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

cs.CVAug 22, 2026

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

Authors: Sanjay Bhandari, Nawazish Khan, Alzbeta Novotna, Tiffany Jeong, Loretta Bowman, Michael Hernandez, Tobi Somorin, Viraj Govani,

Organizations: Scientific Computing and Imaging Institute, University of Utah, SLC, UT, USA · Kahlert School of Computing, University of Utah, Salt Lake City, UT, USA · Division of Pediatric Plastic Surgery, UPMC Children’s Hospital of Pittsburgh, PA, USA

Abstract

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.

Figures & tables

Figure 1: Overview of the Template-constrained Robust Artifact-aware Correspondence Estimation (TRACE) framework. (A) A point-cloud backbone encodes the input scan, global-alignment stage predicts initial template-control-point displacements, and two Surface-Aware Deformation (SAD) stages refine landmarks onto the observed head surface. (B) A SAD stage takes the current reference landmarks and query embeddings. For each landmark, it gathers the $k_{s}$ nearest target points and backbone features, uses cross-attention to form a neighborhood-aware query, softly projects the landmark as a weighted average of nearby target points, and applies self-attention plus an MLP to predict a residual update. (C) TPS warping deforms the dense template mesh using the transformation defined from template control points toward predicted landmarks, reconstructing head anatomy while suppressing artifacts.

Figure 2: Quantitative comparison of our modular architecture against prior SSMs on the held-out test set. We report point-to-face distance (P2F), surface-to-surface distance (S2S), and topology preservation metric (Topo), as well as the statistical shape model metrics, compactness (Comp.), generalization (Gen.), and specificity (Spec.). Lower is better for all metrics. Bold characters denote the best quantitative result for a metric.

Figure 3: Qualitative Results: (A) Qualitative comparison between Point2SSM++ and DG-TRACE. (B) Best-performing, median-performing, and worst-performing DG-TRACE test cases ranked by the sum of P2F, S2S, and topology error.

Figure 4: Shape Model: Side-view and Facial front view of the first two PCA modes of population shape variation for Point2SSM++ and DG-TRACE.

Figure 5: Qualitative ablation analysis: (A) Prediction and warping progression across the deformation cascade. (B) Effect of removing individual loss terms to show the importance of each loss term.

Method	P2F ( $mm$ ) $\downarrow$	S2S ( $mm$ ) $\downarrow$	Topo $\downarrow$	Comp. $\downarrow$	Gen. $\downarrow$	Spec. $\downarrow$
	Surface Sampling		Topology Preservation	SSM Metrics
Global-Alignment Only	3.22 ± 0.88	3.32 ± 0.86	0.76 ± 0.34	4	0.65	1.93
Global-Alignment + 1 SAD	0.39 ± 0.06	0.74 ± 0.07	1.05 ± 0.49	11	2.60	3.97
Global-Alignment + 2 SAD	0.29± 0.04	0.67 ± 0.05	1.10 ± 0.49	11	2.66	3.83

Table 1: Ablation study of the deformation cascade for DG-TRACE variant. These variants differ only in the number of refinement stages applied after the global-alignment stage.

Method	P2F ( $mm$ ) $\downarrow$	S2S ( $mm$ ) $\downarrow$	Topo $\downarrow$	Comp. $\downarrow$	Gen. $\downarrow$	Spec. $\downarrow$
	Surface Sampling		Topology Preservation	SSM Metrics
P2S zeroed	2.24 ± 0.13	1.35 ± 0.08	1.57 ± 0.59	10	2.24	3.60
Sampling zeroed	0.29 ± 0.043	0.67 ± 0.05	1.13 ± 0.50	12	2.92	4.39
Topology zeroed	0.28 ± 0.03	0.68 ± 0.05	1.79 ± 0.47	13	3.10	3.93
Warp zeroed	0.31 ± 0.04	0.69 ± 0.05	1.11 ± 0.51	11	2.65	3.74
Repulsion zeroed	0.19 ± 0.02	0.57 ± 0.03	0.79 ± 0.32	9	2.09	3.52

Table 2: Ablation study of training losses for DG-TRACE model. Each row removes one term from the full objective while keeping the same backbone, cascade, and training protocol.

Statistical Shape Modeling

Statistical shape modeling represents and analyzes population-level geometric variation to support shape comparison, reconstruction, and synthesis.

See all 23 papers

SCALP: Semi-Supervised Statistical Shape Modeling from Imperfect 3D Photogrammetry via Landmark-Anchored Spectral Warp
Jul 31, 2026Nawazish Khan, Sanjay Bhandari, Sarang Joshi +8Point Cloud Learning Statistical Shape Modeling
Scientific Computing and Imaging Institute, University of Utah, Salt Lake City, UT 84112, USA · Kahlert School of Computing, University of Utah, Salt Lake City, UT 84112, USA · Department of Biomedical Engineering, University of Utah, Salt Lake City, UT 84112, USA +1
Articulating then Matching: Zero-Shot Shape Matching for Uncurated Data
Jun 28, 2026Qilong Liu, Qinfeng Xiao, Chenyuan Yi +2Non-Rigid Shape Correspondence 3D Shape Matching
Hong Kong Polytechnic University, HK SAR · Artificial Intelligence in Design, HK SAR
Non-Rigid 3D Shape Correspondences: From Foundations to Open Challenges and Opportunities
Apr 1, 2026Aleksei Zhuravlev, Lennart Bastian, Dongliang Cao +12Functional Maps Non-Rigid Shape Correspondence
MPI for Informatics · Imperial College London · Technical University of Munich +5

Jul 31, 2026J/K move · Enter open · S save

SCALP: Semi-Supervised Statistical Shape Modeling from Imperfect 3D Photogrammetry via Landmark-Anchored Spectral Warp

Nawazish Khan, Sanjay Bhandari, Sarang Joshi, Alzbeta Novotna+7

Scientific Computing and Imaging Institute, University of Utah, Salt Lake City, UT 84112, USA · Kahlert School of Computing, University of Utah, Salt Lake City, UT 84112, USA · Department of Biomedical Engineering, University of Utah, Salt Lake City, UT 84112, USA+1

Point Cloud Learning Statistical Shape Modeling Non-Rigid Shape Correspondence Semi-Supervised Learning

Open paper AskPDF arXiv

Correspondence-based statistical shape modeling (SSM) is vital for population-level morphometric analysis, but conventional pipelines assume clean, fully registered surfaces. Real-world clinical photogrammetry scans are often noisy, partial, and cluttered, hindering the adoption of radiation-free surface imaging as a safe alternative to computed tomography (CT) for infant craniosynostosis. We present SCALP (Semi-supervised Correspondence via lAndmark Localization and sPectral warping), a two-stage framework that constructs consistent shape models directly from raw, imperfect surface scans. First, a semi-supervised Point Transformer leverages a small expert-annotated dataset alongside a large unlabeled cohort to accurately localize craniofacial landmarks with minimal annotation overhead. Second, these landmarks anchor a Laplace--Beltrami spectral deformation of an anatomical template, generating dense correspondences while naturally isolating the cranium from peripheral scanning clutter without manual preprocessing. Experiments on infant photogrammetry scans demonstrate that SCALP consistently outperforms state-of-the-art unsupervised point-cloud approaches, offering a clinically practical pathway toward objective, radiation-free head shape analysis.

Abstract

Statistical Shape Modeling

Explore similar work

SCALP: Semi-Supervised Statistical Shape Modeling from Imperfect 3D Photogrammetry via Landmark-Anchored Spectral Warp

Articulating then Matching: Zero-Shot Shape Matching for Uncurated Data

Non-Rigid 3D Shape Correspondences: From Foundations to Open Challenges and Opportunities