Convolutional mesh autoencoders for the 3-dimensional identification of FGFR-related craniosynostosis

Eimear O’ Sullivan, Lara S. van de Lande, Athanasios Papaioannou, Richard W.f. Breakey, Owase Jeelani, Allan Ponniah, Christian A. Duncan, Silvia Schievano, Roman Hossein Khonsari, Stefanos Zafeiriou, David Dunaway

Abstract

Clinical diagnosis of craniofacial anomalies requires expert knowledge. Recent studies have shown that artificial intelligence (AI) based facial analysis can match the diagnostic capabilities of expert clinicians in syndrome identification. In general, these systems use 2D images and analyse texture and colour. They are powerful tools for photographic analysis but are not suitable for use with medical imaging modalities such as ultrasound, MRI or CT, and are unable to take shape information into consideration when making a diagnostic prediction. 3D morphable models (3DMMs), and their recently proposed successors, mesh autoencoders, analyse surface topography rather than texture enabling analysis from photography and all common medical imaging modalities and present an alternative to image-based analysis. We present a craniofacial analysis framework for syndrome identification using Convolutional Mesh Autoencoders (CMAs). The models were trained using 3D photographs of the general population (LSFM and LYHM), computed tomography data (CT) scans from healthy infants and patients with 3 genetically distinct craniofacial syndromes (Muenke, Crouzon, Apert). Machine diagnosis outperformed expert clinical diagnosis with an accuracy of 99.98%, sensitivity of 99.95% and specificity of 100%. The diagnostic precision of this technique supports its potential inclusion in clinical decision support systems. Its reliance on 3D topography characterisation make it suitable for AI assisted diagnosis in medical imaging as well as photographic analysis in the clinical setting.

Introduction

Diagnosing syndromic craniosynostosis (SC)—a group of genetic conditions where a baby’s skull bones fuse prematurely—is notoriously difficult. It requires highly specialized expert knowledge, and delays in diagnosis can lead to permanent neurological or visual impairment. While AI has previously been used for facial analysis, most existing systems rely on 2D images. These "texture-based" models are easily fooled by lighting or skin tone and, more importantly, cannot be used with standard medical imaging like CT, MRI, or ultrasound. Furthermore, 2D images discard the rich geometric information of the skull and face that is often the "smoking gun" for these genetic disorders. This research proposes a geometric deep learning framework to make diagnosis modality-agnostic and shape-centric.

Methods

The core of this research is the move from standard image-based CNNs to Geometric Deep Learning using Convolutional Mesh Autoencoders (CMAs). Unlike standard autoencoders that work on grids (pixels), CMAs work directly on the mesh graph.

Data Preprocessing

CT scans were converted into 3D soft-tissue meshes. To ensure consistent comparison, the team used non-rigid Iterative Closest Point (ICP) registration and Gaussian Processes to map every head to a common template. This "dense correspondence" ensures that specific vertices on one patient’s model correspond to the same anatomical points on another’s.

Architecture and Classification

The CMAs utilized spiral convolutions to learn local shape features. The autoencoder compresses the complex 3D shape into a 128-dimensional "latent vector." These vectors were then fed into a Support Vector Machine (SVM) to classify the subject as healthy or belonging to one of three syndromes: Apert, Crouzon, or Muenke. The study utilized clinical datasets from Great Ormond Street Hospital and Necker-Enfants Malades Hospital, alongside large-scale public datasets (LSFM & LYHM) to provide a normative baseline across a wide age range.

Results

The framework achieved exceptional performance in binary classification (Syndromic vs. Healthy), reaching 99.98% accuracy, 99.95% sensitivity, and 100% specificity.

t-SNE visualization of the latent space showing distinct clustering of healthy and syndromic populations.

In multi-class classification, the model effectively distinguished between the three syndromes. Visualization using t-SNE demonstrated that the syndromes naturally form distinct clusters in the model's latent space. Notably, Crouzon syndrome appeared "closest" to the healthy population, confirming its often milder physical presentation.

Confusion matrix for multi-class syndrome identification using head shape data.

One of the most significant findings was that facial shape is more diagnostic than head shape for these specific syndromes. The model even successfully identified Crouzon syndrome in a child who had been clinically judged as "unaffected" by experts, suggesting the AI can detect sub-clinical phenotypic variations invisible to the human eye.

Conclusion

By focusing purely on dysmorphology and ignoring texture, this CMA framework is applicable across multiple imaging modalities, including CT scans and 3D photography. While currently limited by the scarcity of 3D medical data for rare phenotypes, future work aims to expand the library of syndromes and explore prenatal applications via 3D ultrasound. This technology holds potential as a real-time decision support system to help primary care physicians flag at-risk infants for specialist referral.

Summary updated

Abstract

Abstract Clinical diagnosis of craniofacial anomalies requires expert knowledge. Recent studies have shown that artificial intelligence (AI) based facial analysis can match the diagnostic capabilities of expert clinicians in syndrome identification. In general, these systems use 2D images and analyse texture and colour. They are powerful tools for photographic analysis but are not suitable for use with medical imaging modalities such as ultrasound, MRI or CT, and are unable to take shape information into consideration when making a diagnostic prediction. 3D morphable models (3DMMs), and their recently proposed successors, mesh autoencoders, analyse surface topography rather than texture enabling analysis from photography and all common medical imaging modalities and present an alternative to image-based analysis. We present a craniofacial analysis framework for syndrome identification using Convolutional Mesh Autoencoders (CMAs). The models were trained using 3D photographs of the general population (LSFM and LYHM), computed tomography data (CT) scans from healthy infants and patients with 3 genetically distinct craniofacial syndromes (Muenke, Crouzon, Apert). Machine diagnosis outperformed expert clinical diagnosis with an accuracy of 99.98%, sensitivity of 99.95% and specificity of 100%. The diagnostic precision of this technique supports its potential inclusion in clinical decision support systems. Its reliance on 3D topography characterisation make it suitable for AI assisted diagnosis in medical imaging as well as photographic analysis in the clinical setting.
BibTeX
@inproceedings{sullivan2022convolutional,
  author = {Eimear O’ Sullivan and Lara S. van de Lande and Athanasios Papaioannou and Richard W.F. Breakey and Owase Jeelani and Allan Ponniah and Christian A. Duncan and Silvia Schievano and Roman Hossein Khonsari and Stefanos Zafeiriou and David Dunaway},
  title = {{Convolutional mesh autoencoders for the 3-dimensional identification of FGFR-related craniosynostosis}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-021-02411-y},
}
Figure