AI-Driven Three-Dimensional Reconstruction and Quantitative Analysis for Burn Injury Assessment

Accurate, reproducible burn assessment is critical for treatment planning, healing monitoring, and medico-legal documentation, yet conventional visual inspection and 2D photography are subjective and limited for longitudinal comparison. This paper presents an AI-enabled burn assessment and management platform that integrates multi-view photogrammetry, 3D surface reconstruction, and deep learning-based segmentation within a structured clinical workflow. Using standard multi-angle images from consumer-grade cameras, the system reconstructs patient-specific 3D burn surfaces and maps burn regions onto anatomy to compute objective metrics in real-world units, including surface area, TBSA, depth-related geometric proxies, and volumetric change. Successive reconstructions are spatially aligned to quantify healing progression over time, enabling objective tracking of wound contraction and depth reduction. The platform also supports structured patient intake, guided image capture, 3D analysis and visualization, treatment recommendations, and automated report generation. Simulation-based evaluation demonstrates stable reconstructions, consistent metric computation, and clinically plausible longitudinal trends, supporting a scalable, non-invasive approach to objective, geometry-aware burn assessment and decision support in acute and outpatient care.

💡 Research Summary

The paper addresses the long‑standing problem of subjective and inconsistent burn assessment, which hampers fluid resuscitation planning, surgical decision‑making, and medico‑legal documentation. Conventional visual inspection and two‑dimensional (2D) photography suffer from perspective distortion, observer bias, and difficulty in longitudinal comparison. To overcome these limitations, the authors present an end‑to‑end, AI‑enabled burn management platform that combines a guideline‑driven clinical workflow with a fully automated three‑dimensional (3D) reconstruction and quantitative analysis pipeline using only consumer‑grade cameras (smartphones or digital cameras).

The system is organized into several modules. The clinical workflow module digitizes primary and secondary surveys, captures patient demographics, injury mechanism, and burn descriptors, and stores them in a structured, time‑stamped database. It also includes a fluid‑resuscitation calculator based on the Parkland formula, automatically deriving total and phase‑specific fluid volumes from weight and %TBSA.

Image acquisition accepts either a set of at least six overlapping photographs taken around the wound or a short video from which up to fifteen frames are extracted. Prior to reconstruction, the software validates image quality by checking resolution, blur (Laplacian variance), and exposure, discarding unsuitable frames.

For 3D reconstruction, the pipeline employs Scale‑Invariant Feature Transform (SIFT) to detect and match keypoints across views, followed by Structure‑from‑Motion (SfM) using COLMAP to estimate camera intrinsics, extrinsics, and a sparse point cloud. Bundle adjustment refines these parameters by minimizing robust reprojection error. Multi‑View Stereo (MVS) then densifies the point cloud and generates a triangulated mesh. Because photogrammetric reconstructions are defined only up to an unknown similarity transform, metric scaling is achieved by placing a known‑length calibration object (e.g., a ruler) in the scene; the scale factor s = dr/dm is computed and applied to all vertices, converting areas and volumes to real‑world units (cm², mm³).

Burn segmentation is performed on each 2D image using deep convolutional networks such as U‑Net or Mask RCNN. The loss combines Dice loss and binary cross‑entropy to handle class imbalance. The resulting binary masks are back‑projected onto the 3D mesh by intersecting camera rays with the surface; a probabilistic fusion across views yields a robust 3D burn label. Surface area is calculated by summing the areas of all burned triangles (A = ∑½‖(V₂‑V₁)×(V₃‑V₁)‖), perimeter by summing the lengths of boundary edges, and depth/volume proxies are derived from surface curvature or by measuring volumetric differences between successive reconstructions.

Longitudinal tracking is enabled by aligning successive 3D models using Iterative Closest Point (ICP) for rigid registration, with optional non‑rigid refinement to accommodate soft‑tissue deformation. After registration, changes in area, perimeter, and volume directly reflect wound contraction and depth reduction, providing clinicians with objective healing trajectories.

The authors evaluate the platform through simulation studies that vary lighting, background, and skin tone. Results show reconstruction root‑mean‑square error below 2 mm, surface‑area error within ±3 %, and registration error under 1 mm, indicating sufficient precision for clinical use. The entire pipeline processes a typical case in 5–10 minutes on a standard workstation, making it feasible for emergency departments, outpatient burn clinics, and tele‑medicine settings.

Key contributions include: (1) elimination of 2D‑based subjectivity by delivering metric‑scaled 3D measurements; (2) a fully consumer‑device workflow that requires no specialized hardware; (3) integration of AI segmentation with 3D geometry to produce explainable, physically meaningful metrics; (4) built‑in longitudinal registration for objective monitoring of healing; and (5) automated report generation that embeds clinical data, calculated metrics, and visualizations for documentation and legal purposes.

Limitations are acknowledged: depth estimation relies on RGB‑only proxies, so true tissue depth cannot be measured without additional modalities such as ultrasound, multispectral imaging, or depth sensors. Future work will focus on multimodal data fusion to improve depth and volume accuracy, real‑time mobile implementation, and large‑scale clinical validation across diverse patient populations.