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 rare genetic disorders like Syndromic Craniosynostosis (SC)—which includes Muenke, Crouzon, and Apert syndromes—is a high-stakes clinical challenge. These conditions involve the premature fusion of skull sutures, which can lead to severe complications such as vision loss, respiratory issues, and cognitive defects if not addressed early. While AI has made strides in identifying facial phenotypes using 2D photography, these systems are often limited by their reliance on texture and color, discarding rich 3D structural information and complicating integration with standard medical imaging like CT or MRI. This study introduces a shape-based diagnostic tool designed to provide objective, high-precision screening across various 3D imaging modalities.
Methods
The framework utilizes Convolutional Mesh Autoencoders (CMAs), which operate directly on 3D meshes rather than pixel grids. The process involves:
- Dense Correspondence: CT scans and 3D photos are converted into meshes and re-parameterized so that every mesh shares the same number of vertices in identical semantic locations.
- Spiral Convolutions: The CMA architecture employs spiral convolutions to learn local geometric features, compressing high-dimensional meshes into a low-dimensional "latent vector"—a unique numerical fingerprint of head shape.
- Classification: These latent vectors are processed by a Support Vector Machine (SVM) to classify subjects as healthy or syndromic.
- Regional Analysis: The researchers evaluated three models—face-only, head-only, and a combined head-and-face model—to determine the most diagnostically significant regions.
Datasets
The study leveraged a diverse dataset comprising 3D data from four primary sources, ensuring the model could distinguish between normal age-related growth and syndromic malformations.
The dataset included CT data from 122 patients with Apert, Crouzon, and Muenke syndromes, alongside control groups of healthy infants and adults from the Large Scale Face Model (LSFM) and Liverpool-York Head Model (LYHM).
Key Results
The CMA framework demonstrated exceptional quantitative performance, particularly in binary classification between healthy and syndromic individuals.
The face-only model achieved a staggering 99.98% accuracy. In multi-class classification (distinguishing between specific syndromes), the models maintained high accuracies: 98.3% for the face, 97.9% for the head, and 98.2% for the combined model. Furthermore, the models reconstructed 3D faces with a mean error of only 1.4mm, proving their anatomical fidelity.
Discussion and Future Work
A key insight from this research is that 3D topography provides a more robust diagnostic signal than 2D texture, as it is unaffected by lighting or skin tone. The model even demonstrated the ability to detect "clinically undetectable" variations; in one case, it correctly identified Crouzon syndrome in a child previously deemed unaffected by experts, a finding later confirmed by genetic testing.
While the results are groundbreaking, the study notes limitations such as data scarcity for certain syndromes and the need for greater demographic diversity. Future applications could include integration into 3D ultrasound for prenatal diagnosis or deployment on mobile devices with 3D cameras for accessible primary care screening.
Summary updated