Wound3DAssist: 3D Wound Assessment

• Wound3DAssist is a cutting-edge framework that generates accurate 3D wound models from short smartphone videos to overcome 2D imaging limitations.
• It integrates photogrammetric 3D reconstruction, advanced segmentation, and periwound mapping in a modular workflow, delivering full assessments within 20 minutes.
• Evaluations report millimeter-level accuracy and reliable tissue analysis, supporting both snapshot measurements and longitudinal monitoring.

Searching arXiv for the named framework and closely related wound-analysis papers to ground the article in current literature. to=arxiv_search tool 大发彩票网 json {"query":"Wound3DAssist practical framework for 3D wound assessment arXiv", "max_results": 5} to=arxiv_search tool аанацҳауеит json {"query":"wound assessment 3D wound segmentation consumer-grade videos wound arXiv", "max_results": 10} Wound3DAssist is a practical framework for 3D wound assessment using monocular consumer-grade videos that addresses limitations of 2D digital videometry, including perspective distortion, a limited field of view, and an inability to capture wound depth, especially in anatomically complex or curved regions (Chierchia et al., 25 Aug 2025). It generates accurate 3D models from short handheld smartphone video recordings and integrates 3D reconstruction, wound segmentation, tissue classification, and periwound analysis into a modular workflow. In reported evaluations across digital models with known geometry, silicone phantoms, and real patients, the framework supports high-quality wound bed visualization, millimeter-level accuracy, and reliable tissue composition analysis, with full assessments completed in under 20 minutes (Chierchia et al., 25 Aug 2025). The broader research context places Wound3DAssist at the intersection of automated wound classification and segmentation (Jahangir et al., 2024), mobile wound analysis (Borst et al., 2024), multi-view consistent 3D wound modeling (Chierchia et al., 23 Jan 2026), RGB-D wound reconstruction (Harkämper et al., 26 Jan 2026), and benchmarked consumer-video reconstruction for wound surfaces (Chierchia et al., 2024).

1. Concept and clinical scope

Wound3DAssist is motivated by the inadequacy of 2D wound documentation for curved anatomy, wound cavities, and view-dependent measurements (Chierchia et al., 25 Aug 2025). The reported clinical requirements include non-contact, reproducible, and view-independent measurements of wound bed area, perimeter, length/width, depth, and volume; tissue composition; and periwound characterization, together with standardized longitudinal tracking and clear visualization of wound bed geometry and surrounding tissues (Chierchia et al., 25 Aug 2025). This framing is consistent with prior wound-analysis literature that emphasizes prompt and accurate segmentation for effective treatment and automated wound care management (Jahangir et al., 2024).

The framework is situated within a continuum of wound-AI systems. Earlier work emphasized automated 2D wound classification and segmentation using modified U-Net architectures for foot ulcers, pressure ulcers, leg ulcers, and infected wounds (Jahangir et al., 2024). Mobile-focused studies investigated lightweight segmentation architectures such as ENet, UNeXt, and TopFormer for smartphone deployment (Borst et al., 2024). Other work introduced remote pressure-ulcer assistance through Residual U-Net with attention (Chae et al., 2021), smartphone wound localization with YOLOv3 and tiny-YOLOv3 (Anisuzzaman et al., 2020), and multimodal referral support combining wound images and clinical notes (Fard et al., 22 Jan 2025). Wound3DAssist extends this trajectory by making 3D reconstruction from handheld video the core geometric substrate for wound measurement and interpretation (Chierchia et al., 25 Aug 2025).

A crucial distinction in the literature is between 2D segmentation consistency and 3D geometric consistency. Multi-view consistent neural-field methods were proposed precisely because inconsistent 2D segmentations across views lead to unreliable measurement and tissue assessment, especially for irregular topologies and undermined edges that are inherently 3D (Chierchia et al., 23 Jan 2026). This suggests that Wound3DAssist is best understood not simply as a segmentation system, but as a modular measurement framework in which segmentation is only one stage in a larger geometry-aware pipeline.

2. Pipeline architecture and modular workflow

The reported Wound3DAssist pipeline is modular and end-to-end, beginning with short handheld smartphone video acquisition and proceeding through key-frame selection, 3D reconstruction, wound segmentation, 2D-to-3D mapping, periwound analysis, measurement extraction, and visualization/report generation (Chierchia et al., 25 Aug 2025). Video acquisition follows a structured protocol to ensure multiple angles around the wound and capture fiducial markers (ArUco) for metric scale (Chierchia et al., 25 Aug 2025). Key-frame selection automatically retains sharp frames with good wound visibility and marker presence, typically $N \approx 50$ frames, balancing coverage and runtime (Chierchia et al., 25 Aug 2025).

For 3D reconstruction, the framework uses photogrammetry-based SfM/MVS via Meshroom (AliceVision), configured for single-camera video and high-density SIFT with up to 50k features per image (Chierchia et al., 25 Aug 2025). The camera model is the standard pinhole formulation

$x = K [R \mid t] X,$

and bundle adjustment minimizes reprojection error,

$$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$

(Chierchia et al., 25 Aug 2025). Metric scaling uses ArUco markers with known side length:

$s = \frac{L_{\text{known}}}{L_{\text{recon}}},$

after estimating reconstructed side lengths from triangulated marker corners (Chierchia et al., 25 Aug 2025).

Segmentation is performed in 2D using SegFormer with MiT-b5 encoder for wound bed, periwound, and tissue classes (Chierchia et al., 25 Aug 2025). The 2D predictions are then rasterized onto the reconstructed mesh, and per-face labels are aggregated via weighted majority voting using view-angle weights (Chierchia et al., 25 Aug 2025). For a mesh face $f_j$, the weight from image $i$ is

$w_j^i = |\langle n_{f_j}, \pi_{f_j}^i \rangle|,$

and the face label is computed as

$$L_j = \arg\max \sum_{i=1}^{N} \mathbf{1}_{w_j^i \ge 1/2} \log\left(\frac{w_j^i}{1-w_j^i}\right) p_i$$

(Chierchia et al., 25 Aug 2025). Oblique views are excluded for tissue mapping through the threshold $w_j^i \ge \cos(\pi/3)=1/2$, while periwound mapping uses all views because coverage is sparser (Chierchia et al., 25 Aug 2025).

The modularity of this workflow aligns with related studies that separate localization, segmentation, and higher-level analysis. Smartphone localizers based on YOLOv3 and tiny-YOLOv3 were proposed specifically to isolate wound ROIs for downstream segmentation and classification (Anisuzzaman et al., 2020). A plausible implication is that Wound3DAssist can incorporate such localization modules upstream when field-of-view clutter or privacy constraints require wound-centric cropping before reconstruction or segmentation.

3. Geometric reconstruction and 3D measurement formulation

A defining property of Wound3DAssist is that measurement is performed on reconstructed 3D meshes rather than directly on 2D masks (Chierchia et al., 25 Aug 2025). Surface area is computed on triangulated meshes as

$$A = \sum_{t \in T} \frac{1}{2} \| (v_{t2} - v_{t1}) \times (v_{t3} - v_{t1}) \|,$$

either for wound bed faces or periwound faces according to labels (Chierchia et al., 25 Aug 2025). Volume can be derived for a closed mesh by

$x = K [R \mid t] X,$0

or by integrating depth relative to a fitted cover surface (Chierchia et al., 25 Aug 2025).

Perimeter is extracted from boundary edges of connected wound components, then represented as a periodic $x = K [R \mid t] X,$1-smooth B-spline $x = K [R \mid t] X,$2, with arc-length

$x = K [R \mid t] X,$3

(Chierchia et al., 25 Aug 2025). Length and width are geodesic measures defined on a wound cover surface $x = K [R \mid t] X,$4, itself estimated from perimeter points by thin-plate interpolation:

$x = K [R \mid t] X,$5

The interpolation weights are obtained by minimizing

$x = K [R \mid t] X,$6

(Chierchia et al., 25 Aug 2025). Geodesic distances between perimeter vertices are then used to define wound length and width:

$x = K [R \mid t] X,$7

(Chierchia et al., 25 Aug 2025).

Depth is defined relative to the cover surface after PCA-based reorientation of the mesh:

$x = K [R \mid t] X,$8

Negative values indicate depression below the cover surface, whereas positive values indicate protrusion above it (Chierchia et al., 25 Aug 2025). This formulation is explicitly designed to remain clinically consistent on curved anatomy, where planar approximations would confound projection error with true cavity geometry (Chierchia et al., 25 Aug 2025).

Parallel research offers alternative geometric constructions. A fast RGB-D pipeline based on Intel RealSense D435, RGB-D odometry, and B-spline surface reconstruction computes perimeter, surface area, and wound dimensions from reconstructed wound bed meshes, with “depth” defined as the smallest extent of the minimal-volume bounding box enclosing the wound bed (Harkämper et al., 26 Jan 2026). Multi-view neural-field work instead reconstructs a signed distance field $x = K [R \mid t] X,$9, extracts the wound surface as the zero level set $$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$0, and computes area, volume, and maximum depth relative to a periwound reference plane (Chierchia et al., 23 Jan 2026). These alternative formulations underscore that Wound3DAssist belongs to a family of geometry-first wound systems, but its reported implementation favors photogrammetric mesh reconstruction with explicit surface-based measurements (Chierchia et al., 25 Aug 2025).

4. Segmentation, tissue analysis, and periwound mapping

Segmentation in Wound3DAssist is not limited to wound-bed delineation. The framework uses SegFormer (MiT-b5) models for binary wound bed segmentation, binary periwound segmentation, and multiclass tissue segmentation comprising granulation, slough, necrotic, and epithelial classes (Chierchia et al., 25 Aug 2025). The validation clinical set includes 9 patients, with tissue occurrences reported as granulation (23), slough (37), necrotic (3), and epithelial (2), and the paper notes that necrotic and epithelial were insufficiently represented for robust quantitative analysis (Chierchia et al., 25 Aug 2025).

Although the exact training losses for Wound3DAssist are not explicitly reported, the paper states that typical formulations include Dice loss,

$$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$1

and cross-entropy or focal loss for multiclass tissues to handle class imbalance (Chierchia et al., 25 Aug 2025). In the broader wound-segmentation literature, BCE and CrossEntropyLoss are common for binary and multiclass wound tasks, respectively (Jahangir et al., 2024, Borst et al., 2024). A plausible implication is that Wound3DAssist can interchange segmentation backbones or training objectives without changing its downstream 2D-to-3D aggregation and measurement logic.

Periwound analysis is treated as a first-class output rather than an afterthought. Periwound labels are predicted in 2D and transferred to 3D via the same weighted mapping mechanism, and periwound surface area is then quantified as the sum of areas of faces labeled as periwound (Chierchia et al., 25 Aug 2025). The paper further notes that erythema, edema, and maceration are clinically relevant color/texture metrics, but specific thresholds and scales for such indices are not reported and remain future extensions (Chierchia et al., 25 Aug 2025). This is significant because many earlier wound systems focused on wound boundaries or tissue types alone, whereas Wound3DAssist explicitly expands the assessment space to surrounding tissue context (Chierchia et al., 25 Aug 2025).

Related work supports this emphasis on tissue semantics and multi-view regularization. WoundNeRF represents wound tissues through a semantic head over a 3D field, producing per-class logits for background and wound tissues: granulation, slough, necrotic, epithelial, and unknown (Chierchia et al., 23 Jan 2026). It then derives a wound bed logit by log-sum-exp aggregation over the five tissue classes,

$$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$2

and renders consistent tissue distributions across views (Chierchia et al., 23 Jan 2026). This suggests a possible future direction in which Wound3DAssist could replace per-view 2D rasterization with volumetric semantic rendering while preserving its current modular interpretation of tissue and periwound analytics.

5. Evaluation, accuracy, and longitudinal use

The reported evaluation of Wound3DAssist spans digital wounds, silicone phantoms, and clinical wounds (Chierchia et al., 25 Aug 2025). Digital wounds are based on Syn3DWound, a synthetic dataset with 2D and 3D annotations designed to benchmark wound reconstruction and morphometry (Lebrat et al., 2023). Silicone wounds are based on SALVE, a benchmark of realistic wound phantoms captured with consumer-grade videos from an iPhone 14 Pro Max and a Logitech webcam, with ground truth obtained by Revopoint POP 3D scans (Chierchia et al., 2024). Clinical validation uses retrospective annotated images for 2D model development and prospective videos with Revopoint scans as geometric reference for a subset of cases (Chierchia et al., 25 Aug 2025).

In the clinical dataset subset with paired Revopoint scans, the framework reports reconstructed mesh accuracy of approximately 1 mm average point-wise error (AD), despite manual alignment and varying field of view (Chierchia et al., 25 Aug 2025). Errors are concentrated near wound boundaries and peripheral skin, where cropping artifacts and irregular geometry are most pronounced (Chierchia et al., 25 Aug 2025). The framework further reports that the 2D-to-3D projected segmentations consistently matched or outperformed single-view 2D segmentations for wound bed and periwound, with Wilcoxon signed-rank test $$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$3 (Chierchia et al., 25 Aug 2025). For granulation and slough, positive trends were observed, but sample sizes limited statistical significance; necrotic and epithelial were omitted from quantitative comparison due to the very low counts (Chierchia et al., 25 Aug 2025).

Longitudinal examples illustrate the framework’s clinical use. In one healing case, wound area decreased from $$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$4 to $$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$5 over approximately 3 months, while slough transitioned to granulation (Chierchia et al., 25 Aug 2025). In a non-healing case, the area increased from $$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$6 to $$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$7, with irregular tissue-composition changes suggestive of impaired healing (Chierchia et al., 25 Aug 2025). These examples are descriptive rather than statistically exhaustive, but they demonstrate that Wound3DAssist is intended not only for snapshot measurement but for trajectory-aware monitoring.

The evaluation literature around wound reconstruction provides useful context for these claims. SALVE found that photogrammetry approaches do not provide smooth surfaces suitable for precise clinical measurements of wounds, whereas neural rendering approaches show promise in addressing this issue (Chierchia et al., 2024). On realistic wound phantoms, Neus-facto and Neuralangelo achieved sub-millimeter AD and high normal consistency, though with much longer runtimes than photogrammetry (Chierchia et al., 2024). A separate RGB-D method reported sub-millimetre 3D reconstruction accuracy on silicone wound phantoms and mean surface area variability below approximately $$\min \sum_i \sum_j \| x_{ij} - \pi(K, R_i, t_i, X_j) \|^2$$8 of lesion area across repeats (Harkämper et al., 26 Jan 2026). Wound3DAssist’s reported millimeter-level clinical accuracy therefore occupies a pragmatic operating point between strict geometric fidelity and end-to-end clinical throughput (Chierchia et al., 25 Aug 2025).

6. Practical deployment, variants, and research directions

The reported runtime for Wound3DAssist is approximately 18 minutes end-to-end: initial 3D visualization requires about 8 minutes, tissue assessment about 10 minutes, and measurement extraction only seconds, with the full workflow consistently under 20 minutes (Chierchia et al., 25 Aug 2025). The evaluation hardware is an AMD EPYC 7302P with 32 cores, 540 GB RAM, and an NVIDIA RTX A6000 with 48 GB, but the core pipeline uses less than 8 GB RAM/GPU at peak and is described as viable on consumer-grade GPUs with longer runtimes (Chierchia et al., 25 Aug 2025). This reported runtime is markedly lower than many neural rendering pipelines benchmarked on SALVE, where Neus-facto required roughly 1.23 hours and Neuralangelo roughly 17.9 hours at 4K inputs (Chierchia et al., 2024).

The framework is designed around accessible devices and routine workflows. Acquisition guidance specifies an orbit around the wound at approximately 20–40 cm, multiple viewpoints, inclusion of ArUco markers, avoidance of severe motion blur, and diffuse illumination to reduce specular highlights (Chierchia et al., 25 Aug 2025). This emphasis on consumer-grade capture parallels mobile wound-segmentation work that sought smartphone-based objective therapy monitoring (Borst et al., 2024), as well as remote medical-assistance systems for pressure ulcers and postoperative surveillance (Chae et al., 2021, Shenoy et al., 2018).

Several variants and adjacent frameworks illuminate possible extensions. A fast RGB-D wound system based on Intel RealSense D435 uses RGB-D odometry and B-spline surface fitting to achieve wound assessment in tens of seconds rather than minutes, at the cost of requiring a depth sensor and a four-frame acquisition protocol (Harkämper et al., 26 Jan 2026). AiD Regen combines RGB-D capture, attention-augmented DeepLabv3 segmentation, boundary-guided point-cloud processing, 3D wound reconstruction, flattening, and G-code generation to produce intraoperative diabetic-foot-ulcer patches (Chae et al., 2022). WoundNeRF provides multi-view consistent wound bed and tissue segmentation in a neural field, with smoother boundaries and higher recall under mask perturbations than 2D or rasterization baselines (Chierchia et al., 23 Jan 2026). Each of these alternatives emphasizes a different trade-off among hardware assumptions, geometric fidelity, semantic consistency, and workflow latency.

The reported limitations of Wound3DAssist include scarce clinical 3D ground truth, device-dependent image formation, scale dependence on visible ArUco markers, sensitivity to specular highlights and shadows, and class imbalance for underrepresented tissue types such as necrotic and epithelial (Chierchia et al., 25 Aug 2025). Future improvements identified in the paper include better scale calibration through multi-marker strategies or anthropometric priors, reflectance modeling and active illumination to mitigate glare, domain generalization for device-agnostic segmentation, uncertainty quantification for per-face labels and depth, and richer periwound analytics based on standardized color/texture indices (Chierchia et al., 25 Aug 2025). Other recent work on wound assessment suggests parallel expansions toward multimodal referral prediction from images and clinical notes (Fard et al., 22 Jan 2025), hospitalization-risk modeling from image-derived wound variables and structured clinical data (Oota et al., 10 Feb 2026), and biochemical wound monitoring with swab-based multi-analyte assays and a controlled iOS imaging pipeline (Sikha et al., 30 May 2025). This suggests that future versions of Wound3DAssist may evolve from a geometry-centered framework into a broader multimodal wound-assessment platform in which 3D reconstruction, tissue semantics, periwound analysis, clinical metadata, and biochemical signals are fused into a unified longitudinal decision-support system.