TRACE: Artifact-Robust Statistical Shape Modeling from Imperfect Surface Scans - A Case Study in Craniosynostosis 3D Photography

1 Introduction

• CT-derived models are a reference for quantitative cranial morphology, but radiation limits frequent monitoring; 3D stereophotogrammetry is repeatable and radiation-free, and Bruce et al. found its metopic severity scores closely matched CT-derived scores.Prior photograph-based workflows nevertheless required extensive manual cleaning, cropping, alignment, and landmark annotation.Source: §1 · p. 2
• Raw clinical photographs contain non-head surfaces and irregular boundaries, while prior deep-learning SSMs were developed mainly for pre-segmented, artifact-free surfaces and can place correspondences on artifacts.The stated technical gap is learning anatomically meaningful head-surface correspondences directly from artifact-contaminated scans.Source: §1 · p. 2
• TRACE is an unsupervised framework that predicts sparse head-surface control points and uses thin-plate spline deformation of a clean template to reconstruct anatomy while suppressing non-head artifacts.The template constraint keeps dense correspondences on clinically relevant head anatomy despite shoulders, clothing, hands, hair, and scanner artifacts.Source: §1 · pp. 2–3
• Across PointNet, DGCNN, and Point Transformer V3, TRACE variants outperform prior SSMs and improve surface sampling, topology preservation, and shape-model quality.The variants perform similarly to one another, suggesting the gain comes primarily from the template-constrained correspondence framework rather than a specific encoder.Source: §1 · p. 3

2 Literature Review

• Craniosynostosis SSMs quantify severity against normative atlases, demographic normative spaces, or PCA-derived scores, making anatomical consistency of correspondences essential.SSMs represent population variation through dense correspondences and PCA; trustworthy models from raw 3D photographs are a prerequisite for automated photograph-based severity analysis.Source: §2 · p. 3
• Point-cloud backbones span PointNet, DGCNN, and transformer models, supporting learned correspondence and shape modeling.PointNet uses shared MLPs and max pooling, DGCNN uses dynamic nearest-neighbor graphs and edge convolution, and transformers aggregate features with attention.Source: §2 · pp. 3–4
• Prior shape-modeling approaches use learned latent representations, point-cloud correspondences, landmarks, template deformation, keypoints, and consistency learning [3, 5, 24, 15, 9, 19, 38, 4, 6].Examples include DPC, ISR points, landmark-based TPS with PCA, Mesh2SSM, SC3K, Point2SSM, and Point2SSM++ [6]; autoencoder shape representations were explored by Achlioptas et al. and Adams et al..Source: §2 · p. 4
• Prior correspondence methods are evaluated mainly on clean surfaces and may encode full-scan variation rather than head anatomy in raw artifact-contaminated photographs; TRACE constrains modeling to the head surface.This limitation motivates a template-constrained correspondence-and-deformation framework.Source: §2 · p. 4

3.1 Problem Formulation and Overview

• TRACE samples fixed-size point clouds from variable-sized clinical meshes, predicts subject-specific head landmarks, and uses them to deform a clean, topologically consistent template.The input scans may contain non-cranial structures such as shoulders and hands; the original mesh vertices remain the deformation target.Source: §3.1 · pp. 4–5

Feature initialization and Global Landmark Alignment:

• A point-cloud backbone supplies per-point features that are mean-pooled into a global descriptor, which is combined with encoded template control points to predict initial landmark displacements.This global-alignment stage captures coarse head shape and provides reference landmarks for the first SAD stage.Source: Feature initialization and Global… · pp. 5–6

Surface-Aware Deformation Stage:

• Two coarse-to-fine Surface-Aware Deformation stages project landmarks toward local surface neighborhoods, then use attention-based residual updates to refine their positions.The first stage uses a larger neighborhood to address residual scale and shape mismatch; the second uses a smaller one to sharpen localization and preserve template ordering.Source: Surface-Aware Deformation Stage · p. 6

3.3 Template Warping via Thin-Plate Spline

• TRACE fixes the TPS source control points to an artifact-free template and predicts target landmarks on the contaminated scan, warping the dense template into subject-specific head anatomy.Because the deformation depends on predicted head landmarks, non-head artifacts have minimal influence and the warped mesh preserves template topology, though large or non-uniform displacements can distort it.Source: §3.3 · p. 7

3.4 Loss Functions

• Unsupervised training combines single-directional surface-sampling losses with topology, repulsion, and sampling-consistency regularizers across the global-alignment and SAD outputs.The losses encourage landmarks and the warped template to explain head anatomy while maintaining template organization and avoiding landmark collapse or drift toward artifacts.Source: §3.4 · pp. 7–8

Dataset

• The 201-mesh CranioRate dataset was split into 151 training, 25 validation, and 25 test samples, with 5,000 uniformly sampled input vertices per mesh.Meshes were manually aligned around the head structure rather than the full scan to account for artifacts.Source: §4 · p. 9

Implementation

• TRACE and baselines used 2,048 correspondence points and the same data splits, while TRACE warped from fixed template control points to predicted target landmarks.The mean outer-head template came from CraniumPy, based on Schaufelberger et al.; the pipeline uses head-surface data without CT.Source: §4 · p. 9

Evaluation

• Compactness, generalization, and specificity were used to assess SSM quality rather than downstream clinical severity scores.The stated aim is to establish an automated shape-modeling foundation before large-scale severity scoring can be clinically validated.Source: §4 · pp. 9–10

5.1 Quantitative Results

• DG-TRACE achieved 0.29 mm P2F and 0.67 mm S2S, while TRACE variants had topology errors of 1.10–1.13 versus the best baseline value of 5.14.Point2SSM++, the strongest baseline, obtained 1.28 mm P2F and 1.90 mm S2S.Source: §5.1 · p. 10, Figure 2 · p. 11
• All TRACE variants required 11 PCA modes and achieved the lowest generalization and specificity errors; PT-TRACE had the best generalization score of 2.50, while PT-TRACE and DG-TRACE tied at 3.83 specificity.PN-AE and DG-AE used the fewest modes, but their higher generalization and specificity errors indicated a restricted or less representative shape space.Source: §5.1 · pp. 10–11, Figure 2 · p. 11

5.2 Qualitative Results

• Compared with Point2SSM++, DG-TRACE kept control points on the head, suppressed lower-surface artifacts, and preserved head surfaces and facial contours, though not fine details such as eyes and mouth.Point2SSM++ placed control points on shoulders and the lower neck, so its warped meshes preserved those artifacts and less reliably preserved fine head structures.Source: Figure 3 · p. 11, §5.2 · p. 12
• DG-TRACE’s best, median, and worst test cases had combined errors of 1.234, 2.194, and 3.135, respectively; the worst case primarily reflected reduced local correspondence organization.The cases were ranked by the sum of P2F, S2S, and topology error; even the worst case produced a coherent head reconstruction.Source: Figure 3 · p. 11, §5.2 · p. 12
• DG-TRACE’s PCA modes localized variation to head anatomy, with the first capturing global head scale and age-related change and the second facial slenderness coupled with head shape.Point2SSM++ modes mixed head and facial variation with neck, shoulder, and lower-surface artifacts.Source: Figure 4 · p. 12, §5.2 · p. 13

6 Ablation Studies

• Adding the first and second SAD stages lowers P2F from 3.22 mm to 0.39 mm and 0.29 mm, and S2S from 3.32 mm to 0.74 mm and 0.67 mm.Global-alignment-only has the lowest topology error but produces template-like shapes with poor surface sampling; the second SAD stage further sharpens landmarks and subject-specific facial structure, with a modest topology-error increase.Source: §6 · pp. 13–14, Figure 5 · p. 13
• Removing P2S raises P2F from the full-model range of roughly 0.29 mm to 2.24 mm and degrades S2S, topology, and SSM quality.The remaining losses do not anchor landmarks reliably to the head surface; Fig. 5(B) shows inconsistent landmarks, mostly underneath the target surface, and an improperly structured warp.Source: Figure 5 · p. 13, §6 · p. 14, Table 2 · p. 14
• Removing topology loss produces the largest topology error, 1.79, and worsens SSM metrics despite low surface distances.This supports the need for local template relationships to maintain anatomically ordered correspondences.Source: §6 · p. 14, Table 2 · p. 14
• Removing sampling consistency has little effect on direct surface sampling but worsens generalization and specificity, while removing warp loss has a smaller effect.The full objective balances surface anchoring, topology preservation, dense warping, and anatomical coverage.Source: §6 · pp. 14–15
• Removing repulsion yields the best reported aggregate scores but reduces coverage of peripheral anatomy, including ears, lower jaw, and neck.Landmarks can concentrate on regions that minimize nearest-surface distances; the study retains repulsion to prioritize broader head-surface coverage despite slightly worse numerical scores.Source: Figure 5 · p. 13, §6 · pp. 14–15

7 Limitations and Future Work

• Evaluation is limited to 25 of 201 photographs; larger, diverse, multi-institutional cohorts and expert-severity validation await a sufficiently large annotated multi-phenotype dataset.Clinical validation is planned as the immediate follow-up study. Bruce et al. established that 3D photography can reproduce CT-based severity scoring, while this work establishes an automated SSM foundation for scalable analysis.Source: §7 · p. 15
• The normocephalic infant template may bias TRACE toward its correspondence organization over extreme or atypical morphology; the worst-performing example in Fig. 3(B) partly indicates this limitation.Source: §7 · p. 15
• Although TRACE captures overall craniofacial geometry for SSM, it does not fully preserve fine facial details around the eyes and mouth; future work will improve local correspondence and reconstruction fidelity.Source: §7 · p. 15
• TRACE is demonstrated on craniosynostosis 3D photographs, and evaluation on additional anatomies is planned to empirically validate its anatomy-agnostic design.Source: §7 · p. 15

• TRACE predicts sparse head-surface control points, refines them through a coarse-to-fine SAD cascade, and warps a clean template with TPS to reduce the influence of non-head artifacts.The design constrains correspondence learning to anatomically meaningful head surfaces.Source: §8 · pp. 15–16
• On clinical 3D scans, TRACE consistently improves surface sampling, topology preservation, and SSM quality over prior deep learning-based SSM methods, yielding cleaner, more interpretable population shape spaces.Source: §8 · p. 16
• The correspondence module is backbone-agnostic, and a clean surface template is the only anatomy-specific element, enabling application to other anatomies and imperfect clinical scans.Source: §8 · p. 16
• TRACE provides a scalable step toward objective craniosynostosis modeling from radiation-free 3D photography and a foundation for future longitudinal and multi-institutional severity-analysis workflows.Source: §8 · p. 16

Ethics

• The study was IRB-approved (STUDY20110396) and used retrospective clinical 3D photographs deidentified and stripped of texture information to protect privacy.The data were clinically acquired.Source: Ethics · p. 16

Acknowledgments

• The authors acknowledge NIH support (NIDCR-R01DE032366), the CranioRate Consortium, and UPMC Children’s Hospital of Pittsburgh for data and clinical expertise.The authors state that the content is solely their responsibility and does not necessarily represent official NIH views.Source: Ethics · p. 16

Disclosure of Interests

• The authors declare no competing interests relevant to the article’s content.Source: Ethics · p. 16