🤖 AI

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

This paper introduces TRACE, an unsupervised framework that constructs robust statistical shape models directly from artifact-contaminated clinical 3D head photographs by predicting sparse anatomical correspondences and refining them through a template-constrained deformation cascade, thereby enabling radiation-free craniosynostosis analysis without the need for curated scans.

Original authors: Sanjay Bhandari, Nawazish Khan, Alzbeta Novotna, Tiffany Jeong, Loretta Bowman, Michael Hernandez, Tobi Somorin, Viraj Govani, Jesse Goldstein, Shireen Elhabian

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sanjay Bhandari, Nawazish Khan, Alzbeta Novotna, Tiffany Jeong, Loretta Bowman, Michael Hernandez, Tobi Somorin, Viraj Govani, Jesse Goldstein, Shireen Elhabian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where a doctor can understand the shape of a baby's skull without ever exposing the child to the radiation of a computerized tomography scan. For decades, the most reliable way to measure the severity of craniosynostosis—a condition where the bones of an infant's skull fuse too early, restricting brain growth and distorting the head's shape—has been to take X-ray images of the skull itself. While effective, this method carries a risk of cancer and is not suitable for the frequent monitoring that many children need. A safer, faster alternative exists in the form of 3D photography, which captures the surface of the skin rather than the bone. However, these photographs are messy. They often include the child's shoulders, hands, hair, and clothing, along with the noise of the scanner itself. For a computer to use these images to measure the head accurately, it must first learn to ignore everything that is not the head, a task that has proven difficult for existing software.

Researchers at the University of Utah and UPMC Children's Hospital of Pittsburgh have developed a new method to solve this problem, allowing computers to build accurate models of a child's head directly from these imperfect, raw photographs. Their approach, called TRACE, acts as a filter that separates the signal from the noise. Instead of trying to clean up the entire image manually, the system starts with a perfect, digital template of a normal infant head. It then looks at the messy photograph and predicts where specific points on that clean template should land on the actual child's head. By focusing only on these key points and ignoring the shoulders, hands, and hair that clutter the photo, the system can warp the clean template to match the child's unique shape. This process creates a detailed, three-dimensional map of the head that is free from the artifacts that usually confuse computer models.

The team tested this method using a dataset of 201 real-world 3D scans from patients, which included children of various ages and with different types of skull deformities. They compared their new system against several existing methods that were designed for clean, pre-sorted images. The results showed a clear advantage for the new approach. When the older methods tried to map the head from a raw photo, they often placed their measurement points on the child's shoulders or neck, mixing the anatomy of the head with the surrounding environment. This led to distorted models that did not truly reflect the skull's shape. In contrast, the new system kept its focus strictly on the head. On average, the points predicted by the new method were less than 0.3 millimeters away from the actual surface of the head, a significant improvement over the previous best methods, which were off by more than 1.2 millimeters.

Beyond just getting the points closer to the surface, the new method produced a much more reliable representation of the head's overall structure. The researchers measured how well the resulting models could describe the natural variations in head shape across a population. The models built with the new system were able to capture these variations with greater precision and were better at reconstructing unseen head shapes than models built with older techniques. This matters because the goal is to use these models to assign an objective severity score to a child's condition, which helps doctors decide on the best treatment. If the model includes the child's shoulders or clothing in its calculation, the severity score could be wrong, potentially leading to unnecessary surgery or a delay in care. The new system ensures that the score reflects only the shape of the skull.

The researchers also explored how the system works under different conditions to understand its strengths and limits. They found that the method works best when it refines its predictions in steps, first getting a rough idea of where the head is and then sharpening that focus to match the fine details of the skin. They discovered that while the system is very good at capturing the overall shape of the skull, it still struggles slightly with the very fine details around the eyes and mouth, which are complex areas to map. Furthermore, because the system relies on a template of a normal head, it may have some difficulty with extreme cases where the skull shape is very different from the average. Despite these minor limitations, the study demonstrates that it is possible to build high-quality, radiation-free models of cranial anatomy directly from messy clinical photos.

This work represents a significant step toward making the monitoring of craniosynostosis safer and more accessible. By removing the need for manual cleaning of images and the reliance on radiation-heavy scans, the new method opens the door for frequent, long-term tracking of a child's development. The researchers plan to test the system on larger groups of patients and to validate how well the computer-generated scores match the judgments of expert surgeons. If these future studies confirm the findings, the technology could become a standard tool in pediatric clinics, allowing doctors to monitor head growth with a simple camera rather than a scanner, ensuring that every child receives the care they need without unnecessary risk.

Technical Summary: TRACE – Artifact-Robust Statistical Shape Modeling from Imperfect Surface Scans

Problem Statement
Craniosynostosis, a condition involving the premature fusion of cranial sutures, requires objective severity analysis to guide surgical intervention. While Statistical Shape Models (SSMs) derived from Computed Tomography (CT) provide a reference for quantitative morphology, CT exposes infants to ionizing radiation, making it unsuitable for frequent longitudinal monitoring. Three-dimensional (3D) stereophotogrammetry offers a radiation-free alternative; however, raw clinical 3D photographs often contain significant artifacts, including shoulders, hands, clothing, hair, and scanner noise. Existing deep learning-based SSM methods are primarily designed for pre-segmented, artifact-free surfaces. When applied directly to raw clinical scans, these methods often establish correspondences on non-anatomical artifacts, corrupting the resulting shape models and limiting their clinical utility. The core technical challenge is learning anatomically meaningful head-surface correspondences directly from artifact-contaminated point clouds without extensive manual cleaning.

Methodology: The TRACE Framework
The authors propose TRACE (Template-constrained Robust Artifact-aware Correspondence Estimation), an unsupervised framework for constructing SSMs directly from raw clinical 3D head photographs. The pipeline consists of three main stages:

Global Alignment and Control Point Prediction:
- An input point cloud is encoded by a point-cloud backbone (e.g., PointNet, DGCNN, or Point Transformer V3) to generate per-point features.
- A fixed set of pre-computed template control points (derived from a clean, normocephalic head template) is encoded with positional information.
- A global-alignment MLP predicts initial displacements for these template control points, yielding a globally aligned landmark configuration. This stage captures coarse scale and global shape but does not resolve fine surface details.
Surface-Aware Deformation (SAD) Cascade:
- The framework employs a coarse-to-fine cascade of $S$ SAD stages to refine the landmarks onto the observed head surface.
- In each stage, for every reference landmark, the model gathers the $k_s$ nearest target points within a neighborhood radius $\eta_i$ .
- Using cross-attention, the model forms a neighborhood-aware query. A softmax mechanism produces "soft projection weights" to project the landmark as a weighted average of nearby target points.
- A self-attention block and an MLP then predict a residual deformation to update the landmark position.
- This process separates coarse anatomical placement from local surface localization, progressively sharpening landmark positions while maintaining template ordering.
Thin-Plate Spline (TPS) Warping:
- The final predicted landmarks are used to drive a TPS transformation that deforms a clean, dense template mesh.
- Crucially, the source of the correspondence is fixed (the clean template), while only the target landmarks are predicted from the noisy scan. This constrains the dense correspondences to clinically relevant head anatomy, effectively suppressing non-head artifacts like shoulders and clothing during reconstruction.

Loss Functions
TRACE is trained without correspondence annotations using a composite loss function evaluated in a normalized frame:

Surface-sampling losses: Point-to-surface (P2S) and warping losses (single-directional Chamfer distances) ensure landmarks lie on the target anatomy and the warped mesh recovers only the head surface.
Topology-preserving loss: Enforces edge-length consistency between predicted landmarks and the fixed template to maintain local geometric organization.
Repulsion loss: Prevents landmarks from clustering at single locations, encouraging uniform spatial coverage.
Sampling-consistency loss: Ensures the network outputs consistent landmark configurations for independent samplings of the same mesh.

Key Contributions

Unsupervised Artifact-Robust Framework: TRACE introduces a method to build SSMs directly from noisy clinical 3D photographs, eliminating the need for heavy manual cleaning, cropping, or alignment.
Template-Constrained Strategy: By predicting sparse control points on the raw scan and using them to warp a clean template, the method isolates head anatomy from acquisition artifacts (shoulders, hands, clothing).
Modular Backbone Design: The correspondence-prediction module is decoupled from the point-cloud encoder. The authors instantiated TRACE with PointNet, DGCNN, and Point Transformer V3 backbones, demonstrating that performance gains stem from the framework itself rather than a specific encoder.
Coarse-to-Fine Refinement: The SAD cascade effectively refines landmark positions from global alignment to local surface localization.

Results
Experiments were conducted on a dataset of 201 clinical 3D head meshes (151 training, 25 validation, 25 testing). The authors evaluated surface sampling, topology preservation, and SSM quality (compactness, generalization, specificity) against prior methods including Point2SSM++, Point2SSM, and various autoencoders.

Surface Sampling: TRACE variants significantly outperformed baselines. The best variant (DG-TRACE) achieved a Point-to-Face (P2F) distance of 0.29 mm and a Surface-to-Surface (S2S) distance of 0.67 mm, compared to the strongest baseline (Point2SSM++) which achieved 1.28 mm and 1.90 mm, respectively.
Topology Preservation: TRACE reduced the average topology error to 1.10, a substantial improvement over the best baseline of 5.14.
SSM Quality: All TRACE variants required 11 PCA modes to explain 95% of variation (comparable to the best baselines) but achieved significantly lower generalization and specificity errors (e.g., PT-TRACE achieved a generalization error of 2.50 vs. 8.70 for Point2SSM++). This indicates that TRACE produces population shape spaces that better reconstruct unseen anatomy and generate samples closer to realistic shapes.
Qualitative Analysis: Visualizations confirmed that while baselines often placed correspondences on shoulders and necks, TRACE successfully constrained correspondences to the head surface, reconstructing coherent cranial anatomy while suppressing artifacts.

Significance and Claims
The paper claims that TRACE establishes a necessary prerequisite for extending quantitative severity scoring to 3D photography at scale. By providing a scalable foundation for artifact-robust SSM construction, the framework enables longitudinal and multi-institutional monitoring of craniosynostosis without radiation exposure. The authors emphasize that the framework is anatomy-agnostic; while demonstrated on craniosynostosis, the design relies only on the availability of a clean surface template, suggesting applicability to other anatomical structures with imperfect scans.

The study acknowledges limitations, including the relatively small held-out test set (25 subjects) and the potential for template bias in extreme phenotypes. Future work is planned to validate the framework on larger, multi-institutional cohorts and to connect SSM quality directly to expert-defined clinical severity scores. The authors do not claim immediate clinical deployment for severity scoring but position TRACE as the foundational tool required to make such validation possible.

Technical Summary: TRACE – Artifact-Robust Statistical Shape Modeling from Imperfect Surface Scans

More like this