The 3D skull 0–4 years: A validated, generative, statistical shape model

Eimear O' Sullivan, Lara S. van de Lande, Anne-Jet C. Oosting, Athanasios Papaioannou, Owase Jeelani, Maarten J. Koudstaal, Roman Hossein Khonsari, David Dunaway, Stefanos Zafeiriou, Silvia SchievanoBone 2021 · DOI

TL;DR

This paper introduces a novel 3D statistical shape model (3DMM) of the pediatric skull (0-4 years old), addressing the scarcity of normative data for this age group. This model not only accurately represents existing skull shapes but can also generate realistic synthetic skulls, paving the way for improved in silico medical applications, surgical planning, and device development.

What Problem Does This Paper Address?

Developing accurate computational models for medical applications, especially in pediatrics, is often hampered by a lack of sufficient, high-quality data. This is particularly true for the developing pediatric skull, where ethical considerations, the risks of radiation from CT scans, and the rapid pace of growth make collecting large datasets challenging. Existing methods often rely on 2D measurements, which fail to capture the full complexity of 3D skull anatomy. This paper aims to bridge this gap by creating a robust, generative 3D statistical shape model of the pediatric skull (0-4 years) that can serve as a valuable resource for research and clinical applications.

Methods

The researchers developed a 3D statistical shape model (3DMM) using a dataset of 178 pediatric CT scans. The process involved:

  • 3D Skull Reconstruction: CT scans were processed to create 3D meshes of the skull.
  • Dense Correspondence: All skull meshes were aligned to a common template using a non-rigid iterative closest point algorithm, ensuring that corresponding points across different skulls represented the same anatomical features.
  • Principal Component Analysis (PCA): PCA was applied to the aligned meshes to identify the principal modes of variation in skull shape within the population. This allowed for the creation of a compact model that captures the majority of the shape variance with a limited number of components.
  • Validation: The model was rigorously validated through:
  • Intrinsic Validation: Assessing the model's compactness (how well shape variance is captured by principal components), generalisation error (ability to represent unseen data), and specificity (realism of generated synthetic shapes).
  • Extrinsic Validation: Comparing anthropometric measurements extracted from the 3DMM (both the mean shape and synthetic samples) against established normative data from literature (Waitzman et al., 1992) and the original study population.

Datasets

  • Training and Evaluation Dataset: 178 CT scans of healthy children aged 0-48 months (mean age 20.8 ± 12.9 months).
  • Normative Data for Comparison: Anthropometric measurements from Waitzman et al. (1992), a widely recognized source for craniofacial skeletal measurements in children.

Key Insights and Ideas

  • Generative Power for Scarce Data: The core innovation is the creation of a generative model. This means it can produce new, realistic skull shapes, effectively overcoming the limitations imposed by the scarcity of real patient data for research and development.
  • Capturing Developmental Nuances: The model is specifically designed for the 0-4 year age range, a period of rapid skull development. This makes it particularly relevant for understanding and addressing early craniofacial anomalies.
  • Bridging 2D and 3D: By validating against established 2D anthropometric measures, the study demonstrates how a complex 3D model can align with and extend traditional clinical measurements, offering a more comprehensive view.
  • Age Predictability: The model shows a good ability to predict a child's age from their skull shape, highlighting the strong correlation between skeletal development and morphology in this age group.

Key Results

  • Compact Model: 90% of the shape variance in the pediatric skull population (0-4 years) was captured by the first 10 principal components, indicating a compact and efficient model.
  • Accurate Generalisation: The generalisation error was as low as 0.47 mm when using all model components, demonstrating the model's ability to accurately represent skull shapes not included in the training set.
  • Realistic Synthetic Shapes: Specificity values below 0.7 mm confirmed that the synthetic skull instances generated by the model are realistic.
  • Good Agreement with Normative Data: Anthropometric measures extracted from the 3DMM mean shape and synthetic samples showed good agreement with the study population and published normative data, with minor deviations attributed to landmark localization consistency.
  • Age Prediction Accuracy: A root mean square error of 6.1 months and an R² score of 0.77 were achieved when predicting age from skull shape, indicating reasonable accuracy.
  • Limited Gender Predictability: Gender prediction accuracy was only slightly better than chance (60.1%), suggesting skull shape alone is not a strong indicator of gender in this age group.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that smaller standard deviations were observed for generated samples, likely due to the omission of size effects from the model.
  • Gender prediction accuracy was limited, suggesting skull shape is not a primary differentiator in this age group.
  • Future Work:
    • The authors initiated the construction of a mandible model and aim to combine it with soft tissue 3DMMs of the face.
  • This combined model could lead to more refined surgical planning tools, advanced craniofacial diagnosis algorithms, and better postoperative outcome assessments.
  • The normative 3DMM can be used to evaluate normal and abnormal skull shapes for diagnostic purposes and to potentially plan personalized surgical treatments for craniofacial syndromes.
  • The model can also support the design of surgical devices and implants, and potentially inform the development of safety equipment like helmets.

Summary updated

Abstract

This study presents a reliable statistical shape model of the paediatric skull 0-4 years that adheres to known skull morphometric measures, can accurately represent unseen skull samples not used during model construction and can generate novel realistic skull instances, thus presenting a solution to limited availability of normative data in this field.
BibTeX
@inproceedings{sullivan2021d,
  author = {Eimear O' Sullivan and Lara S. van de Lande and Anne-Jet C. Oosting and Athanasios Papaioannou and Owase Jeelani and Maarten J. Koudstaal and Roman Hossein Khonsari and David Dunaway and Stefanos Zafeiriou and Silvia Schievano},
  title = {{The 3D skull 0–4 years: A validated, generative, statistical shape model}},
  booktitle = {Bone},
  year = {2021},
  doi = {10.1016/j.bonr.2021.101154},
}
Figure