{"id":"302239aa-1c83-432d-b6a0-909465e40a37","arxiv_id":"2508.17635","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Wound3DAssist produces 3D models of wounds from monocular smartphone video and combines 3D reconstruction with 2D segmentation to measure area, perimeter, depth, and tissue composition.","lead":"This paper describes Wound3DAssist, a smartphone-based system that turns a short video of a wound into a 3D model and automatically measures its size, depth, and tissue types. If it works in practice, nurses and doctors could track chronic wounds more accurately and quickly than with rulers or tracing paper.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth/volume claims rest on an unvalidated surface-cover assumption; no experiment compares computed depth to known ground truth.","rationale":"The reader's weakest_assumption correctly identifies the TPS/RBF surface cover as the load-bearing premise, and I agree. The central advertised capability is automated, view-independent 3D wound measurement, and the metric that cannot be obtained from 2D images—depth—depends entirely on the unvalidated surface cover. The paper's own limitations section concedes the absence of ground-truth annotations for real clinical data, and the clinical validation compares only against Wintape/ruler, which do not measure depth. The strongest available evidence, the digital dataset with known geometry, is used only for reconstruction accuracy, not for metric accuracy. A targeted experiment on that dataset would directly test the assumption and either support or undermine the depth claims. I do not see an internal inconsistency in the pipeline, and the reconstruction and segmentation components are plausibly adequate; the issue is that the headline accuracy claim goes beyond what is currently validated. Since the reader already returned CONDITIONAL, my read does not change that verdict, but it sharpens the condition: depth accuracy must be demonstrated against known geometry before the framework's core clinical claims are accepted.","tokens_in":13960,"tokens_out":6737,"duration_ms":71757,"concrete_test":"Run the framework's metric pipeline on the Syn3DWound digital dataset (Sec. II-A, [23]) for wounds placed on both flat and strongly curved anatomical regions, using the known pre-injury skin surface as ground truth. Compute the surface cover via Eq. 3 and maximum depth via Sec. IV-C4, then compare to ground-truth depth and local surface curvature. If depth error exceeds about 1-2 mm on curved regions while staying small on flat ones, the surface-cover assumption is the limiting factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the TPS/RBF wound bed surface cover (Sec. IV-C, Eq. 3) equals the pre-injury skin surface. The surface is fitted only to wound-perimeter vertices, minimizes bending energy, and is described as 'intended to represent the skin's original location prior to injury'—but no experiment validates this equivalence. All depth measurements (Sec. IV-C4) are vertical distances from wound-bed vertices to this virtual surface. On curved anatomy, a minimal-curvature TPS surface through boundary vertices bridges concavities and flattens convex regions, so depth errors can be large even with a perfect reconstruction. The clinical comparison (Sec. V-C1) uses Wintape and ruler, neither of which measures depth, and the four clinical cases have no ground-truth depth. The digital dataset (Sec. II-A) has known geometry but is used only to evaluate reconstruction, not metric accuracy; detailed silicone results are deferred to prior work [24]. Therefore the headline 'millimeter-level accuracy' claim is not supported for the depth measurement, the component that most directly justifies 3D over 2D assessment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Wound3DAssist, a modular framework for 3D wound assessment from monocular smartphone videos. The pipeline combines photogrammetric reconstruction (Meshroom), a fine-tuned SegFormer 2D segmentation network whose predictions are rasterized onto the 3D mesh, and a geometry-processing module that computes wound bed perimeter, length, width, surface area, depth, and tissue composition. The authors evaluate the framework across three tiers: a synthetic digital wound dataset with known geometry, silicone phantoms, and a prospective clinical dataset from nine patients (with four reconstructions). They report sub-millimeter precision on the phantom dataset (deferred to prior work [24]), an average point-wise reconstruction error on the four clinical cases, improved 2D-to-3D segmentation relative to 2D baselines with statistical significance for wound bed and periwound, and longitudinal tracking of wound progression on two patients. The full documentation pipeline is reported to run in approximately 18 minutes on a high-end workstation.","tokens_in":14195,"tokens_out":4623,"duration_ms":48528,"significance":"If the depth and volume measurements were properly validated, the framework would be a practically relevant contribution: it uses consumer-grade hardware, produces colored 3D models, integrates segmentation and tissue classification, and demonstrates longitudinal tracking. The modular design and the systematic comparison of 2D versus 2D-to-3D segmentation mapping are useful methodological contributions. However, the principal claim that distinguishes this work from 2D assessment—millimeter-accurate depth measurement—is not supported by the evidence in the manuscript. The digital and silicone validation tiers are deferred to the authors' prior papers [23], [24], and the clinical evaluation has no depth ground truth, relies on manual alignment, and uses only four cases. The paper is best characterized as a promising systems paper whose central quantitative claim requires additional experiments.","major_comments":[{"comment":"The wound bed depth metric is defined as the vertical distance between wound bed vertices and the TPS/RBF surface cover fitted to the perimeter vertices. This surface is described as 'intended to represent the skin's original location prior to injury,' but no experiment validates that the fitted surface matches the pre-injury skin geometry. The digital dataset with known geometry (Sec. II-A) is used only for reconstruction evaluation (Sec. V-A1), and the clinical comparison (Sec. V-C1) uses ruler and Wintape, neither of which measures depth. Consequently, the abstract's claim of 'millimeter-level accuracy' is not supported for the depth measurement, which is the principal justification for 3D assessment over 2D. Please add a validation experiment, for example by computing depth errors against the known geometry in the digital dataset or by comparing the TPS cover to a ground-truth skin surface on a phantom.","section":"Sec. IV-C4 and Eq. (3)"},{"comment":"The clinical reconstruction evaluation is limited to four cases and requires manual alignment of the reconstruction to the Revopoint reference because the ArUco markers were unreliable in the RGB stream. The manuscript does not state the numeric value of the average point-wise error (the symbol is missing), nor does it quantify the uncertainty introduced by manual alignment. Moreover, the reported errors are concentrated at the wound bed boundaries (Fig. 7), which is exactly the region used for perimeter and depth measurement. Please report per-case errors with the exact average, and provide a sensitivity analysis of the manual alignment, or downgrade the clinical accuracy claim accordingly.","section":"Sec. V-A3"},{"comment":"The segmentation evaluation is performed on 2D re-projections of the 3D segmentation against 2D expert annotations; there is no ground-truth 3D segmentation. The Wilcoxon signed-rank test yields statistical significance for wound bed and periwound, but not for granulation or slough, which are the two tissue classes with sufficient samples. The claim that the 2D-to-3D method improves tissue segmentation is therefore only partially supported. Please provide additional evidence for the tissue classes, or refine the claim to reflect the classes for which significance was shown.","section":"Sec. V-B"}],"minor_comments":[{"comment":"The manuscript contains many garbled characters and placeholder values (e.g., Sec. V-A3 'approximately � mm', Sec. IV-A1 '� � �� frames', and the figure captions for Figs. 5, 10, 11, 12, and 13). These must be corrected to allow proper review and to make the reported results readable.","section":"Throughout"},{"comment":"The B-spline perimeter estimate introduces a smoothness regularization parameter 's', but the manuscript does not report the value used in the experiments or any sensitivity analysis for this parameter.","section":"Sec. IV-C1, Eq. (5)"},{"comment":"The geodesic distance is computed by projecting intermediate Euclidean points onto the surface cover and fitting a B-spline; please clarify how the projection is well-defined when the line segment between perimeter vertices lies partially outside the convex hull of the fitted surface domain.","section":"Sec. IV-C2, Eq. (7)"},{"comment":"The statement 'Similar outcomes were observed across additional cases' is not backed by a quantitative comparison. Please include a table reporting framework and manual measurements (area, perimeter, and depth, where available) for all four prospective cases.","section":"Sec. V-C1"},{"comment":"The weighting factor in Eq. (1) and the voting scheme in Eq. (2) are described with notation that is partially illegible in the submitted version; please ensure all symbols and index ranges are typeset correctly.","section":"Sec. IV-B2, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a preliminarily formatted draft with many unrendered equations and missing numeric values; the version sent to reviewers should be thoroughly copy-edited. The reliance on prior conference papers [23], [24] for two of the three validation tiers leaves this paper with thin independent evidence for its headline claims; the authors should either integrate the key results from those papers or clearly state that the overall accuracy claim is cumulative across works. The depth-validation gap is the central concern and should be addressed with a direct experiment before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, it is a competent systems-integration paper: smartphone video in, a textured 3D mesh out, with 2D-to-3D segmentation and a set of clinically motivated wound metrics. The genuinely new piece is the normal-weighted 2D-to-3D majority voting and the TPS surface-cover used to estimate depth, area, and volume. That combination is not in the cited prior work. Second, the paper's headline claim, \"millimeter-level accuracy\", overreaches for the depth metric specifically, because depth is never checked against any ground truth. The stress-test note is right: the surface cover is fitted to perimeter vertices with a bending-energy-minimizing TPS, and all depth/volume numbers are distances to that virtual surface. On curved anatomy that surface can deviate a lot from the true pre-injury skin, and no experiment quantifies that error. The clinical comparison uses Wintape and ruler, both of which ignore depth. So the abstract's accuracy claim is really only supported for reconstruction geometry, not for the depth measurement that justifies 3D over 2D.\n\nCredit where it is due: the authors are unusually candid about limitations. They explicitly say the ArUco alignment failed on clinical data, that manual alignment was used instead, that the clinical set is only four cases, and that necrotic and epithelial classes are omitted due to sparse data. The modular design and the decision to stick with Meshroom after comparing against NeRF and Gaussian Splatting is reasonable. The longitudinal case studies, while qualitative, are a useful sanity check. The 18-minute runtime is a practical claim that is backed by measurements.\n\nSoft spots, in proportion. The four-case clinical evaluation is small, and the approximate 1 mm AD is reported without error bars or per-case detail. Two of the three validation tiers (digital and silicone) are explicitly deferred to the authors' own prior papers [23] and [24]; as a reviewer you would want those numbers reproduced or at least summarized. The statistical significance claim for the 2D-to-3D segmentation improvement is based on a Wilcoxon test, but the text says it only holds for wound bed and periwound, not for granulation and slough, so the claim that 2D-to-3D outperforms 2D is weaker than it looks. None of these are fatal, but they add up.\n\nWho this is for: a clinical imaging researcher or a referee for a medical-vision venue. It is a useful integration of existing building blocks and it provides a reasonable template for validation on real patient data. It deserves peer review, but the depth and volume claims need either a phantom with known ground-truth depth or a validation study against a structured-light scanner or CT, plus more than four clinical cases, before the abstract's headline can stand. I would send it out, with a request for major revisions.","headline":"A transparent, well-engineered system paper whose real novelty is the surface-cover depth estimation, but the headline millimeter-level accuracy claim is not yet supported for depth because that component has no ground-truth validation.","tokens_in":770,"tokens_out":791,"would_cite":false,"duration_ms":24097,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A short smartphone video can yield a 3D wound model with automated depth, area, and tissue metrics.","keywords":["3D wound assessment","monocular photogrammetry","wound segmentation","tissue classification","depth measurement","smartphone video","longitudinal wound monitoring","3D mesh analysis"],"falsifier":"On a curved calibration phantom with a known cavity geometry, record a handheld video, run the full pipeline, and compare the computed depth to caliper-measured depth; if depth error grows systematically with surface curvature, the smooth cover does not represent pre-injury skin.","tokens_in":13807,"feed_emoji":"🩹","tokens_out":5968,"duration_ms":62390,"temperature":0.7,"pith_summary":"This paper sets out to show that accurate 3D wound assessment no longer requires specialized scanners: a short handheld smartphone video, processed by photogrammetry and neural segmentation, can be turned into a colored 3D model from which area, perimeter, length, width, depth, and tissue composition are computed automatically. The motivation is that current manual methods, such as ruler and tracing techniques, are subjective, view-dependent, and cannot capture wound depth on curved anatomy. The authors validate the claim on synthetic models with known geometry, silicone phantoms, and real patient videos, reporting millimeter-level reconstruction accuracy and statistically improved 3D segmentation over single-view 2D segmentation for wound bed and periwound regions. Full documentation runs in under 20 minutes, which the paper argues makes the workflow feasible for routine clinical use.","feed_headline":"Phone video yields millimeter-accurate 3D wound maps","feed_subtitle":"A short handheld clip becomes a 3D wound model with depth, area, and tissue metrics in under 20 minutes.","key_machinery":"The load-bearing object is the wound bed surface cover: a virtual smooth surface fitted to the wound perimeter vertices using radial basis function interpolation with a thin-plate spline kernel. This surface stands in for where the skin was before injury, and every depth and volume-related measurement is computed as the vertical distance from a wound-bed vertex to this cover. It also carries the geodesic length and width calculations, since those paths are projected onto the cover before their arc lengths are measured.","core_discovery":"On its own terms, the paper argues that monocular videogrammetry of a wound is sufficient: after key-frame selection, an open-source structure-from-motion pipeline builds a textured 3D mesh; a fine-tuned transformer-based segmentation network labels wound bed, periwound, and tissue classes in 2D; majority voting over visible views, weighted by viewing angle, projects those labels onto the mesh; and all clinical measurements are then made directly on the mesh. Depth is defined relative to a smooth surface cover fitted to the wound perimeter, and scale is recovered from fiducial markers placed in the scene. The evaluation claims sub-millimeter precision on phantoms, millimeter-level surface agreement on real wounds, view-consistent segmentation that is at least as good as single-view 2D segmentation, and longitudinal trends that match both healing and non-healing courses.","pith_inferences":["If the perimeter-cover assumption holds on curved anatomy, the same measurement layer could be applied to other cavity-like surface conditions, such as pressure injuries or surgical wounds, without retraining the geometry pipeline.","The view-weighted voting rule predicts a testable property: deliberately adding oblique or blurry frames should barely change the final 3D labels, because low-confidence views are down-weighted; a perturbation study varying acquisition angle would quantify this robustness.","A natural stress test the paper does not run is scan-rescan repeatability: recording the same wound twice after repositioning the camera should yield nearly identical depth and area, which would separate reconstruction noise from true healing change.","As neural rendering methods mature, the same cover-and-voting machinery could operate on their meshes, making the measurement step independent of whichever reconstruction backend proves fastest or most accurate."],"forward_implications":["Wound area, perimeter, length, width, and depth can be measured without touching the wound, from a video captured on an ordinary phone, making repeated assessment practical in telehealth and home care.","Because measurements are taken on the reconstructed 3D surface rather than a single photograph, they should be view-independent and robust to camera motion during capture.","The same reconstructed mesh can be revisited across consultations; the paper shows area shrinkage and a slough-to-granulation transition in a healing patient and area growth in a non-healing patient.","View-weighted 2D-to-3D label voting yields wound bed and periwound segmentation that is at least as accurate as single-view 2D segmentation, so the 3D step does not sacrifice tissue classification quality.","The modular design lets reconstruction or segmentation components be swapped without changing the measurement layer, so future accuracy improvements can enter the pipeline incrementally."],"supporting_citations":[{"why":"Supplies the synthetic wound dataset with known geometry used for quantitative reconstruction evaluation.","marker":"[23]"},{"why":"Companion evaluation of reconstruction methods on consumer wound videos that informs the chosen reconstruction backend.","marker":"[24]"},{"why":"Open-source photogrammetry system adopted for camera pose estimation and mesh generation.","marker":"[32]"},{"why":"Transformer segmentation architecture that the paper fine-tunes for wound bed, periwound, and tissue classes.","marker":"[30]"},{"why":"Rasterization method used to project mesh faces onto 2D segmentation masks for label voting.","marker":"[33]"},{"why":"Provides the scale-ambiguity result and the direct linear transform used to turn fiducial marker corners into metric scale.","marker":"[36]"},{"why":"Earlier multi-view wound segmentation work whose 2D-to-3D labeling approach the paper extends and compares against.","marker":"[6]"},{"why":"Defines the manual ruler and tracing measurement practice used as coarse sanity-check benchmarks.","marker":"[37]"},{"why":"Recent photogrammetry-plus-segmentation system used as a scope baseline for the contribution table.","marker":"[5]"}],"fun_headline_variants":["Smartphone video builds 3D wound map in minutes","Phone clip yields millimeter-accurate 3D wound model","3D wound assessment from a quick phone video","Handheld clip makes 3D wound measurements possible","Video-only 3D wound mapping with mm precision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The depth and volume numbers all rest on the assumption that the smooth surface fitted to the wound perimeter is where the skin was before injury; if that cover is wrong on curved anatomy, every depth metric is wrong regardless of reconstruction quality.","fun_headline_variants_meta":{"raw":{"variants":["Smartphone video builds 3D wound map in minutes","Phone clip yields millimeter-accurate 3D wound model","3D wound assessment from a quick phone video","Handheld clip makes 3D wound measurements possible","Video-only 3D wound mapping with mm precision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1367,"prompt_tokens":908,"completion_tokens":459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":381}},"tokens_in":524,"tokens_out":459,"duration_ms":4599,"temperature":1.0,"reasoning_tokens":381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:02:42.256835+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a curved calibration phantom with a known cavity geometry, record a handheld video, run the full pipeline, and compare the computed depth to caliper-measured depth; if depth error grows systematically with surface curvature, the smooth cover does not represent pre-injury skin.","supporting_citations":[{"cited_title":"Wound measurement by RGB-D camera,","cited_arxiv_id":null,"evidence_quote":"Supplies the synthetic wound dataset with known geometry used for quantitative reconstruction evaluation."},{"cited_title":"Three-dimensional wound reconstruction using point descriptors: A comparative study,","cited_arxiv_id":null,"evidence_quote":"Companion evaluation of reconstruction methods on consumer wound videos that informs the chosen reconstruction backend."},{"cited_title":"Autonomous robot-driven chronic wound 3D reconstruction and analysis system,","cited_arxiv_id":null,"evidence_quote":"Open-source photogrammetry system adopted for camera pose estimation and mesh generation."},{"cited_title":"Develop- ment of a robust photogrammetric metrology system for monitoring the healing of bedsores,","cited_arxiv_id":null,"evidence_quote":"Transformer segmentation architecture that the paper fine-tunes for wound bed, periwound, and tissue classes."},{"cited_title":"System design for 3D wound imaging using low-cost mobile devices,","cited_arxiv_id":null,"evidence_quote":"Rasterization method used to project mesh faces onto 2D segmentation masks for label voting."},{"cited_title":"FUSeg: The Foot Ulcer Segmentation Challenge","cited_arxiv_id":"2201.00414","evidence_quote":"Provides the scale-ambiguity result and the direct linear transform used to turn fiducial marker corners into metric scale."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier multi-view wound segmentation work whose 2D-to-3D labeling approach the paper extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Recent photogrammetry-plus-segmentation system used as a scope baseline for the contribution table."}],"review_version":2}