PaperlayerView exact revision on arXiv

Source-linked AI summary

Repeatability Characterisation and Error Budget of a Consumer Structured-Light Scanner for 3-D Wound Geometry:A Rigid-Phantom Study

Pushkal Kumar, Aadit Aggarwal, Karlen Aleksanyan

arXiv:2608.30143v1cs.CV

TL;DR

Whether wound descriptors can detect clinically meaningful change depends on their measurement noise, which is difficult to separate from biological change in patients. The paper uses rigid phantoms to characterise repeat-scan error and finds low reliability across descriptors, while capture extent is an important addressable source of variance.

  • Problem

    The study addresses the limited comparison between clinically meaningful change in wound descriptors and the measurement noise of those descriptors.

    Source: §I · p. 1

  • Method

    The authors repeatedly free-hand scan rigid wound phantoms and decompose descriptor variance, including an analysis of capture extent while holding the analysis region fixed.

    Source: §I · pp. 1–2

  • Results

    44 of 45 descriptors fall below an intraclass correlation of 0.50, and none of 248 descriptor–pipeline combinations reaches 0.75; surface-area repeatability limit is a factor of 4.8.

    Source: §VII · p. 8

  • Takeaways & Limitations

    A 25 mm-radius crop reduces the surface-area limit to 2.13, meeting rather than clearing the clinical threshold ratio; capture extent, not sensor resolution, bounds these measurements.

    Source: §VII · p. 8

    Every repeatability, correlation, precision, region, and error-budget result characterises the Revopoint RANGE alone; the study reports no reproducibility across instruments.

    Source: §VI · p. 7

Abstract

from arXiv · show

We characterise the measurement error of a free- hand consumer structured-light scanner used to derive three- dimensional geometric wound descriptors. Rigid wound-care phantoms cannot change, so every difference between repeat scans of one site is measurement error; all repeats come from a single scanner unit, nine sites and 23 scans, so this charac- terises one instrument. The 95 percent repeatability limit for reconstructed surface area is a factor of 4.8, with a confidence interval from 3.0 to 7.0, so rescanning an unchanged site can shift the reading from a 79 percent decrease to a 377 percent increase. Forty-four of 45 descriptors fall below an intraclass correlation of 0.50, and none of 248 descriptor and pipeline combinations reaches 0.75. An error budget formed by holding the analy- sis region fixed, leaving sensor and reconstruction untouched, bounds the share of variance attributable to how much surface the operator captured at 75 percent for surface area, 91 percent for hull area and 95 percent for bounding-box diagonal; only hull volume is majority instrumental, at 48 percent, so a better sensor would buy little. No wound is delineated anywhere in the chain, so the comparison against the four-week area reduction used clinically to predict healing, a ratio near 2.1, is a lower bound on the noise an unsegmented pipeline must overcome, not a measurement of wound-area reproducibility. Standardising the analysis region cuts the limit to 2.13, meeting that ratio rather than clearing it. Statistical outlier removal imposes a measured systematic area deficit near 11 percent.

I. INTRODUCTION

Rigid phantoms make repeat-scan variation attributable to the acquisition-to-descriptor chain, enabling measurement noise to be compared with clinically meaningful change. The paper characterises descriptor noise and identifies capture extent as an addressable source of error.

Motivation

  • Ruler-based length × width overestimates area by roughly 40% relative to digital planimetry, and reported inter-rater ICC is 0.34 versus 0.77 for a smartphone planimetry application.Margin identification, rather than tracing itself, dominates error.Source: §I · p. 1
  • Prior 3-D systems report area ICCs above 0.95, but commercial cameras underestimate volume by 23 to 58% and clinical depth agreement can fall to R = 0.49.Systematic reviews describe 3-D wound technologies as under-evidenced rather than established,,.Source: §I · p. 1
  • A descriptor supports monitoring only when clinically meaningful change exceeds measurement noise, a comparison difficult to isolate in patients because biology changes.Rigid synthetic models provide an unchanging geometry, making repeat-scan spread attributable to the acquisition-to-descriptor chain.Source: §I · p. 1

Contributions

  • The study characterises 45 geometric descriptors, estimates per-descriptor noise floors, and benchmarks them against wound-care area-reduction thresholds.Its variance decomposition separates between-site morphological spread from within-site measurement error.Source: §I · pp. 1–2
  • Capture extent accounts for 75% of surface-area variance and 95% of bounding-box-diagonal variance; standardising the analysis region cuts the surface-area limit from 4.6 to 2.1.The paper presents capture extent, rather than sensor improvement, as the more addressable error source.Source: §I · p. 2
  • A fixed-scale curvature estimator is proposed because repeat scans can differ by up to a factor of 15 in vertex density, making per-vertex curvature incomparable across captures.The estimator is described as resolution-independent over a measured window.Source: §I · p. 2

II. MATERIALS

The study uses two commercial infrared structured-light scanners and rigid wound-care phantoms, with repeatability measured from free-hand scans on one device. The corpus does not support cross-device agreement or longitudinal analysis.

A. Sensor Taxonomy

  • The paper distinguishes triangulation-based structured light from time-of-flight sensing and states that neither scanner is LiDAR.The Apple TrueDepth camera is also described as structured light, unlike the rear-facing Apple LiDAR Scanner.Source: A. Sensor Taxonomy · p. 2

B. Scanners

  • Both scanners use infrared structured light and triangulation; the paper distinguishes point distance, a sampling density, from accuracy.Their stated minimum scan sizes place typical wounds near or below the devices’ declared operating bounds.Source: B. Scanners · p. 2, A. Sensor Taxonomy · p. 2

C. Phantoms

  • The targets are commercially manufactured training models representing staged pressure injuries, ulcers, dehisced surgical wounds, callus, and fissures.Their rigid, catalogue geometry makes repeat-scan differences identifiable as measurement error; polymer optical behaviour differs from living tissue.Source: C. Phantoms · p. 2

D. Corpus

  • The corpus contains 66 meshes: 26 from the Einstar and 40 from the RANGE, with nine RANGE sites repeated two or three times for 23 repeatability scans.Only two staging blocks were scanned on both devices, too few for cross-device agreement analysis; no longitudinal analysis is supported.Source: D. Corpus · p. 2

III. METHODS

Meshes are cleaned and smoothed before descriptor computation, but outlier removal creates a systematic area deficit. Descriptors are computed over captured mesh regions without wound delineation, so they do not measure isolated wound beds.

A. Preprocessing

  • Meshes undergo statistical outlier removal followed by Taubin smoothing, whose measured effect is reported rather than assumed.The outlier rule removes vertices beyond the mean neighbour distance plus two standard deviations; smoothing uses ten iterations.Source: A. Preprocessing · pp. 2–3
  • Statistical outlier removal discards 6.1% to 12.7% of true surface area on analytic surfaces, leaving reconstructed object area about 11% low.The authors state that the common-mode deficit does not affect headline ratios, but absolute areas are not calibrated measurements.Source: A. Preprocessing · pp. 2–3

B. Descriptors

  • All 66 meshes are open surfaces, so enclosed mesh volume and hull-minus-mesh estimated wound volume are undefined; convex hull volume is computed instead.Enclosed volume is reported as missing rather than assigned a geometrically meaningless value.Source: B. Descriptors · p. 3
  • No wound is segmented: surface area is measured over whatever mesh region the operator captured, often a limb or sacral region rather than a wound bed.These measurands differ from clinical planimetry, and comparisons are framed as a lower bound on noise an unsegmented pipeline must overcome.Source: What the descriptors are computed over · p. 3
  • The 45 descriptors include extensive measures that scale with captured surface and intensive ratios or per-unit measures intended to be scale-free.Per-vertex curvature excludes boundary vertices because the estimators are undefined there and boundary inclusion would add a crop-dependent artefact.Source: What the descriptors are computed over · p. 3

C. Scale-Normalised Curvature

Per-vertex curvature changes with mesh sampling, which varies substantially across repeat scans; the paper estimates curvature at fixed spatial scales and defines a variance framework for repeatability and morphology.

C. Scale-Normalised Curvature

  • Repeat scans differ by a median factor of 4.6 in vertex density, which varies by a factor of 43 across the corpus and within Revopoint scans.Discrete per-vertex curvature therefore reflects different effective sampling scales across scans of the same rigid site.Source: C. Scale-Normalised Curvature · p. 3
  • The scale-normalised estimator samples 4,000 points uniformly by area, fits local quadrics within a fixed radius, and evaluates radii of 2, 3, or 5 mm.Mean and Gaussian curvature are derived from the first and second fundamental forms; fixing sample count and support radius removes inherited mesh-resolution dependence over a measured window.Source: C. Scale-Normalised Curvature · p. 3

D. Variance Decomposition

  • The analysis estimates precision under conditions between repeatability and intermediate precision, but not trueness because no certified artefact was available.The conditions use the same operator, procedure, and instrument on an unchanged measurand.Source: D. Variance Decomposition · p. 3
  • A one-way random-effects model treats within-site variance as measurement error and between-site variance as morphological spread, with intraclass correlation representing morphology’s share of observed variance.The one-way form is used because sites share a procedure without a crossed rater factor; exact F-based confidence intervals follow Koo and Li,.Source: D. Variance Decomposition · p. 3

E. Precision and Detectable Change

The section distinguishes cohort-dependent descriptor reliability from within-site precision, which determines how large a change must be to exceed repeat-scan noise. Because errors are proportional for most descriptors, it reports primary limits on a log scale and back-transforms them.

E. Precision and Detectable Change

  • The repeatability coefficient RC = 2.77 σw is the 95% limit on the difference between two measurements of an unchanged wound.Within-site precision, rather than ICC, sets the change threshold for monitoring one wound over time.Source: E. Precision and Detectable Change · p. 4
  • Log-transformed limits are back-transformed because proportional errors and CVw above 50% make symmetric percentage limits implausible or awkward.A bounded quantity cannot fall by more than 100%.Source: E. Precision and Detectable Change · p. 4
  • Cluster-bootstrap confidence intervals use 2,000 site-level resamples because repeats within a site are not independent.The section also reports σb/σw because ICC and CVw can diverge when a descriptor varies little across wounds.Source: E. Precision and Detectable Change · p. 4

A. Capture Variability

Repeat captures of rigid sites vary substantially, and surface-area noise exceeds the clinical change scale used for comparison. The section cautions that this is a lower bound for an unsegmented pipeline, not a measure of wound-area reproducibility.

A. Capture Variability

  • The median within-site largest-to-smallest reconstructed surface-area ratio is 2.12, while vertex-count and vertex-density ratios are 3.5 and 4.6.Fig. 1 plots reconstructed surface area for repeat scans at each site, annotates within-site max/min ratios, and uses a logarithmic axis.Source: A. Capture Variability · p. 4, Figure 1 · p. 4

B. Precision and Detectable Change

  • Surface area has CVw = 54.4% (95% CI 38.1–69.0), median same-wound scan difference 60.8%, and minimum detectable change 151%; convex hull volume has CVw = 61.4% and MDC = 170%.Fig. 2 compares within-site CV and cluster-bootstrap 95% confidence intervals against the four-week area-reduction decision scale.Source: B. Precision and Detectable Change · p. 4, Figure 2 · p. 4
  • Surface area has a multiplicative 95% repeatability limit R95 = 4.77 (95% CI 3.02–6.95), corresponding to shifts from −79% to +377% on rescanning an unchanged wound.Table I presents multiplicative 95% repeatability limits with cluster-bootstrap confidence intervals; 1.0 would mean perfect repeatability.Source: B. Precision and Detectable Change · p. 4, Table I · p. 4
  • No descriptor reaches ICC 0.75 under any processing pipeline, while the sole descriptor with MDC below 50%, concave fraction, is nearly constant across sites.Concave fraction has CVw = 2.9%, ranges from 0.49 to 0.62 across 66 scans, and has signal-to-noise ratio 0.57.Source: Table II · p. 5, B. Precision and Detectable Change · p. 5
  • The surface-area noise factor of 4.77 is 2.2 times wider than the factor of 2.13 corresponding to a 53% four-week area reduction.This is a lower bound on noise an unsegmented pipeline must overcome, not wound-area reproducibility, because descriptors use the reconstructed patch rather than a delineated wound bed.Source: B. Precision and Detectable Change · p. 5

C. Variance Decomposition

Most descriptors show poor site discrimination, while standardising the analysis region improves precision. The best informative crop reaches, but does not clear, the clinical decision ratio; smaller and larger crops have distinct limitations.

C. Variance Decomposition

  • 44 of 45 descriptors fall below ICC 0.50, and none reaches 0.75.Mean ICC(1, 1) is 0.016 for extensive descriptors versus 0.195 for intensive descriptors.Source: C. Variance Decomposition · p. 5

1) Standardising the analysis region:

  • R95 falls from 4.59 as-acquired to 2.13 at 25 mm and 1.63 at 40 mm, but the 25 mm limit only matches the 2.13 clinical decision ratio.The table frames region standardisation as a trade-off between precision and surviving site-specific signal.Source: Table III · p. 5, 1) Standardising the analysis region · p. 5

D. Do the Obvious Remedies Help?

  • Mean ICC(1, 1) for extensive descriptors rises from 0.016 to 0.204 with cropping, but ICC understates the precision improvement because it is bounded by cohort spread.Precision is the relevant measure for whether a change is detectable.Source: 1) Standardising the analysis region · p. 5

1) Standardising the analysis region:

  • At 40 mm, the retained surface nearly fills the crop disc and surface-area and bounding-box-diagonal ICCs are both 0.000, so the crop—not the wound—is being measured.The authors do not treat its R95 of 1.63 as evidence of clinical performance.Source: 1) Standardising the analysis region · p. 5
  • At 15 mm, R95 worsens to 11.14 and four scans lose the region entirely.Region standardisation is therefore bounded at the small-radius end as well.Source: 1) Standardising the analysis region · p. 5

2) Removing the resolution dependence:

The fixed-scale curvature estimator reduces sampling-density dependence, but its accuracy holds only over a measured window and it does not improve reliability. Automated cavity delineation performs poorly because its mesh-derived reference inherits capture extent.

2) Removing the resolution dependence:

  • On a 20 mm-radius sphere, the estimator recovers H = 0.0502 mm−1 versus 0.05 and K = 0.002 52 mm−2 versus 0.0025; on a 15 mm-radius cylinder, H = 0.0336 versus 0.0333 and K ≈0.These are analytic-ground-truth validations of the fixed-scale estimator.Source: 2) Removing the resolution dependence · pp. 5–6
  • Mean ICC(1, 1) across the curvature block is 0.107 at r = 3 mm, and the maximum is 0.462; making curvature comparable across resolutions does not improve reliability.The estimator addresses comparability, not signal absent from the acquisition.Source: 2) Removing the resolution dependence · p. 6

3) Would segmenting the wound fix it?:

  • At a 2 mm threshold, automated delineation finds a cavity in only 12 of 23 repeat-set scans, and paired areas differ by a median factor of 64.5, about an order of magnitude worse than the whole-patch R95 of 4.59.The mesh-derived reference inherits capture extent; the authors identify manual tracing by trained raters as the appropriate fix.Source: 3) Would segmenting the wound fix it? · p. 6

4) Where the fixed-scale estimator stops working:

  • At r = 3 mm, median absolute bias grows from 4.2 %–5.3 % at densities of 16 to about 4 mm−2 to 40.3 % at 0.71 mm−2; r = 5 mm stays within 2.6 %–8.6 % across the ladder.Larger support extends the resolution window but truncates genuine small-scale curvature, so support must suit both density and feature size.Source: 4) Where the fixed-scale estimator… · p. 6

5) The combined screen:

Across descriptor–pipeline combinations, the combined screen finds no reliable site discriminator; controlling capture extent substantially improves precision, while smoothing preserves bulk geometry but alters curvature and roughness. The error budget attributes most variance in several size descriptors to capture extent rather than the instrument.

5) The combined screen:

  • None of 248 descriptor–pipeline combinations reaches ICC(1, 1) = 0.75, and none combines CVw < 25 % with σb/σw > 1.Only mean curvature as-acquired and Gaussian-curvature kurtosis under 25 mm region standardisation have between-site signal at least as large as within-site noise, and both are too imprecise to act on.Source: 5) The combined screen · p. 6
  • Region standardisation improves longitudinal precision and brings surface-area precision close to sufficient, but neither tested remedy makes a descriptor a reliable discriminator between wound sites.The section distinguishes within-site precision for monitoring from between-site discrimination in this cohort.Source: 5) The combined screen · p. 6

E. Error Budget

  • Capture extent accounts for 75 % of surface-area variance, 91 % of hull-area variance, and 95 % of bounding-box-diagonal variance under 25 mm region standardisation.The corresponding repeatability limits fall from 4.59 to 2.13, 4.89 to 1.61, and 2.51 to 1.23; enclosed hull volume is the exception, with 48 % of its variance attributable to extent.Source: E. Error Budget · p. 6
  • Eliminating instrument error would reduce the surface-area limit only from 4.59 to about 2.1, while the residual error in enclosed volume is linked to reconstructing a surface that is nowhere watertight.Region standardisation already achieves about 2.1 for surface area, so the paper concludes that a better sensor would buy little for that descriptor.Source: E. Error Budget · p. 6

F. Effect of Smoothing

  • After ten iterations, Taubin smoothing changes convex-hull volume by a median −0.03 %, compared with +1.29 % for Laplacian smoothing.Both values are measured on 14 meshes; the reported IQRs are −0.07 to −0.00 % and +0.49 to +2.84 %, respectively.Source: F. Effect of Smoothing · p. 6
  • Ten Taubin iterations attenuate root-mean-square curvature by 49.8 %, while Laplacian smoothing raises it 435 % at ten iterations and by two orders of magnitude by thirty.The Laplacian increase is attributed to triangles approaching degeneracy and the angle-deficit estimator dividing by vanishing vertex areas.Source: F. Effect of Smoothing · p. 6
  • Roughness indices depend jointly on the wound and smoothing parameters, so values are not comparable across studies that smooth differently.Smoothing suppresses high-frequency content that roughness descriptors measure.Source: F. Effect of Smoothing · p. 7

V. DISCUSSION

In free-hand consumer scanning, capture variability—not sensor resolution—is the principal constraint identified for these measurements, and acquisition controls are presented as practical remedies. The discussion situates the results against prior work and emphasizes boundaries involving the small repeat-site sample, instrument, phantom, and measurement design.

A. What Limits These Measurements

  • The binding constraint is inconsistent free-hand coverage: repeated captures of one site vary twofold in area and up to seventeenfold in vertex count.Size-dependent descriptors inherit coverage variance directly; intensive descriptors are affected indirectly through resolution-dependent estimators and boundary effects.Source: A. What Limits These Measurements · p. 7
  • The authors identify fixed standoff, a registration fiducial, a defined wound margin, and fixed-density resampling as acquisition remedies, favoring control at capture time.They describe these post hoc remedies as individually insufficient and propose controlling framing during acquisition.Source: A. What Limits These Measurements · p. 7

B. Relation to Prior Work

  • Prior studies’ high 3-D area reliability is attributed largely to fixed or tripod-mounted capture, defined margins, and trained operators, unlike the free-hand consumer regime studied here.The authors also place their findings alongside reviews describing 3-D wound technology as under-evidenced and reports of limitations in depth and volume measurement.Source: B. Relation to Prior Work · p. 7

VI. LIMITATIONS

  • Only nine repeat sites support the variance decomposition, and 43 of 45 ICC lower bounds are zero, so the claim that no descriptor reaches good reliability concerns point estimates.The intervals cannot exclude several descriptors clearing that bar in a larger study; within-site precision estimates are better constrained, based on 14 within-site pairs.Source: §VI · p. 7
  • The pooled repeatability estimate blends same-session and cross-session sites, but session is confounded with phantom identity.Surface-area repeatability limits are 5.57 for five same-session sites and 3.44 for four cross-session sites; the contrast cannot isolate a session effect.Source: §VI · p. 7
  • On the open meshes, the retained sphericity descriptor reaches 8.9 despite a theoretical bound of 1, so it is not sphericity in the usual sense.It is retained because it appears in comparable studies and its instability is informative.Source: §VI · p. 7
  • Rigid polymer phantoms differ optically from living tissue, and without a certified geometry or reference measurement the study estimates precision, not trueness.Whether phantom repeatability is optimistic or pessimistic for patients is not established, although the absence of patient movement suggests it may be optimistic.Source: §VI · p. 7
  • All repeatability and error-budget results characterize the Revopoint RANGE; the study reports neither cross-instrument reproducibility nor smartphone capture or longitudinal healing prediction.The Einstar contributes single scans without repeat pairs, smartphone scans lack mesh topology, and reliable acquisition timestamps are unavailable.Source: §VI · p. 7

VII. CONCLUSION

This study characterizes repeatability for geometric wound descriptors from free-hand consumer structured-light scans of rigid phantoms. It reports large surface-area repeat differences and weak descriptor reliability, while identifying capture extent as a key source of error and 25 mm region standardisation as a partial remedy.

VII. CONCLUSION

  • The surface-area repeatability limit is a factor of 4.8, 44 of 45 descriptors fall below ICC 0.50, and none of 248 descriptor–pipeline combinations reaches 0.75.The surface-area limit corresponds to a −79 to +377 % difference between scans of an unchanged wound.Source: §VII · p. 8
  • A 25 mm analysis region cuts the surface-area limit to 2.13, meeting rather than clearing the ratio represented by the clinical threshold.The fixed-scale curvature estimator is resolution-independent across the repeat-set density range but adds no reliability of its own in this cohort.Source: §VII · p. 8
  • The authors recommend reporting descriptor noise floors and checking healing-prediction claims against them.The conclusion frames this as a reporting and evaluation implication for descriptors of this kind.Source: §VII · p. 8
Loading 2608.30143v1…