Validation of laser scanner for the collection of anthropometric measurement

2.1Population sampling

This study is a proof of concept validation to test if a 3D low resolution laser scanner (Unique Solutions, Inc.) would adequately measure linear and circumferential anthropometric measurements, especially in severely obese populations. Consequently, participants composed a small-scale sample of convenience of the target population. This stratified sampling allows to determine the systematic error of the scanner over a large range of possible measurements, and larger variety of body shape, prior to determination of agreement between manual and 3D scan extracted measures in obese participants. The stratification of the participants was based on the following BMI categories: < 18.5, 18.5 to 24.0, 25 to 29.9, > 30.

2.2Data collection

All participants underwent the same measurement protocol consisting in one landmark identification station and two measurement stations. The order of stations was the following: Landmark Identification, Laser Scan acquisition and Linear Manual measurements.

2.3Laser scanner post data collection processing

Following the data collection, each laser scan was converted to a point cloud of three-dimensional data points (Unique Solutions, Inc.). Custom software was developed to identify landmarks and begin the measurement process. The custom software identified the distal and proximal landmarks of the length measurements in Table 1 and measured the distances between them. In cases where the ground was the terminal point of the measurement, the lowest data point on the foot was determined to be the end-point. Furthermore, to create the circumferences, the software used the landmarks described in Table 1 to create planes segmenting the body into individual segments.

The custom analysis software extracted measurements based on ISO and research standards such that the scanner could be validated based on compatible comparisons. Briefly, the custom software processed the point cloud data, identified anthropometric landmarks, and performed linear as well as circumferential measurements. Furthermore, a graphical user interface (GUI) was developed to allow non-technical users to interact with point cloud data.

The GUI dashboard was the main workspace for user interaction with point cloud data. It displayed the three-dimensional point cloud and allowed users to select a series of anthropometric landmarks from the plot, from which measurements were extracted. Once the user loaded the point cloud file, they then visually identified the 1-cm diameter landmark markers and clicked them in a pre-determined order. The index locations of the three-dimensional coordinates selected were automatically stored. Once the user had selected all 43 landmarks, the user inspected each of the major segments in the segment viewer each point cloud segment was saved in a separate file for measurement, if necessary. For all segments measured, the vertical (superior-inferior) axis was the Y-axis, the anterior-posterior axis was the Z-axis, and the medial-lateral axis was the X-axis. Because arm segments were abducted in the standard posture, an algorithm in the custom software calculated the abduction angle, and rotated the local long axes of these segments to be aligned with the global Y axis, orienting the arm segments vertically.

To calculate linear measurements, Euclidian distances were calculated between three-dimensional coordinates of appropriate three-dimensional landmarks. Circumferences were determined as the maximum distance around the segment. Each major segment was viewed as 1-cm transverse slices of body segments taken along each long axis and each slice was “flattened” into a two-dimensional cross-section. An algorithm identified the vector connecting the maximally displaced data points as the major axis of the slice, and rotated the slice data to align the line joining the points with the horizontal axis. The cross-section was divided horizontally into two independent sets of data which were curve fitted using an 8th order polynomial, to account for the contours of the various body surface, while reducing the risking of over-fitting the data. The complex curvature drove the decision to use an 8th order polynomial to model the upper and lower regions separately. The arc length of each polynomial was extracted and summed to approximate the circumference of that slice. The arc length of a polynomial was calculated according to the Eq. (1).

The lower and upper limits of integration in the case above were defined to be the origin of the X-Z plane, and the X-axis value farthest from the origin, respectively. In the cases where the upper and lower regions were discontinuous at the end of the curves away from the origin, a straight-line fit was calculated between the distal points, and the length was added to the total circumference. Finally, using the index positions of the identified landmark markers, the anthropometrically relevant circumferences were extracted.

2.4Statistical analyses

The data (linear and circumferential) collected using manual measurement methods and extracted using three-dimensional imaging was compared statistically to confirm agreement between the two measurement methodologies. All data from the manual and scanner was tested for normality using the Shapiro-Wilkes test using a critical value equal to 1 (α= 0.1).

4.1Image quality improvement

Observation of the scans yielded a source of error related to limitations of field of view. A hard-coded algorithm in the scanner software was designed to exclude data points at the limit of the scan volume to ensure that data points representing the walls of the scanner are not included in the participant’s scan file. Often this resulted in the truncation of hand segment data points. Furthermore, due to the placement of the scan heads, a resulting reduction of field of view was observed at the upper and lower regions of the scan volume along the vertical Y-axis. Despite some mechanical alterations to the scanner hardware, the field of view was only increased a small amount. This resulted in the removal of data points in the region of some participants’ heads and feet. While this was a known limitation of the scans after analysis, the truncated segments were included in the regressions because this evaluation serves as the measure of the performance of the scanner as it currently exists. Furthermore, without improving the quality of the scans, exclusion criteria can be designed that will prevent users over a certain height to be measured in the scanner (∼ 1.8 m).

Aside from mechanical improvements used to increase the scanner’s field of view, several sources of error in the acquisition of scan images were also identified. Radial distortion was most notable at the periphery of the scan volume. While this was not directly corrected by improving the quality of the scan itself, the regression of scanned measurements substantially improved the quality of the measurements at the periphery. Based on the early work using a phantom [9], it was clear that radial distortion was a problem in the images. This is an example of an error that is camera specific that should be corrected early within the measurement process. A systematic regression was employed to compensate for the distortion across participants, and the results were validated [9]. The scanner limitations were most notably experienced at the periphery of the scan volume. There is in fact little mention in the literature of these errors or their effect if any on the scan quality of prior studies [10]. Thus, it is not known if the error was not present in their measures or what of any processes were used to minimize their effects.

4.2Comparison of scan-extracted and manual measurements

Several research groups have employed 3D laser imaging in the collection of segment and total body volumes, and have published their work; however, notably absent from the literature is a validation of these measurements beyond mannequin measurements, and large segment volumes; in 2006, Wang published results of repeated measures on a single mannequin [10]. However, until this point, no comprehensive study has been performed that compares segment volumes acquired through manual methodologies to those computationally extracted from 3D scans on a sample of human participants, or even on a variety of body shapes. This validation is essential to determine the reliability of 3D laser scanning apparatus, as its eventual application is a clinical environment.

4.3Other sources of scan-extracted and manual testing errors

The primary error sources in the process can be separated into five major types. Image quality and radial distortion, field-of-view (FOV) limitations and manual volumetry issues have been discussed previously. In addition to those sources of errors there are motion artifacts (sway, breathing), and post-scan digitization.

4.4Stepwise analysis of statistical validation process

From the correlation relationships between measurement methods, it was apparent that one-dimensi-onal scan-extracted measurements are highly correlated to manual measurements.

Bland-Altman analysis allows for analysis of trends in errors between measurement modalities, and to determine the limits of agreement between measurement modalities. When the differences were calculated between manual and scan extracted measurements, and confidence intervals were calculated (95% CI), the confidence intervals far exceeded the allowable range as presented in the ISO. Due to an inability to improve scan quality through improved intrinsic and extrinsic calibration techniques (software licensing issues with the partner company), regression was used as a means of using scan-extracted measurements to predict the true measurements.

In addition to the ISO, ANSUR has provided a secondary standard for allowable error in human anthropometry – called “Maximum Allowable Error” (MAE) between repeated linear manual measurements. Its application was originally applied to the repeated measurements of manual anthropometry: a measure of precision, not a validation across measurement methodologies, and was extended for use in the evaluation of the two scanners employed during the CAESAR studies [1]. However, because this work comprises a validation of a technology that compares two measurement methodologies, not simply the precision of a single apparatus, this error standard is insufficient.

In addition to calculating confidence intervals, the Bland-Altman plots were used to determine trends between the magnitude of the error and the magnitude of true measurements. The strongest correlation can be seen in the neck circumference measurements (r=-0.717), demonstrating that the magnitudes of the errors in this measurement increases as true measurements increase in magnitude. The strong correlation between neck circumference difference values and magnitude may have been brought on by the inclusion of beads in the circumference measurements, which presented most strongly in this smaller circumference measurement.

A novel method of regression analysis was applied to the scan-extracted measurements, to improve the quality of predicted measurements. In place of difference calculations in the ISO documentation, the 95th percentile confidence interval was taken about the mean of the residuals of the regressions.

The regressions measurements included the manual measurements of mass and stature (frequently collected clinically) as predictive factors in the regressions, to test for a systematic effect caused by the size of the participant.

The resulting linear models incorporated the new variables almost exclusively additively to the regression equations, implying an under-prediction in the absence of the compensatory variables: as the mass and stature of the participants increased, the correction factors increased the compensatory effects in the regression by linearly increasing the magnitude of the predicted measurement (besides hip breadth, where mass was inversely proportional). The correlations between residuals and manual measurements were non-significant after manual stature and mass were included in the regressions. The implication of this is that one-dimensional metrics of patient size easily captured clinically, can substantially improve the quality of scans.

Several limitations with these regressions should be noted. Firstly, the manual measurements of stature and mass are significantly and moderately correlated (r= 0.754, p> 0.000) across this sample. While the regressions that used these predictors may describe the sample recruited in this study, issues of multicollinearity may increase the volatility of these regressions if used as correction regressions for scan-extracted measurements in the future. Further testing of these regressions on another sample of participants is advised.

4.5Limits of agreement

In the initial work, Bland-Altman confidence intervals were compared the ANSUR standards to determine whether a linear scan-extracted measurement was valid; however, ANSUR maximum allowable error (MAE) values only apply to repeated manual measurements and the objective of this research is to compare across measurement methodologies. The ISO standard of comparison quantifies measurement validity, which is a combination of measurement accuracy and precision.

Whereas the Bland-Altman formula (applied to ANSUR standards) for determining limits of agreement utilized standard deviation, the ISO standard utilized standard error to develop the limits of agreement between measurement methodologies. Both ANSUR and ISO methods were used to compare the limits of agreement of linear measurements to absolute thresholds of allowable error. The clinical threshold for allowable error developed for this study was 5.3%, against which confidence intervals were converted to relative errors for comparison. An examination of the results shows that most of the measurements were close to, or met, the criterion standards for each measurement. Again, unsurprising was the result that both stature and neck circumference grossly exceeded the ISO standard for linear measurement. One proposed source of neck circumference variability discussed prior is the inclusion of landmark markers in the circumference calculation. Again, discussed prior was the effect in several taller participants where stature was underpredicted due to the truncation of the point cloud near the apex of the head.

While prior discussion has shown that breathing has little impact on the error in torso volumes, one study has shown that tidal breathing artifacts can affect circumference measurements by as much as 2.0 cm, greater than the ISO standard for allowable variation in chest circumference [13]. This may explain the error in the linear measurements in chest depth, chest circumference and chest breadth.

5.1Improvements to the scan process

The main goal of this study was the validation of a commercial three-dimensional laser scanner. The intended use of the scanner for clinical applications, the custom nature of the software developed for this project, and the novel application of the Unique Solutions, Inc scanner lead to many improvements for the future application of the three-dimensional laser scanner.

Firstly, algorithms related to image acquisition should be modified to accommodate improved intrinsic lens parameters, to correct radial lens distortion. This will impact all measurements; however, those at the most distal locations – such as the hands and feet – would likely experience the greatest increase in measurement quality. While improved intrinsic parameters may result in somewhat improved distortion, the replacement of the cameras on the scanners may further improve the quality of scan-extracted measurements. In the measurement of smaller circumferences, such as the neck circumference, it would likely be beneficial to improve upon the design of landmark markers such that they do not interfere with the measurement of circumferential measurements. Slight improvements to the field of view of upper scan heads may further improve the total field of view, and improve stature, and other head and neck measurements. Experimentation in the algorithm used to model circumference measurements, such as optimizing the order of polynomial, or incorporating spline fits into the circumference fits to scan data, may increase the quality of scan-extracted circumferential measurements.

5.2Improvements to methodology

Retrospective consideration of the second purpose of this study, the validation of a process by which laser scanners can be validated, has given rise to several concerns that may require improvement in the future.

Statistical methodologies have been shown to improve the predictive ability of three-dimensional laser imaging for measurements related to obesity. There remain some proprietary hardware and software issues related to the scan process that have been identified and will likely require resolution before the scanner can be used for all measurements. With the limitations identified, the scanner is unlikely to become useful for hand and foot measurements in the near future; however, with minor modifications, clinically relevant measurements taken in the arm, torso and thigh regions will be the most useful measurements extracted by this scan system.