Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Feb 7.
Published in final edited form as: J Biomech. 2025 Jun 6;189:112810. doi: 10.1016/j.jbiomech.2025.112810

An accuracy assessment of SlicerAutoscoperM – software for tracking skeletal structures in multi-plane videoradiography datasets

Amy M Morton a, John D Holtgrewe a, Jillian E Beveridge a, Dajung Yoon b, Michael J Rainbow b, Cesar Lopez c, Kristin D Zhao c, Beatriz Paniagua d, Jean-Christophe Fillion-Robin d, Anthony J Lombardi d, Douglas C Moore a, Joseph J Crisco a,*
PMCID: PMC12880572  NIHMSID: NIHMS2142560  PMID: 40516372

Abstract

Tracking skeletal joint kinematics in vivo with biplane videoradiography (BVR) can rigorously address a range of important questions in musculoskeletal research. Here we report on SlicerAutoscoperM (SAM), an upgrade of the markerless tracking software (Autoscoper) and its migration into the established computing environment of 3DSlicer and addition of a comprehensive pre-processing module to provide a standardized workflow. We present the accuracy and agreement in tracking four skeletal joints by four research groups. Accuracy was assessed by comparing marker-generated and SAM kinematics for bones of the foot, knee, shoulder, and wrist. Bland-Altman analyses quantified bias (mean error) and limits of agreement (LOA). Tracking accuracy was robust for all joints. In the foot, mean error (bias) was less than 0.5° (1.8°) and 0.8 mm (3.1 mm). In the knee, mean error was less than 1.0° (1.5°) and 0.4 mm (0.8 mm). In the shoulder, mean translational error for both the humerus and scapula was less than 0.2 mm (0.7 mm). Rotational error was highest in Roll and Pitch for the humerus, 1.9° (4.8°) and 1.7° (4.6°), respectively, and Yaw was 0.3° (2.1°). The scapula rotational bias was less than 0.2° (0.7°). In the wrist, the error was less than 0.05° (1.2°) and 0.5 mm (1.2 mm). Our data demonstrate that SAM is an accurate image-based skeletal motion tracking tool. With broad adoption, SAM will promote collaboration, simplify the harmonization of methods between study sites for large multi-center research studies, lower the entry bar for early-stage investigators, and facilitate translations toward clinical use.

Keywords: Autoscoper, Kinematics, 3DSlicer, Videoradiography, Skeletal Tracking

1. Introduction

Over 250 million people in the US are affected by musculoskeletal disorders, including arthritis, trauma, osteoporosis, and spine pathology, which cause significant functional disability and treatment-related costs approaching 1 trillion dollars annually ($874B in 2015) (Watkins-Castillo, 2016). Insights into the effects of musculoskeletal disorders, disease, and injury on joint function, as well as effects of treatment and interventions such as joint arthroplasty, can be achieved by quantifying joint kinematics (Setliff and Anderst, 2024). Robustly quantifying joint kinematics has improved our understanding of pathological gait (Andriacchi et al., 2004) and upper extremity function (Murgia et al., 2010), and it has been used as the basis for novel surgical treatments (e. g. tendon transfers (Arnold et al., 2006)). These findings have impacted joint replacement implant designs (Fantozzi et al., 2004), and led to the development of state-of-the-art rehabilitation programs, especially when coupled with advanced musculoskeletal modeling (Delp and Loan, 2000; Ong et al., 2019; Remy and Thelen, 2009).

While optical motion capture and wearables can provide metrics of gross limb motion and activity levels, image-based methods are the state-of-the-art for accurate non-invasive quantification of dynamic skeletal kinematics, especially joint-level measurements, or arthrokinematics (Wang et al., 2015). The modalities most often used to measure arthrokinematics include biplane videoradiography (BVR), static and dynamic computed tomography (3DCT and 4DCT, respectively), and magnetic resonance imaging (MRI). The accuracy of image-based methods has, generally, been reported to be ≤ 3.1° of rotation and ≤ 1.2 mm of translation (Setliff and Anderst, 2024). In addition to measuring kinematics, these modalities provide data that can be used to calculate ligament elongations (Englander et al., 2020, Englander et al., 2019), joint spacing (a surrogate for cartilage thickness), and joint contact. Since the measurement of dynamic three-dimensional kinematics from sequences of biplane radiographs was first described in 2003 (Tashman and Anderst, 2003), there are now currently more than 45 custom BVR systems in use (Setliff and Anderst, 2024), with several additional groups adapting clinical dual-fluoroscopy systems. Growth of the field is reflected by the steadily increasing number of publications utilizing image-based motion tracking methods for musculoskeletal research, growing from 22 per year between 2012 and 2017 to 35 per year between 2017 and 2022 (Setliff and Anderst, 2024). 3DCT and, to some extent, 4DCT are more broadly available, as CT is integral to standard patient care.

Sophisticated software is required to reduce acquired images to arthrokinematics data. A major barrier to more widespread utilization of image-based tracking is the lack of accurate, affordable, and user-friendly software. Image-based skeletal tracking software is challenging to implement, time-consuming, and expensive to maintain. The only commercial package for model-based image tracking from biplane radiographs (DSX Suite™. C-Motion, Inc. Germantown, MD) has ceased product development and support as of December 2024 (“DSX Overview,” 2024). There is presently no “industry-standard” software or workflow template, resulting in inefficient lab-specific implementations, duplicate effort, and a lack of transparency. The most widely available software solution has been Autoscoper, a free, open-source image registration software developed at Brown University in the early 2010s (Akhbari et al., 2019a; Brainerd et al., 2010). However, Autoscoper lacked detailed documentation, support, and updates (legacy Autoscoper, version 2.7.1 Appendix A.1). Other groups have shared their custom-written software with colleagues, but troubleshooting, feature requests, and bug fixes remained out of reach, lacking an ongoing, centralized development effort.

To address the need for accurate and affordable software for image-based skeletal tracking, we have developed SlicerAutoscoperM (SAM), an extensive upgrade to the markerless rigid body tracking software Autoscoper, as a free, open-source software solution. Software development was accomplished as a collaboration between leading academic research groups and Kitware Inc., a highly regarded technology company involved in open-source software research and development and the core developers of 3DSlicer, the widely used software ecosystem for medical image visualization, processing, and analysis (Fedorov et al., 2012). SAM was built by updating and refining legacy Autoscoper and integrating it as part of a 3DSlicer extension. SAM includes a comprehensive pre-processing module intended to provide a standardized framework for research workflow. Documentation includes written and video tutorials, sample data and an expert-monitored user discourse forum (Appendix A.2).

The goal for SAM is to provide a free, extensible, open-source software solution for image-based skeletal motion tracking. The specific purpose of this manuscript is to report the accuracy of tracking using SAM, and to evaluate the agreement of kinematics calculated using SAM to kinematics previously calculated using legacy Autoscoper. Four research groups performed these assessments across four joints (foot and ankle, knee, shoulder, and wrist).

2. Methods

The accuracy and agreement of SAM were evaluated using BVR datasets previously acquired for other experiments of the foot and ankle, knee, shoulder, or wrist (Fig. 1). Separate datasets were used for the accuracy (markered) and for the agreement (markerless) analyses. The accuracy datasets included BVR image sequences of human joints (mostly cadaver) to which tantalum beads or optical markers were rigidly affixed for gold standard marker-based tracking (Knörlein et al., 2016). The bones in these datasets were additionally tracked using SAM to yield frame-by-frame rigid-body transformations post-processed into kinematic output metrics. Accuracy was defined as the difference in kinematic outputs computed from SAM data vs previously calculated kinematic outputs resolved via marker tracking. The agreement datasets included BVR image sequences from live human subjects performing dynamic tasks. As with the accuracy assessment, bones in the agreement datasets were tracked using SAM. Agreement was defined as the difference in kinematic outputs calculated from SAM data and the previously calculated kinematics using legacy Autoscoper. Each dataset was generated and analyzed by a research group that specialized in the biomechanics of the respective joints: Queen’s University, Kingston, ON (foot/ankle), Mayo Clinic, Rochester, MN (shoulder), and Brown University, Providence, RI (knee and wrist/hand). All data acquired using human subjects were IRB approved at the respective institution, and all study subjects provided informed consent.

Fig. 1.

Fig. 1.

Open in a new tab

Overall approach to BVR acquisitions (top row), paired, undistorted radiographs (blue and purple image sets) of each joint analyzed (middle row), and partial volumes (PV) processed from segmentation of the CT volume images (bottom row).

2.1. Image datasets and analysis Overview

The BVR hardware configurations used to generate the raw videoradiography data varied at each institution, but they were all conceptually similar: paired X-ray sources, image intensifiers, and video acquisition systems (Table 1). Bone tracking in SAM requires the following input files: undistorted biplane videoradiographs, 2D-to-3D projection camera model calibration files, and processed CT partial volumes (.tiff stacks) of the bones being tracked (Appendix A.3). In all cases, the videoradiographs were undistorted in XMALab (https://xromm.org/xmalab/) using parameters determined by analyzing images of precision-machined grids, and model calibration files were determined by analyzing images of precision-fabricated calibration cubes obtained in multiple orientations (Knörlein et al., 2016). CT segmentation and partial volume processing of the bone being tracked was completed in Materialise Mimics (v21–23) or Analyze Pro. CT scan resolution was < 0.5 mm in-plane and ≤ 0.625 along the z-axis, for all studies. Each partial volume, saved as a.tiff stack, included grayscale image voxels from within a bone’s segmented mask and null value (0) external to the bone. Digitally Reconstructed Radiographs (DRRs) are computed in SAM from the partial volumes using ray casting algorithms. Partial volume renderings are visible in the 3D World view (Fig. 2. A, C), and DRRs displayed in the projected image view (Fig. 2. B, D).

Table 1.

Biplane videoradiogarphy (BVR) imaging parameters and configuartions for each research group.

Group Queen’s Univ. Brown Univ. Mayo Clinic Brown Univ.
Anatomy Foot/Ankle Knee Shoulder Hand/Wrist
CT Image Resolution 0.441 × 0.441 mm in-plane 0.625 mm, z-axis 0.278 × 0.278 mm, in-plane 0.625 mm, z-axis 0.75 mm slice thickness and 0.35 mm slice 0.391 × 0.391 mm in-plane 0.625 mm, z-axis
BVR Imaging System Skeletal Observation Laboratory, Queen’s University Keck Facility, Brown University Clinical neuro-angiographic fluoroscopy system, Mayo Clinic Keck Facility, Brown University
BVR Source-to-detector Distance 140 cm ~185 cm ~110 cm ~130 cm
Interbeam angle 115 deg ~55 deg ~90 deg ~110 deg
X-ray Settings 70 kV, 100 mA. Continuous 70–91 kV, 160–250 mA. Continuous 70–110 kV, 255–370 mA (dynamically modulated) Continuous 68–76 kV, 80 mA Continuous
Frame Rate, Pulse width 250 Hz, 125 Hz, 1.25 ms 250 Hz, 2.0 ms 15 Hz, 3.2 ms 200 Hz, 5.0 ms
Open in a new tab
1)

Queen’s University, Skeletal Observation Laboratory: Varian model G-1086 x-ray tubes, 2 EMD Technologies model EPS 45–80 pulsed x-ray generators, 2 Dunlee model TH9447QXH590 image intensifiers (16″ diameter), and 2 Photron FastCam Mini high-speed digital video cameras (resolution 2000 × 2000.

2)

Brown University, Keck Facility System: Two Varian G-1086 x-ray tubes, two EMD Technologies EPS 45–80 pulsed x-ray generators, two Dunlee TH9447QXH590 image intensifiers (16″ diameter), and two Phantom v10 high-speed digital video cameras.

3)

Mayo Clinic, Clinical neuro-angiographic fluoroscopy system: Artis zee integrated clinical biplane fluoroscopy system (Siemens Healthcare). This fluoroscope system is designed for orthogonal neuro-angiographic imaging and features 30 × 40 cm flat-panel detectors.

Fig. 2.

Fig. 2.

Open in a new tab

Videoradiographic images, digital reconstructed radiographs (DRR), and partial volumes as loaded into SAM. (A) 3D World view (Partial volume (PV) Femur (white), Radiograph pairs background). (B) Projected image view. Femur and Tibia DRRs (orange), radiograph pairs (blue). (C) Following key frame and optimization in 3D World view. Femur and Tibia (white). (D) Femur and Tibia DRRs display white when overlapped with underlying radiograph.

Tracking in SAM was semi-automated. To start, each DRR was initially aligned with the underlying radiograph in selected key frames of the videoradiographs (Fig. 2. B). Initial key frame alignment was set either manually or using pre-existing registration data (Appendix A.4) The key frame intervals were different for each group, with a range of 2 to 14 frames. After the DRRs were initially aligned in the key frames, interpolation provided initial pose estimates for the intermediary frames. At every frame, particle swarm optimization with a normalized cross-correlation heuristic(Sharma and Singh, 2010) was used to maximize correspondence of the projected DRRs and both radiographs (Fig. 2 C, D). The particle swarm optimization input parameters included perturbations of ≤ 3 mm, ≤ 3 deg, ≤ 1,000 fit iterations, and a stall limit of 25 iterations.

2.2. Accuracy analysis

Experimental parameters including specimens, tracking approach, simulated movements, and number of frames for the SAM accuracy analyses varied by group (Table 2).

Table 2.

Experimental parameters for the accuracy studies by each research group.

Group Queen’s Univ. Brown Univ. Mayo Clinic Brown Univ.
Anatomy Foot/Ankle Knee Shoulder Hand/Wrist
Biological Replicates n = 1. Live human foot/ankle (M, 49 yrs) n = 1. Frozen human cadaver knee (M, 76 yrs) n = 1. Human cadaver torso/shoulder (M, 82 yrs) n = 1. Human cadaver hand/ wrist (F, 78 yrs)
Bead/Marker, Placement (number of beads per bone) Tantalum beads (0.8 mm). Tibia (5), Talus (4), Calcaneus (3) Tantalum beads (0.7 mm). Femur (8), Tibia (6) Tantalum beads (1.0 mm). Left and Right Scapula (4), Left and Right Humerus (4) Retro-reflective markers (9.5 mm) Radius (4), 3rd Metacarpal (4)
Tasks/Movements Hopping, walking, jogging, running Simulated hop landing, vertical drop Internal/external humeral rotation, scaption, simulated wheelchair propulsion Wrist radial-ulnar deviation, flexion-extension, and circumduction
Technical Replicates n = 1 per task/movement n = 7; vertical drop (3), hop (4) n = 1 trials per task/movement/side n = 1 trials per task/movement
Data Reduction
BVR image frames tracked 999 349 311 150
Undistortion and calibration software XMALab (Version 2.1.0) XMALab (Version 2.1.0) XMALab (Version 2.1.0) XMALab (Version 2.1.0)
Marker Tracking Software XMALab (Version 2.1.0) XMALab (Version 2.1.0) XMALab (Version 2.1.0) Qualisys Track Manager, Visual 3D (v6)
Markerless Tracking SAM-Slicer pre-release 5.70 SAM-Slicer pre-release 5.70 SAM-Slicer pre-release 5.70 SAM-Slicer pre-release 5.70
Bead Removal for Markerless Tracking “Remove Beads” fill algorithm in Gimp (GNU Image Manipulation Program)S. O. Laboratory, skelobslab/RemoveBeads. (Feb. 12, 2024). Python. [Online]. Available: https://github.com /skelobslab/ RemoveBeads XMA LAB 2D bead locations as input into custom-written Matlab script. https://bitbucket.org/ xromm/xromm_autoscopertools /src/master/ Beads removed in CT scans by manual selection and pixel adjustment. ImageJ https://imagej.net/ij/ N/A
Dataset References (Behling et al., 2024) (Welte et al., 2023) (Holtgrewe et al., 2024) (Mozingo et al., 2019, Mozingo et al., 2018) (Akhbari et al., 2019c)
Grant Funding Ontario Early Researcher Award, an NSERC Discovery Grant (RGPIN/04688–2015), an NSERC Postgraduate Scholarship—Doctoral and the Pedorthic Research Foundation of Canada National Institutes of Health: National Institute of Arthritis and Musculoskeletal and Skin Diseases (K99/R00-AR069004, R01-AR047910, R01-AR074973), National Institute of General Medical Sciences (P30-GM122732, P20-GM139664), and the Lucy Lippitt Endowment. National Institutes of Health (NIH) National Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS) under T32 Training Grant T32 AR56950 and Minnesota Partnership for Biotechnology and Medical Genomics (MNP IF #14.02) National Institutes of Health P20-GM104937 and a grant from the American Foundation for Surgery of the Hand (AFSH)
Open in a new tab

The foot/ankle dataset consisted of images from a single live human volunteer who had tantalum beads previously implanted into his tibia, talus and calcaneus, and was imaged as he hopped, walked, jogged, and ran. The knee, shoulder, and hand/wrist data sets were obtained by imaging cadaveric specimens as they were moved through the BVR field of view. The knee was moved through the field of view to simulate hop landing and a vertical drop. The scapulas and humerii were manipulated to simulate internal/external humeral rotation, scaption, and wheelchair propulsion. The hand/wrist were imaged during wrist flexion–extension, radial-ulnar deviation, and circumduction.

Bead tracking and rigid body transform calculations for the foot/ankle, knee, and shoulder were resolved using XMALAb (Version 2.1.0). 3D marker locations in the wrist study were acquired using optical motion capture (8 Oqus 5 + cameras, Qualisys Track Manager), and analyzed using Visual 3D (v6). The registration transform from optical marker to radiograph space was calculated from a custom-designed cross-calibration cylinder containing retroreflective and radiopaque markers, which was acquired in both systems simultaneously.

SAM tracking was performed using modules downloaded from the extension index as part of Slicer pre-release 5.7.0 (https://download.slicer.org/). To minimize potential bias in SAM tracking, the tantalum bead pixel and voxel intensities were algorithmically replaced with adjacent intensities in the foot/ankle and knee videoradiographs and partial volumes (Table 2). Similarly, bead pixels in the shoulder accuracy CT data were manually replaced with adjacent intensities using ImageJ; the bead pixels were not altered in the shoulder accuracy videoradiographs.

Following tracking, users exported a tracking (.tra) data file for each bone. Tracking data files contained the 4-by-4 rigid body transformation matrices of the bone’s pose at each videoradiograph frame in 3D space with respect to the global origin established during BVR system calibration.

Accuracy analysis was specific to each research group and joint studied, reflecting their particular workflows. For the foot/ankle data, a transformation matrix quantifying the difference between the SAM-calculated pose and the bead-tracked pose was computed for each of the three bones at each frame. The computed matrices were converted to helical axis of motion parameters rotation (phi) and translation (t). For the knee, the SAM tracking data for each bone was converted to quaternions and filtered using a low-pass 2nd-order recursive Butterworth filter with a cut-off frequency of 10 Hz. Tibiofemoral motion was expressed as the 6 degree of freedom (DOF) motion of the tibia relative to the femur using an x-y-z Euler sequence, corresponding to flexion/extension, abduction/adduction, and internal/external rotations, and inferior/superior, anterior/posterior, and medial/lateral translations. The shoulder data was analyzed similar to that of the foot/ankle. For the shoulder, transformation matrices were computed for the scapula and humerus that described the difference between the bead-based poses to the poses computed by SAM. The differences were reported as Euler angles using the z-y-x rotation sequence and x, y, z translations (mm). For the wrist/hand, radiocarpal motion was expressed as helical axis of motion parameters rotation (phi) and translation (t) quantifying the pose of the third metacarpal with respect to the radius.

2.3. Agreement analysis

All agreement kinematic data were previously generated using legacy Autoscoper. For the current analysis, the same data and skeletal structures were tracked in SAM and compared. Joint-specific experimental parameters, including the number of subjects, types and number of movements/trials tracked, and the number of frames processed are provided in Table 3.

Table 3.

Experimental parameters for the Agreement analysis by each research group.

Group Queen’s Univ. Brown Univ. Mayo Clinic Brown Univ.
Anatomy Foot/Ankle Knee Shoulder Hand/Wrist
Tasks/Movements 5 motion tasks: task subject (s) 1 task, 2 sides:
1-leg hop for distance (each L/ R hop direction)		 4 tasks:seat pressure relief (push up), propulsion on stationary passive wheelchair ergometer, shoulder scaption no load, shoulder scaption weighted (4 kg) 4 tasks: radial-ulnar deviation, flexion–extension, circumduction, and non-contact hammering motion
walking 3
run 3
hop 120 bpm 1
hop 156 bpm 1
hop self-selected pace 1
Total Processed Trials n = 9 n = 36 n = 40 n = 36
BVR image frames tracked 997 5,579 1,929 14,400
Dataset References (Welte et al., 2022) (Beveridge et al., 2024) (Mozingo et al., 2019) (Akhbari et al., 2021, Akhbari et al., 2020)
Grant Funding NSERC Discovery Grant (RGPIN/04688–2015) and the Ontario Early Researcher Award. National Institutes of Health: National Institute of Arthritis and Musculoskeletal and Skin Diseases (K99/R00-AR069004, R01-AR047910, R01-AR074973), National Institute of General Medical Sciences (P30-GM122732, P20-GM139664), and the Lucy Lippitt Endowment. National Institutes of Health (NIH) National Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS) under T32 Training Grant T32 AR56950 and Minnesota Partnership for Biotechnology and Medical Genomics (MNP IF #14.02) National Institutes of Health P30GM122732 (COBRE Bioengineering Core) and a grant from the American Foundation for Surgery of the Hand (AFSH).
Open in a new tab

For the foot/ankle data, a transformation matrix was computed that quantified the difference between the SAM-calculated pose and the legacy Autoscoper tracked pose for each of the three bones at each frame. The differences were reported as Euler angles using the z-y-x rotation sequence and x, y, z translations (mm). For the knee, tibiofemoral Euler angles and translations were computed as in the accuracy assessment. In the shoulder, transforms for each bone and each frame were expressed as Euler rotation (z-y-x) and translations from global. As in wrist accuracy, helical axis of motion phi and translation values were compared for the third metacarpal relative to the radius.

2.4. Statistical analysis

Bland-Altman analyses (Bland and Altman, 1986; Hamilton and Stamey, 2007) were used to quantify the bias and limits of agreement (LOA) between gold-standard and SAM-derived kinematics for Accuracy assessment and between legacy Autoscoper and SAM-derived kinematics for Agreement assessment. Confidence intervals (CI) of the LOA are summarized for data in the Accuracy assessment.

3. Results

3.1. Accuracy

Accuracy of skeletal tracking with SAM was robust for all joints and across all tasks. For the foot, the translation mean error (bias) was less than 0.8 mm for each of the three bones, the LOA was less than 3.1 mm and LOA CI less than 0.5 mm across all tasks. The bias for rotation tracking of the foot/ankle was less than 0.5 degrees, the LOA was less than 1.8 degrees and rotation LOA CI less than 0.3 degrees. There were no differences amongst the three bones of the foot (Fig. 3A).

Fig. 3.

Fig. 3.

Open in a new tab

Bland-Altman bias (data points) and limit of agreement (95% LOA, error bars) across 4 joints, 9 bones, and 11 tasks. Accuracy in tracking was assessed by comparing markered kinematics with markerless (SAM) kinematics for these four joint complexes.

For the knee, across all tasks, the translation bias was less than 0.4 mm, the LOA was less than 0.8 mm, and LOA CI less than 0.07 mm. Rotation bias was less than 1 degree, with a LOA less than 1.5 degrees and LOA CI less than 0.2 degrees. Neither bias, nor LOA, differed among the tasks (Fig. 3B).

For the shoulder, mean error in translation for both the humerus and scapula was less than 0.2 mm in each direction, with LOA less than 0.7 mm and LOA CI less than 0.2 mm across all tasks. For the humerus, bias in rotation was highest in Roll and Pitch at 1.9 degrees and 1.7 degrees, respectively, and Yaw was 0.3 degrees. The LOA in Roll, Pitch and Yaw for the humerus were 4.8, 4.6 and 2.1 degrees respectively and LOA CI less than 0.6 degrees. The rotation bias of the scapula was less than 0.2 degrees, a LOA less than 0.7 degrees and LOA CI less than 0.1 degrees (Fig. 3C).

In the wrist, across all tasks, the translation bias was less than 0.5 mm with LOA of less than 1.2 mm and LOA CI less than 0.2 mm. The mean rotation error was less than 0.05 degrees and LOA less than 1.2 degrees, (Fig. 3D) and LOA CI less than 0.2 degrees.

3.2. Agreement

The agreement between SAM and legacy Autoscoper was in general good to excellent. However, for some bones there were gross disagreements in some frames. For the foot, the translation and rotation bias between SAM and legacy Autoscoper were less than 1 mm and 1 degree for the calcaneus, talus, and tibia across all trials. The LOA between SAM and legacy Autoscoper were approximately 5 mm and 2 degrees but varied across bones and trials, with LOA of 4 trials over 10 mm and over 5 degrees across all trials. For the knee, the translation and rotation bias were less than 1 mm and 1 degree. The LOA in the knee had mean values across trials of approximately 4 mm and 2 degrees, but there were several positions with LOA translations greater than 20 mm. Similarly, in internal/external rotation (IE) the LOA rotation values were greater than 20 degrees. These outliers come from 51 frames and represent < 1 % of the total dataset. When these frames are excluded from the analyses, the LOAs are 0.73 mm and 0.98 degrees, respectively. For the shoulder, the translation and rotation bias were less than 1 mm and 1 degree except in Yaw of the scapula which was 2.0 degrees. The translation LOAs were notably larger in the scapula than in the humerus with mean LOA greater that 10 mm. The most notably large LOA was in Yaw for both the humerus and the scapula with some values greater than 180 degrees. For the wrist, translation and rotation mean bias values were less than 0.5 mm and 0.5 degrees. The LOA translation and rotation values were less than 1 mm and less than 2 degrees, but with a single outlier greater than 7 mm and 10 degrees.

4. Discussion

The goal of this paper was to introduce SlicerAutoscoperM, a free, open-source software program for image-based markerless skeletal motion tracking, and to provide an estimate of its accuracy and the agreement of its results to those obtained using legacy Autoscoper. To do so, we used Bland-Altman analyses to compare SAM outputs for bones in the foot/ankle, knee, shoulder, and wrist/hand to those obtained using marker-based tracking, and to those obtained using Autoscoper. Our accuracy analysis revealed that SAM tracking was robust, with biases for translation and rotation of ≤ 0.8 mm and ≤ 1.9 degrees and limits of agreement (LOA) ≤ 3.1 mm and 1.8 degrees.

To ensure SAM is a sustainable, robust, and reliable software package, we have followed software development and quality assurance processes as enforced in medical device software standards. All code developed as part of SAM has been fully disclosed and stored into a public software repository in GitHub (https://github.com/BrownBiomechanics/SlicerAutoscoperM) and is fundamentally integrated into rigorous software quality testing procedures. SAM is built and uploaded daily through the 3DSlicer Extensions Index, facilitating direct access to enhancements and bug fixes to users as they are made. SAM represents a collective effort of top research minds and software developers. The success of this collaboration is evident in the accuracy results we present.

A significant strength of this work is that accuracy was assessed across multiple joints using images captured by different research groups and BVR systems. The performance of X-ray-based bone tracking can be affected by overlying soft tissue, soft tissue-to-bone contrast, bone occlusion or overlap, bone velocity, and BVR system configuration. Accuracy was consistent despite the wide range of bone sizes and shapes, the various approaches to assessments, and 4 different research groups. Accuracy data were reduced differently by each of the four groups, which provides a robust and unbiased accuracy assessment. Despite the variation in post-processing, a consistent level of accuracy was achieved throughout. Significant differences occurred in < 1 % of frames. We postulate that these errors were caused by two mechanisms – bones with poorly defined features and gimbal lock. Long cylindrical bones have limited features when projected, thus, multiple orientations may produce similar optimization metrics. Gimbal lock, when two rotational DOF become aligned and “lock” the transformation, may have occurred as part of using Euler sequences in registration and post-tracking kinematic calculations. Between the times when the agreement and accuracy analyses were performed, the method for frame-to-frame pose estimation was changed from Euler sequences to quaternions.. By including multiple joints and research groups, these issues were identified, resolved, and implemented into current and future SAM releases.

SAM accuracy values align well with work by others. In the recent skeletal kinematics resolved with BVR scoping review, the accuracy across the hind foot and ankle, tibiofemoral, patellofemoral, hip, hand and wrist, lumbar spine, cervical spine, and shoulder were summarized from 18 studies (Setliff and Anderst, 2024). Those studies rereported an average translational accuracy of 0.5 mm (0.04–1.4 mm) and a rotational accuracy of 0.9 degrees (0.16–3.1 degrees). As the authors of the review point out, some studies reported accuracy as precision, some as root mean squared error (RMS), and some studies reported accuracy of tracking an individual bone while others reported accuracy of joint kinematics, making comparison challenging. We put forth that Bland-Altman analysis, with bias and limits of agreement, are a more rigorous analysis of tracking errors and could be considered the universal method of accuracy reporting for the field moving forward.

We also considered agreement assessment to be valuable as the results would confirm that modernization of the SAM environment preserved core Autoscoper functionality. Our agreement analysis revealed biases for translation and rotation that were generally on the order of ~ 1.0 mm and ~ 1.0 degrees and limits of agreement generally on the order of ~ 3.1 mm and 2 degrees. Agreement differences may reflect the tracking improvements SAM provides, as further evidenced by the strong accuracy results. In some cases, it may reflect different user workflows. As part of the agreement analysis, we also compared computational speed where we found that speed of SAM was substantially improved – 2.8 times faster than legacy Autoscoper (Appendix A5).

While we assessed SAM agreement and accuracy across four different joints and four different research groups, limitations in the scope of our assessment remain. Other joints that have been previously tracked with BVR, such as the spine (Aiyangar et al., 2014; Anderst et al., 2013; Como et al., 2024), and hip (Johnson et al., 2024) remain to be assessed with SAM. Implant components, such as in wrist (Akhbari et al., 2019b) or hip arthroplasties (D’Isidoro et al., 2023) remain to be assessed with SAM. We also did not assess all kinematic parameters for every joint and task, rather we focused on the parameters deemed most appropriate by each research group. The raw transform outputs were not directly compared among joints or institutions. Rather, we reported the results specific to each laboratory’s workflows because we believed this better reflects how accuracy is considered in the real-world application. Thus, the accuracies reported here include variations among joints due to shape and bone density characteristics, as well as among institutional workflows.

In our shoulder accuracy data, the pixel intensities of the implanted tantalum beads were removed from the CT scans but they were not removed from videoradiograph images. Accordingly, the bead intensity data in the vidoradiographs did not have any corresponding bead pixel data in the DRRs, and thus did not carry any additionally improved matching beyond what would be achievable in bead-erased images. The highest values of rotational error in the shoulder data were recorded for the humerus. Optimizing long bones with limited distinguishing features is a known hardship in 3D pose tracking. Despite this difficulty and relative increase in diminished accuracy, we still report values that should allow studies to detect clinically meaningful differences.

Future work is focused on expanding SAM tracking capabilities to calculate kinematics from sequential 3D CT scans (3DCT) and 4D CT image sequences (4DCT). CT is integral to standard patient care and has been used to study the knee (Forsberg et al., 2016; Rosa et al., 2019), hand and wrist (Choi et al., 2013; Garcia-Elias et al., 2014; Kakar et al., 2016; Shores et al., 2013; Tay et al., 2007; Wang et al., 2018; Zhao et al., 2015), and foot and ankle (Burssens et al., 2016; Collan et al., 2013). Additionally, we intend to improve the tracking process by integrating collision detection into the optimization process; real-time algorithms will calculate points of collision and intersection metrics and use that information to drive the tracking by penalizing bone positions that are in collision.

Our ultimate goal in developing SAM is to build an international user-base of collaborators and contributors to foster innovation and inquiry in musculoskeletal research. Our efforts in developing SAM aim to eliminate a significant barrier to progress in the field of image-based kinematic analysis by increasing access to validated standardized tools. With broad adoption, SAM will promote collaboration, simplify the harmonization of methods between study sites for large multi-center research studies, lower the entry bar for early-stage investigators, and facilitate translation toward clinical use.

Acknowledgements

Research reported in this publication was supported by the National Institute of Arthritis and Musculoskeletal and Skin Diseases of the National Institutes of Health under Award Number AR059185 and AR078924. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

The authors thank Joseph D. Mozingo for original processing and analysis of shoulder data.

Appendices.

A.1. Legacy markerless tracking software, Autoscoper 2.7.1 Last updated October 2019. Binaries built in CUDA backend configuration. A GitHub release was created a posteriori https://github.com/BrownBiomechanics/Autoscoper/releases/tag/autoscoper-v2.7.1.

A.2. https://autoscoperm.slicer.org/.

https://autoscoper.readthedocs.io/en/latest/index.html.

https://discourse.slicer.org/c/community/slicerautoscoperm/30.

A.3. Required input files had been previously generated with all file paths stored in a configuration file as part of the workflow for other, purpose-designed experiments. Previously configured filter parameters (Sober, Contrast, Gaussian) were saved into.vie files and applied in SAM and in agreement data, replicas of those used in paired trials processed in legacy Autoscoper.

A.4. Frame seeding in the foot/ankle utilized randomly perturbed bead tracking data: each bead-tracked pose was rotated 8 degrees about a 3-space vector diagonal to the anatomical coordinate axes of the bone.

A.5. Computation Performance. A computational performance analysis was conducted using a representative BVR wrist dataset, tracking the third metacarpal over 14 frames of a radial-ulnar deviation motion trial. Optimizations were run on a Dell workstation (i9-9900 K CPU @3.6 GHz, 64.0 GB RAM, Nvidia RTX 2080). Optimizations in legacy Autoscoper and SAM were performed using a CUDA back-end configuration and time to convergence was documented. The time to tracking convergence of all 14 frames of the third metacarpal during radial deviation using legacy Autoscoper was 3 mins, 19 s; using SAM, the same dataset converged in 1 min, 11 s.

A.6. SlicerAutoscoperM Extension Expansion and Terminology. Software is an agile product: active, changing and evolving. During completion of this body of work and upon initial submission of this publication, the 3DSlicer Extension: SlicerAutoscoperM solely included an evolved Autoscoper product and Pre-Processing module, and thus the SAM acronym has been used interchangeably in this work to refer to that specific module. In April 2025, an additional module was added to the SlicerAutoscoperM extension, and updated terminology was published in the associated readthedocs online documentation website and SAM homepage (A.2.).

Footnotes

CRediT authorship contribution statement

Amy M. Morton: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Resources, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. John Holtgrewe: Formal analysis, Data curation. Jillian E. Beveridge: Writing – review & editing, Supervision, Funding acquisition. Dajung Yoon: Formal analysis, Data curation. Michael J. Rainbow: Writing – review & editing, Supervision, Funding acquisition. Cesar Lopez: Writing – review & editing, Formal analysis, Data curation. Kristin Zhao: Supervision, Funding acquisition. Beatriz Paniagua: Writing – review & editing, Supervision, Project administration, Funding acquisition. Jean-Christophe Fillion-Robin: Supervision, Software, Project administration. Anthony Lombardi: Software. Douglas C. Moore: Writing – review & editing, Writing – original draft, Funding acquisition. Joseph J. Crisco: Writing – original draft, Visualization, Validation, Supervision, Resources, Project administration, Methodology, Funding acquisition, Formal analysis, Conceptualization.

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.

References

  1. Aiyangar AK, Zheng L, Tashman S, Anderst WJ, Zhang X, 2014. Capturing three-dimensional in vivo lumbar intervertebral joint kinematics using dynamic stereo-X-ray imaging. J. Biomech. Eng 136, 011004. 10.1115/1.4025793. [DOI] [PubMed] [Google Scholar]
  2. Akhbari B, Knörlein B, Loomis A, Howison M, 2019a. Autoscoper. [Google Scholar]
  3. Akhbari B, Morton A, Moore D, Weiss A-P-C, Wolfe SW, Crisco JJ, 2019b. Kinematic accuracy in tracking total wrist arthroplasty with biplane videoradiography using a CT-generated model. J. Biomech. Eng Doi 10 (1115/1), 4042769. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Akhbari B, Morton AM, Moore DC, Crisco JJ, 2021. Biplanar videoradiography to study the wrist and distal Radioulnar joints. J. vis. Exp. Jove 10.3791/62102. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Akhbari B, Morton AM, Moore DC, Weiss A-P-C, Wolfe SW, Crisco JJ, 2019c. Accuracy of biplane videoradiography for quantifying dynamic wrist kinematics. J. Biomech 92, 120–125. 10.1016/j.jbiomech.2019.05.040. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Akhbari B, Morton AM, Shah KN, Molino J, Moore DC, Weiss A-P-C, Wolfe SW, Crisco JJ, 2020. Proximal-distal shift of the center of rotation in a total wrist arthroplasty is more than twice of the healthy wrist. J. Orthop. Res. off. Publ. Orthop. Res. Soc 38, 1575–1586. 10.1002/jor.24717. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Anderst W, Donaldson W, Lee J, Kang J, 2013. Cervical disc deformation during flexion-extension in asymptomatic controls and single-level arthrodesis patients. J. Orthop. Res. off. Publ. Orthop. Res. Soc 31, 1881–1889. 10.1002/jor.22437. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Andriacchi TP, Mundermann A, Smith RL, Alexander EJ, Dyrby CO, Koo S, 2004. A framework for the in vivo pathomechanics of osteoarthritis at the knee. Ann Biomed Engin 33, 447–457. [DOI] [PubMed] [Google Scholar]
  9. Arnold AS, Liu MQ, Schwartz MH, Ounpuu S, Dias LS, Delp SL, 2006. Do the hamstrings operate at increased muscle-tendon lengths and velocities after surgical lengthening? J. Biomech 39, 1498–1506. 10.1016/j.jbiomech.2005.03.026. [DOI] [PubMed] [Google Scholar]
  10. Behling A-V, Welte L, Kelly L, Rainbow MJ, 2024. Human in vivo midtarsal and subtalar joint kinematics during walking, running and hopping. J. R. Soc. Interface 21, 20240074. 10.1098/rsif.2024.0074. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Beveridge JE, Hague M, Parola LR, Costa MQ, Janine Molino, Fleming BC, 2024. Static and Dynamic Constraint in ACL-Reconstructed Patients at 10–15 Year Follow-Up, in: Proceedings of the Orthopaedic Research Society Annual Meeting. Long Beach, CA. [Google Scholar]
  12. Bland JM, Altman DG, 1986. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet Lond. Engl 1, 307–310. [PubMed] [Google Scholar]
  13. Brainerd EL, Baier DB, Gatesy SM, Hedrick TL, Metzger KA, Gilbert SL, Crisco JJ, 2010. X-ray reconstruction of moving morphology (XROMM): precision, accuracy and applications in comparative biomechanics research. J. Exp. Zool. Part Ecol. Genet. Physiol 313, 262–279. 10.1002/jez.589. [DOI] [PubMed] [Google Scholar]
  14. Burssens A, Peeters J, Buedts K, Victor J, Vandeputte G, 2016. Measuring hindfoot alignment in weight bearing CT: a novel clinical relevant measurement method. Foot Ankle Surg. 22, 233–238. 10.1016/j.fas.2015.10.002. [DOI] [PubMed] [Google Scholar]
  15. Choi YS, Lee YH, Kim S, Cho HW, Song H-T, Suh J-S, 2013. Four-dimensional real-time cine images of wrist joint kinematics using dual source CT with minimal time increment scanning. Yonsei Med. J 54, 1026–1032. 10.3349/ymj.2013.54.4.1026. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Collan L, Kankare JA, Mattila K, 2013. The biomechanics of the first metatarsal bone in hallux valgus: a preliminary study utilizing a weight bearing extremity CT. Foot Ankle Surg. 19, 155–161. 10.1016/j.fas.2013.01.003. [DOI] [PubMed] [Google Scholar]
  17. Como CJ, LeVasseur CM, Oyekan A, Padmanabhan A, Makowicz N, Chen S, Donaldson WF, Lee JY, Shaw JD, Anderst WJ, 2024. Dynamic in vivo 3D atlantooccipital kinematics during multiplanar physiologic motions. J. Biomech 173, 112236. 10.1016/j.jbiomech.2024.112236. [DOI] [PubMed] [Google Scholar]
  18. Delp SL, Loan JP, 2000. A computational framework for simulating and analyzing human and animal movement. IEEE Comput. Sci. Eng 2, 46–55. [Google Scholar]
  19. D’Isidoro F, Brockmann C, Friesenbichler B, Zumbrunn T, Leunig M, Ferguson SJ, 2023. Moving fluoroscopy-based analysis of THA kinematics during unrestricted activities of daily living. Front. Bioeng. Biotechnol 11, 1095845. 10.3389/fbioe.2023.1095845. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. DSX Overview [WWW Document], 2024.. Softw. Prod. Doc URL https://wiki.has-motion.com/DSX_Overview (accessed 7.18.24).
  21. Englander ZA, Baldwin EL, Smith WAR, Garrett WE, Spritzer CE, DeFrate LE, 2019. In vivo anterior cruciate ligament deformation during a single-legged jump measured by magnetic resonance imaging and high-speed biplanar radiography. Am. J. Sports Med 47, 3166–3172. 10.1177/0363546519876074. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Englander ZA, Garrett WE, Spritzer CE, DeFrate LE, 2020. In vivo attachment site to attachment site length and strain of the ACL and its bundles during the full gait cycle measured by MRI and high-speed biplanar radiography. J. Biomech 98, 109443. 10.1016/j.jbiomech.2019.109443. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Fantozzi S, Leardini A, Banks SA, Marcacci M, Giannini S, Catani F, 2004. Dynamic in-vivo tibio-femoral and bearing motions in mobile bearing knee arthroplasty. Knee Surg. Sports Traumatol. Arthrosc. off. J. ESSKA 12, 144–151. 10.1007/s00167-003-0384-5. [DOI] [PubMed] [Google Scholar]
  24. Fedorov A, Beichel R, Kalpathy-Cramer J, Finet J, Fillion-Robin J-C, Pujol S, Bauer C, Jennings D, Fennessy F, Sonka M, Buatti J, Aylward S, Miller JV, Pieper S, Kikinis R, 2012. 3D slicer as an image computing platform for the quantitative imaging network. Magn. Reson. Imaging 30, 1323–1341. 10.1016/j.mri.2012.05.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Forsberg D, Lindblom M, Quick P, Gauffin H, 2016. Quantitative analysis of the patellofemoral motion pattern using semi-automatic processing of 4D CT data. Int. J. Comput. Assist. Radiol. Surg 11, 1731–1741. 10.1007/s11548-016-1357-8. [DOI] [PubMed] [Google Scholar]
  26. Garcia-Elias M, Serrallach XA, Serra JM, 2014. Dart-throwing motion in patients with scapholunate instability: a dynamic four-dimensional computed tomography study. J. Hand Surg. Eur 39, 346–352. 10.1177/1753193413484630. [DOI] [PubMed] [Google Scholar]
  27. Hamilton C, Stamey J, 2007. Using bland-altman to assess agreement between two medical devices–don’t forget the confidence intervals! J. Clin. Monit. Comput 21, 331–333. 10.1007/s10877-007-9092-x. [DOI] [PubMed] [Google Scholar]
  28. Holtgrewe J, Fleming BC, Beveridge JE, 2024. Accuracy and precision of model-based bone tracking for a dynamic hop landing activity, in: Orthopaedic Research Society (ORS) Annual Meeting. Long Beach, CA, p. 1064. [Google Scholar]
  29. Johnson CC, Ruh E, Frankston N, Charles S, McClincy M, Anderst W, 2024. Sex-based differences and Asymmetry in hip kinematics during unilateral extension from deep hip flexion. J. Biomech. Eng 146, 124501. 10.1115/1.4066466. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Kakar S, Breighner RE, Leng S, McCollough CH, Moran SL, Berger RA, Zhao KD, 2016. The role of dynamic (4D) CT in the detection of scapholunate ligament injury. J. Wrist Surg 5, 306–310. 10.1055/s-0035-1570463. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Knörlein BJ, Baier DB, Gatesy SM, Laurence-Chasen JD, Brainerd EL, 2016. Validation of XMALab software for marker-based XROMM. J. Exp. Biol 219, 3701–3711. 10.1242/jeb.145383. [DOI] [PubMed] [Google Scholar]
  32. Mozingo JD, Akbari Shandiz M, Marquez FM, Schueler BA, Holmes DR, McCollough CH, Zhao KD, 2018. Validation of imaging-based quantification of glenohumeral joint kinematics using an unmodified clinical biplane fluoroscopy system. J. Biomech 71, 306–312. 10.1016/j.jbiomech.2018.02.012. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Mozingo JD, Akbari-Shandiz M, Van Straaten MG, Murthy NS, Schueler BA, Holmes DR, McCollough CH, Zhao KD, 2019. Comparison of glenohumeral joint kinematics between manual wheelchair tasks and implications on the subacromial space: a biplane fluoroscopy study. J. Electromyogr. Kinesiol 102350. 10.1016/j.jelekin.2019.08.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Murgia A, Kyberd P, Barnhill T, 2010. The use of kinematic and parametric information to highlight lack of movement and compensation in the upper extremities during activities of daily living. Gait Posture 31, 300–306. [DOI] [PubMed] [Google Scholar]
  35. Ong CF, Geijtenbeek T, Hicks JL, Delp SL, 2019. Predicting gait adaptations due to ankle plantarflexor muscle weakness and contracture using physics-based musculoskeletal simulations. PLoS Comput. Biol 15, e1006993. 10.1371/journal.pcbi.1006993. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Remy CD, Thelen DG, 2009. Optimal estimation of dynamically consistent kinematics and kinetics for forward dynamic simulation of gait. J. Biomech. Eng 131, 031005. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Rosa SB, Ewen PM, Doma K, Ferrer JFL, Grant A, 2019. Dynamic evaluation of patellofemoral instability: a clinical reality or just a research field? A literature review. Orthop. Surg 11, 932–942. 10.1111/os.12549. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Setliff JC, Anderst WJ, 2024. A scoping review of human skeletal kinematics research using biplane radiography. J. Orthop. Res. off. Publ. Orthop. Res. Soc 10.1002/jor.25806. [DOI] [PubMed] [Google Scholar]
  39. Sharma A, Singh N, 2010. Object detection in image using particle swarm optimization. Int. J. Eng. Technol 2, 419–426. [Google Scholar]
  40. Shores JT, Demehri S, Chhabra A, 2013. Kinematic “4 dimensional” CT imaging in the assessment of wrist biomechanics before and after surgical repair. Eplasty 13. [PMC free article] [PubMed] [Google Scholar]
  41. Tashman S, Anderst W, 2003. In-vivo measurement of dynamic joint motion using high speed biplane radiography and CT: application to canine ACL deficiency. J. Biomech. Eng 125, 238–245. [DOI] [PubMed] [Google Scholar]
  42. Tay S-C, Primak AN, Fletcher JG, Schmidt B, Amrami KK, Berger RA, McCollough CH, 2007. Four-dimensional computed tomographic imaging in the wrist: proof of feasibility in a cadaveric model. Skeletal Radiol. 36, 1163–1169. 10.1007/s00256-007-0374-7. [DOI] [PubMed] [Google Scholar]
  43. Wang B, Roach KE, Kapron AL, Fiorentino NM, Saltzman CL, Singer M, Anderson AE, 2015. Accuracy and feasibility of high-speed dual fluoroscopy and model-based tracking to measure in vivo ankle arthrokinematics. Gait Posture 41, 888–893. 10.1016/j.gaitpost.2015.03.008. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Wang KK, Zhang X, McCombe D, Ackland DC, Ek ET, Tham SK, 2018. Quantitative analysis of in-vivo thumb carpometacarpal joint kinematics using four-dimensional computed tomography. J. Hand Surg. Eur 43, 1088–1097. 10.1177/1753193418789828. [DOI] [PubMed] [Google Scholar]
  45. Watkins-Castillo S, 2016. The Impact of Musculoskeletal Disorders on Americans – Opportunities for Action, the Executive Summary of the Burden of Musculoskeletal Diseases in the United States: Prevalence, Societal and Economic Cost (BMUS) [WWW Document]. BMUS Burd. Musculoskelet. Dis. U. S. URL http://www.boneandjointburden.org (accessed 3.1.16).
  46. Welte L, Dickinson A, Arndt A, Rainbow MJ, 2022. Biplanar videoradiography dataset for model-based pose estimation development and new user training. J. vis. Exp. Jove 10.3791/63535. [DOI] [PubMed] [Google Scholar]
  47. Welte L, Holowka NB, Kelly LA, Arndt A, Rainbow MJ, 2023. Mobility of the human foot’s medial arch helps enable upright bipedal locomotion. Front. Bioeng. Biotechnol 11, 1155439. 10.3389/fbioe.2023.1155439. [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Zhao K, Breighner R, Holmes D, Leng S, McCollough C, An K-N, 2015. A technique for quantifying wrist motion using four-dimensional computed tomography: approach and validation. J. Biomech. Eng 137. 10.1115/1.4030405. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES