Digital anthropometry for body circumference measurements: Toward the development of universal three‐dimensional optical system analysis software

Abstract Background/Objective Digital anthropometric (DA) assessments are increasingly being administered with three‐dimensional (3D) optical devices in clinical settings that manage patients with obesity and related metabolic disorders. However, anatomic measurement sites are not standardized across manufacturers, precluding use of published reference values and pooling of data across research centers. Subjects/Methods This study aimed to develop universal 3D analysis software by applying novel programming strategies capable of producing device‐independent DA estimates that agree with conventional anthropometric (CA) measurements made at well‐defined anatomic sites. A series of technical issues related to proprietary methods of 3D geometrical reconstruction and image analysis were addressed in developing major software components. To evaluate software accuracy, comparisons were made to CA circumference measurements made with a flexible tape at eleven standard anatomic sites in up to 35 adults scanned with three different commercial 3D optical devices. Results Overall, group mean CA and DA values across the three systems were in good agreement, with ∼2 cm systematic differences; CA and DA estimates were highly correlated (all p‐values <0.01); root‐mean square errors were low (0.51–3.27 cm); and CA‐DA bias tended to be small, but significant depending on anatomic site and device. Conclusions Availability of this software, with future refinements, has the potential to facilitate clinical applications and creation of large pooled uniform anthropometric databases.


| INTRODUCTION
Body size and shape information provides valuable insights into a wide range of topics related to human obesity. [1][2][3][4][5] Anthropometric measurements, such as circumferences that define body size and shape, are inexpensive and safely acquired for evaluating the health and nutritional status of patients with overweight and obesity across the full lifespan. 2,3 The use of these estimates, applied worldwide in highly varied settings, are advocated by numerous scientific organizations and health-related associations as a means of weight trajectories, gauging the risk of developing chronic diseases, and many other topics of clinical and research interest. 5 Moreover, anthropometric measurements defining body size and shape go beyond applications in patients with overweight and obesity and are widely applied in many other clinical nutrition areas.
An important application of conventional anthropometry (CA) is to quantify and monitor somatic features, such as adiposity level and adipose tissue distribution, as part of multicenter trials and survey protocols. 1,[6][7][8] An ambitious goal would be to create large cloudbased anthropometric databases by pooling the information collected in these studies and using them for numerous clinical and investigative purposes. Although building these global databases is a laudable objective, several roadblocks now limit the practicality of this approach. First, CA measurements are often highly variable between observers. 5,9 While implementing standardized protocols and using highly trained staff to perform assessments mitigates inter-observer error, measurement bias from training and technique can still be pronounced. Additionally, even with the employment of competent technicians and staff, participant evaluations are time consuming and can effectively lead investigators to compromise the number and type of measurements to be collected as well as overall study sample size. 5 Another concern is that conventional anthropometric measurements are typically transcribed by hand, thus increasing the risk of data entry errors.
Recent advances in three-dimensional (3D) optical imaging technology are making large datasets possible by bypassing the inefficiency of developing similar resources using CA measurements.
3D scanners ranging from research-grade laser systems to small inhome devices [10][11][12][13] are increasingly popular and widely used. These scanners automatically give consumers and investigators hundreds of digital anthropometric (DA) measurements with little test-retest variation. [10][11][12][13][14] However, through their proprietary software, each manufacturer provides DA measurements unique to their own system and the anatomic site definitions of these measurements are often unclear. Thus, pooling DA data acquired from different systems is largely precluded by variable anthropometric betweenscanner estimates for the same anatomic region. Consequently, a universal definition is required for obtaining DA measurements from different scanners despite their inherent design differences.
An important consideration in establishing appropriate DA measurements is to find definitions that would best match standard CA measurements.
The process of comparing digitally and traditionally acquired body measurements is challenging due to the intrinsic nature in which each method identifies landmarks. For example, CA relies on palpation for anatomic structures to locate measurement sites, whereas DA is limited to superficial information to define measurement locations. As the relationship between standard CA measurements and scanner-specific DA measurements is not available, DA methods might not be usable for health risk prediction from published models that, for example, rely on waist circumference estimates. 10,15 This study aimed to develop "universal" 3D analysis software with a two-fold purpose: to provide clinicians a tool for gathering device-independent standardized anthropometric body dimensions for estimating body composition and heath-risks; and to initiate creation of large pooled anthropometric research-oriented datasets.
The developed universal software tool was evaluated by comparing DA measurements to corresponding CA measurements acquired at well-defined and accepted anatomic locations acquired with a flexible tape by an expert anthropometrist.

| Participants
A convenience sample of healthy adults (age ≥18 years) at or above a body mass index of 18.5 kg/m 2 were asked to arrive at the Pennington Biomedical Research Center Metabolism-Body Composition Laboratory and change into form fitting shorts and in addition, if female, a sports bra. Subjects with hair that extended below their chin were asked to wear a swim cap. Prior to enrollment, all participants provided their written and informed consent. The experimental procedures were approved by the Pennington Biomedical Institutional Review Board as part of the larger Shape Up! Adults study which is publicly listed on ClinicalTrials.gov as ID NCT03637855.

| Experimental design
The study was conducted in two phases. First, a team of engineers at Louisiana State University (Baton Rouge, LA) led by the authors stepped through the key stages of software development, solving technical problems at each stage. Following completion of the prototype software, participants underwent scanning by three commercially available 3D optical scanners as well as CA measurements collected by a highly trained staff member. The universal software was then used to analyze the images acquired by each scanner and generate estimates designed to match those of standard anthropometric definitions. These digital measurements produced by the universal software for each scanner were then compared to those acquired by CA.

| Anthropometry
Accuracy and precision of measurement methods were evaluated as part of the parent Shape Up! Adults study. 11,12

| Conventional
Conventional anthropometric measurements were made on each participant by a single highly trained staff member using a calibrated flexible tape. Circumferences of the chest, waist, hip, upper arms, thighs, calves, and ankles were measured and recorded in triplicate to the nearest 0.1 cm; results were averaged. Replicate measurements differing more than 0.5 cm were discarded and remeasured three additional times. For a description of measurement sites and associated references, see Table 1. 16,17 The coefficient of variation for the CA circumference measurements at our center range from 0.3%-0.9% for repeated measurements. 10,11,13

| Digital
Three commercial optical systems were used to obtain images of Previously published studies implementing these devices show precision estimates ranging from 0.3% to 5.0% for repeated measurements of the same 11 circumferences. 10,11,13

| Universal software development
The software program Matlab (Mathworks) was used to create a prototype of the universal software that consists of a three-step procedure: preprocessing, landmark detection, and DA calculation.
The investigators will make the Matlab version of the software freely available without restrictions upon request. A more general-use version of the software is in development.

| Preprocessing
The data for each 3D scan contains a triangular mesh similar to the one shown in Figure 1, where the 3D cloud points are represented by a list of vertices and meshing is represented by a list of "faces". To ensure comparable analysis results across devices, the developed software initially reformats and repairs the cloud points and the mesh such that the same software can be used for the 3D data obtained from different scanners. This step consists of correcting mesh alignment, adjusting poorly shaped faces and faces causing meshing defects, and reconstructing the mesh where there is missing data known as "holes". Scan Reconstruction for Anthropometry Measurements (ScReAM) software was used for reconstructing the mesh as reported in an earlier study. 17,18

| Landmark detection
After preprocessing, basic landmarks, represented by red stars in

| Calculating DA measurements
Using the landmarks as points of reference, each scan is first segmented into arm, leg, and center segments ( Figure 2). Then each SOBHIYEH ET AL. To explore the level of agreement between the universal software-derived DA measurements and their CA counterparts, the following were tested for each scanner: (1) the magnitude and significance of between-method CA-DA differences (Δs) using paired t-tests for each of the eleven circumferences, (2) the magnitude of correlations between DA and CA measurements using simple linear regression analysis, and (3) if between-method DA-CA bias was present as quantified using Bland-Altman analyses. Paired, two-sided t-tests were used to compare DA circumference estimates derived from the universal software using images obtained from each device to corresponding CA measurements made with the flexible tape.
Mean differences at p < 0.05 were considered statistically significant.
For linear regression and Bland-Altman analyses, significance was set at p < 0.05.

| RESULTS
The All circumferences were collected with the participant standing between the technician and a mirror to ensure the tape remained parallel with the floor. Details of the measurement protocols and anatomic sites can be found in references #16 and #17. The coefficient of variation for these circumferences at our center ranges from 0.3%-0.9% for repeated measurements. 10,11,13

| Algorithm evaluation
Circumference measurement means (�SD) for CA and DA along with their mean differences (Δs) are presented in Table 2. Overall, the group mean values for circumferences evaluated by CA and DA were similar with small but statistically significant differences that varied by system and anatomic site. Absolute mean CA-DA differences were about 2 cm, except for a few outliers, across all three systems and eleven anatomic sites. The mean systematic differences between CA and DA were negative for Styku and positive for Proscanner and SS20. As the mean absolute differences (∼2 cm) were similar across anatomic sites, percentage CA-DA differences were relatively small for the large chest, waist, and hip measurements (i.e., ∼2 cm for sites varying in circumference from ∼85-100 cm or ∼2%-3%) and larger for the small arm (i.e., ∼2 cm for sites varying circumference from ∼30-35 cm or 5%-7%) and ankle measurements (i.e., ∼2 cm for sites with circumferences of ∼20-25 cm or 8%-10%).
The system root-means square errors and linear regression analysis results are presented in Table 3

| DISCUSSION
The current study documents the development pathways and initial validation of software that processes images acquired on commercially available 3D optical devices and produces a standardized output of anthropometric body dimensions. These standardized anthropometric measurements can be used for a wide range of millions. 19 Our group is now developing a cloud-based system that will actualize this concept in collaboration with an international team of investigators.
A key feature of the universal software is that it uses identical anatomic landmarks in defining body circumferences independent of device or manufacturer. The selected landmarks match those used by national and international organizations in conducting health surveys such as NHANES. Even with programming efforts, small and sometimes systematic CA-DA measurement differences were observed.
These differences arise because of scanner-specific features that define the original avatar's shape that's generated by the instrument's proprietary software. One predictable cause of these effects is that the alignment between the CA estimates and automatically identified DA measurements was imperfect. The errors from these small circumference differences can propagate when taking ratios such as those of the waist and hip. With additional analyses, these kinds of measurement differences can be compensated for in future software iterations. As these CA-DA differences were variable across systems, other strategies may be needed such as performing human or phantom "calibration" studies for each new system introduced to the market. This approach is well recognized with other devices that are used as part of multicenter research or clinical programs. One vision is to include scanner-specific adjustments to data stored in cloud sites, thereby automatically making the minor compensations needed to achieve perfect alignment between commercial devices.
Another cause for CA-DA differences involves specific features of the evaluated participants or with device hardware. Of human anatomic features, errors can arise from the difficulty in finding the crotch and armpit landmarks that help the software navigate through different body parts. In the case of participants with obesity, detecting these landmarks is challenging as legs may not be clearly separated in the thigh region and the arms and trunk also can touch each other making it hard to separate the two automatically.
Advanced algorithms were introduced to address these challenging cases. 17  The armpit and crotch are also the regions that often have "holes" in the captured 3D surface. The algorithm for patching these and other regional holes was developed so that the mesh patches follow the curvature of the body and thus do not compromise body shape. Moreover, these areas are not the sites at which DA measurements are made and thus their influence on accuracy is minimal.
Intra-scan movement was consistently less pronounced when device features helped stabilize the participant. For example, more movement and image distortion were apparent in scans from the Styku system that has a turntable but no handlebars. These movement artifacts have been observed in other studies 10,11 ; however, these artifacts rarely occurred in the present study with the systems that provided handles (Proscanner) or that had minimal scan times and remained stationary (SS20). While handles produced more consistent results at distal extremities, they also obstructed light waves in such a way that over exaggerated the size of the user's wrists and forearms.
Although the results reported in this study are adequate to support the initial use of the proposed universal software and its algorithms, a larger sample of subjects is required to better analyze Another concern is that flexible tape measurements were used as the reference to evaluate the accuracy of DA measurements.
However, CA measurements contain human error and can be difficult  Table 3. CA, conventional anthropometric; DA, digital anthropometric to measure accurately for more robust or curvy body shapes. While precision estimates are inherently low due to measuring protocols in place preventing larger deviations, this does not account for measurement accuracy. High resolution laser-based scanners might serve as an alternative to potentially less accurate CA in future studies.
However, these devices tend to be costly and may not be practical for use in large scale trials carried out at multiple research centers.
This study validated the proposed universal software for three optical imaging devices (Styku, Proscanner, and SS20). In the next step, the plan is to extend the application of the developed software to scans obtained by other devices, specifically those optimized for home use. Due to compromises in hardware size and quality to make these devices affordable, scans usually contain more image artifacts.
To ensure the universal software works for all devices, the future plan is to train a Convolutional Neural Network (CNN) with the results obtained by the three evaluated devices. Once trained, the CNN can be tested for producing similar results on other devices.
In conclusion, universal 3D optical scan analysis software was developed and critically evaluated in the current study. The software takes a 3D optical scan in the form of a triangular mesh, first reformats and edits the mesh, then automatically detects landmarks such as the feet, armpits, shoulders and crotch, and lastly calculates DA measurements including chest, hip, waist, mid-upper arm, thigh, calf, and ankle circumferences. The software provides standard definitions for DA measurements that are not only manufacturer-independent, but also closely match CA measurements in practice. Differences between CA and DA were detected, and areas of software and device hardware improvement have been identified. With further software updates and refinements, clinicians who acquire data with 3D optical devices and then process the data with universal software will then be able to reference patient results against published normative values. These developments moving forward will open a path to creating large anthropometric datasets as standardized data can be collected across centers using different 3D optical devices. Such large datasets with diverse samples will create an opportunity for conducting new studies in a wide range of health and fitness topics.