N.; JP21gm6510006 to S.S.), MEXT JSPS KAKENHI, Japan (Grant Number: JP24K19570 to K.O.), the JST FOREST Program, Japan (Grant Number: JPMJFR244P to S.S.), the General Insurance Association of Japan to J. N., S.S., and M.H. and by JSPS Program for Forming Japan's Peak Research Universities (J-PEAKS) Grant Number JPJS00420240016.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.