Source-linked AI summary
TRACE: Artifact-Robust Statistical Shape Modeling from Imperfect Surface Scans - A Case Study in Craniosynostosis 3D Photography
Sanjay Bhandari, Nawazish Khan, Alzbeta Novotna, Tiffany Jeong, Loretta Bowman, Michael Hernandez, Tobi Somorin, Viraj Govani, Jesse Goldstein, Shireen Elhabian
TL;DR
Craniosynostosis shape analysis needs scalable correspondence estimation from raw 3D photographs, which are repeatable but contaminated by artifacts and often require manual curation. TRACE uses sparse head control points, staged deformation, and thin-plate spline warping of a clean template to construct anatomy-focused SSMs. Across multiple backbones, it improves surface sampling, topology preservation, and SSM quality over prior methods, while remaining limited by a small held-out test set and the absence of clinical severity validation.
Problem
Raw clinical 3D photographs offer a radiation-free, repeatable modality, but artifacts and incomplete boundaries corrupt correspondences while prior workflows require extensive manual preprocessing.
Method
TRACE predicts sparse head-surface control points, refines them through Surface-Aware Deformation stages, and uses thin-plate spline warping to deform a clean template mesh.
Results
TRACE variants consistently improve surface sampling, topology preservation, and statistical shape-model quality over prior deep learning-based SSM methods across PointNet, DGCNN, and Point Transformer V3 backbones.
Takeaways & Limitations
TRACE establishes a scalable foundation for objective craniosynostosis shape modeling from radiation-free 3D photography and may extend to other anatomies when a clean template is available.
Takeaways & Limitations
Evaluation uses 25 test subjects from a 201-photograph dataset, and clinical validation against expert-defined severity measures remains future work.
Abstract
from arXiv · showhide
Craniosynostosis severity analysis increasingly relies on statistical shape models (SSMs) to quantify cranial morphology, but most existing workflows depend on computed tomography or heavily curated three-dimensional (3D) photographs. Raw clinical 3D photographs provide a radiation-free and repeatable alternative, yet often contain shoulders, hands, hair, clothing, scanner noise, and incomplete boundaries that corrupt correspondences. We introduce the Template-constrained Robust Artifact-aware Correspondence Estimation (TRACE) framework, an unsupervised method for constructing SSMs directly from artifact-contaminated clinical 3D head photographs. TRACE predicts sparse anatomically corresponding head-surface control points from the raw point cloud, refines them through a coarse-to-fine Surface-Aware Deformation cascade, and uses thin-plate spline warping to deform a clean template mesh into a subject-specific head reconstruction. This template-constrained formulation keeps dense correspondences on clinically relevant head anatomy while suppressing non-head artifacts. The correspondence module is decoupled from the point-cloud encoder, enabling the same deformation pipeline to be paired with different backbones, including PointNet, DGCNN, and Point Transformer V3. Across all backbones, TRACE substantially improves surface sampling, topology preservation, and shape-model quality over prior SSM methods, providing a scalable foundation for photograph-based craniosynostosis shape analysis and a framework that may extend to other artifact-contaminated surface scans when an appropriate clean template is available.
1 Introduction
Craniosynostosis severity assessment needs reproducible 3D shape analysis, but raw clinical photographs contain artifacts that disrupt anatomical correspondences. TRACE addresses this gap by constraining correspondence estimation to head anatomy through a clean-template deformation framework.
- Motivation: Craniosynostosis severity guides surgical indication, timing, technique, and outcome assessment, motivating objective and reproducible 3D shape-based scoring.SSM-derived severity scores have been validated against expert surgeon assessments and correlate strongly with clinical severity ratings.
- Motivation: 3D stereophotogrammetry is radiation-free, repeatable, and suitable for longitudinal or population-scale monitoring, but prior workflows required extensive manual preprocessing.The earlier workflow involved cleaning, cropping, alignment, and landmark annotation before correspondence estimation.
- Technical gap: Raw clinical photographs contain shoulders, necks, clothing, hands, hair, and irregular boundaries that can place correspondence points on non-head artifacts.Existing deep learning SSMs were developed mainly for pre-segmented, artifact-free surfaces.
- TRACE framework: TRACE predicts sparse head-surface control points from artifact-contaminated photographs and uses them to drive thin-plate spline deformation of a clean head template.Its template-constrained formulation avoids heavy manual cleaning while reconstructing subject-specific head anatomy.
- TRACE framework: TRACE uses the same correspondence-and-deformation framework with PointNet, DGCNN, and Point Transformer V3 backbones, with gains attributed primarily to the framework.All three variants substantially outperform evaluated prior SSMs and perform similarly to one another.
- Results: TRACE produces more accurate surface sampling, better topology preservation, and higher-quality statistical shape models than prior deep learning-based SSMs.These improvements establish a foundation for scalable photograph-based craniosynostosis shape analysis.
2 Literature Review
Statistical shape models represent anatomical variation through dense correspondences and PCA, supporting craniosynostosis severity quantification. Existing learned correspondence methods generally assume clean surfaces, whereas TRACE restricts modeling to head anatomy in raw artifact-contaminated photographs.
- Statistical shape modeling: SSMs establish dense population correspondences and apply PCA to produce compact, interpretable representations of anatomical variation.Craniosynostosis applications include normative atlas comparison, demographic projection, and severity scores from PCA modes.
- Deep point-cloud methods: PointNet, DGCNN, and Transformer-based encoders provide distinct approaches to permutation-invariant point-cloud representation learning and correspondence estimation.Their mechanisms include shared multilayer perceptrons, dynamic nearest-neighbor graphs, edge convolution, and attention-based aggregation.
- Research gap: Existing methods are evaluated mainly on pre-segmented, artifact-free surfaces and can encode full-scan variation rather than isolating head anatomy in raw photographs.TRACE addresses this limitation with a template-constrained correspondence-and-deformation framework.
3 Method
TRACE predicts sparse head-surface correspondences from artifact-contaminated point clouds, refines them coarse-to-fine, and warps a clean template mesh to reconstruct subject-specific anatomy.
- Problem formulation: TRACE samples fixed-size point clouds from variable-vertex clinical meshes while retaining meshes as deformation targets.The method treats control points, landmarks, and correspondence points as interchangeable sparse anatomical representations.
- Global landmark alignment: Global alignment predicts initial displacements for template control points, capturing coarse scale and global head shape.Each aligned landmark is computed as aj = cj + Δcj.
- Surface-aware deformation: Each SAD stage combines neighborhood-aware cross-attention, soft projection onto nearby target points, self-attention, and residual deformation prediction.Soft projection represents each landmark as a convex combination of neighboring target points before residual refinement.
- Surface-aware deformation: Two Surface-Aware Deformation stages progressively refine landmarks from coarse anatomical placement to local surface localization.The first stage uses larger neighborhoods, while the second sharpens localization with smaller neighborhoods and preserves template ordering.
- Template warping: TPS warping deforms a clean template from fixed control points toward predicted correspondences, preserving topology while suppressing non-head artifacts.Large or highly non-uniform landmark displacements can still produce local stretching and distortion where control-point support is sparse.
- Loss functions: TRACE is trained without correspondence annotations using point-to-surface, warping, topology, repulsion, and sampling-consistency losses across prediction stages.Topology loss maintains local edge-length organization, repulsion discourages landmark collapse, and sampling consistency enforces stable predictions across independent samplings.
4 Experiments
Experiments use 201 clinical 3D surface meshes and assess correspondence quality through established statistical shape-model criteria rather than downstream clinical severity scores.
- Dataset: The dataset contains 201 3D photograph-derived surface meshes spanning ages, treatment statuses, and craniofacial phenotypes.Each mesh contributes 5000 uniformly sampled vertices as the input point cloud.
- Template and implementation: TRACE uses a normocephalic infant outer-head-surface template with 2048 pre-computed control points and no CT dependency.The same control-point count is predicted for target surfaces during correspondence estimation and TPS warping.
- Implementation: The two SAD stages use neighborhood-radius settings (ks = 64, ρs = 0.36) and (ks = 16, ρs = 0.09), respectively.Training runs for 200 epochs with batch size 4 and initial learning rate 10^-3.
- Evaluation: Evaluation uses compactness, generalization, and specificity as established statistical shape-model quality criteria.These metrics are selected because the paper focuses on automated SSM construction before clinical severity scoring is validated at scale.
5 Results
TRACE variants substantially improve surface sampling, topology preservation, and statistical shape-model quality over prior methods, with DG-TRACE delivering the lowest average surface errors. Qualitative results further show artifact suppression and more anatomically interpretable population shape variation.
- Quantitative Results: TRACE variants use P2F, S2S, topology, compactness, generalization, and specificity to evaluate correspondence and shape-model quality against prior SSMs.P2F measures landmark-to-surface distance, S2S measures warped-template-to-surface distance, and lower values are better for all reported metrics.
- Quantitative Results: 0.29 mm P2F and 0.67 mm S2S are DG-TRACE’s lowest average surface errors, compared with 1.28 mm and 1.90 mm for Point2SSM++.DG-TRACE also achieves a topology error of 1.10, versus the best baseline value of 5.14.
- Quantitative Results: 2.50 is PT-TRACE’s best generalization score, while PT-TRACE and DG-TRACE share the best specificity score of 3.83.All TRACE variants require 11 PCA modes and achieve the lowest generalization and specificity errors.
- Qualitative Results: DG-TRACE keeps predicted control points on the head surface, suppressing shoulder and lower-neck artifacts that Point2SSM++ preserves in warped meshes.The resulting reconstructions preserve head surfaces and facial contours, though fine structures such as the eyes and mouth are not fully preserved.
- Ablation Study: Table 1 ablates the global-alignment stage and two Surface-Aware Deformation refinement stages in DG-TRACE.The study compares variants differing only in the number of refinement stages after global alignment.
- Qualitative Results: DG-TRACE’s first PCA mode reflects global head scale and age-related change, while its second captures facial slenderness coupled with head-shape variation.Point2SSM++ modes mix head and facial variation with neck, shoulder, and lower-surface artifacts, reducing anatomical interpretability.
6 Ablation Studies
The ablations identify the coarse-to-fine deformation cascade and key loss terms as central to TRACE’s sampling, topology, coverage, and warped-mesh quality.
- Deformation cascade: The global-alignment-only model preserves template topology but produces poor surface sampling, with P2F 3.22 mm and S2S 3.32 mm.Its reconstructions remain template-like rather than subject-specific.
- Deformation cascade: Adding the first SAD stage reduces P2F to 0.39 mm and S2S to 0.74 mm by moving landmarks toward the observed surface.
- Deformation cascade: The second SAD stage further reduces P2F to 0.29 mm and S2S to 0.67 mm, sharpening landmark locations and facial structure.Smaller local neighborhoods refine the first-stage predictions.
- Loss terms: Removing the repulsion term can improve aggregate metrics for central cranial coverage, but the full objective better supports broader anatomy including ears, jaw, and neck.The authors retain repulsion despite slightly worse numerical scores to prioritize peripheral head-surface coverage.
- Loss terms: Overall, SAD refinement and P2S anchoring drive the main gains, while sampling, repulsion, and warping terms regulate stability, coverage, and dense mesh quality.
7 Limitations and Future Work
The study’s held-out evaluation is limited to 25 test subjects, and future work will assess larger, more diverse cohorts and clinical severity validation.
- The held-out test set contains 25 subjects from a dataset of 201 clinical 3D photographs because of raw-scan constraints.
- Future work will evaluate TRACE on larger, more diverse, multi-institutional cohorts.
- Clinical validation against expert-defined severity measures awaits a sufficiently large annotated multi-phenotype dataset.
- The normocephalic infant template may bias correspondence organization for phenotypes that differ substantially from the template.Extreme or atypical morphology may therefore be represented less accurately.
8 Conclusion
TRACE constructs head-anatomy SSMs from raw clinical 3D photographs by combining sparse control points, SAD refinement, and TPS warping. It improves key SSM properties over prior methods and provides a foundation for broader clinical and anatomical applications.
- TRACE predicts sparse head-surface control points, refines them through a coarse-to-fine SAD cascade, and warps a clean template with TPS.
- The template-constrained design limits correspondence learning to meaningful head surfaces while reducing influence from shoulders, hands, clothing, hair, and scanner artifacts.
- Experiments show consistent improvements in surface sampling, topology preservation, and SSM quality compared with prior deep learning-based SSM methods.
- TRACE variants outperform prior SSM methods across different point encoders and produce cleaner, more interpretable population shape spaces.
- Because the template is the only anatomy-specific element, the framework may apply to other anatomies with a clean surface template.