Using A Disentangled Neural Network to Objectively Assess the Outcomes of Midfacial Surgery in Syndromic Craniosynostosis

Advancements in artificial intelligence and the development of shape models that quantify normal head shape and facial morphology provide frameworks by which the outcomes of craniofacial surgery can be compared. In this work, we will demonstrate the use of the Swap Disentangled Variational Autoencoder (SD-VAE) to objectively assess changes following midfacial surgery. Our model is trained on a dataset of 1405 3D meshes of healthy and syndromic patients which was augmented using a technique based on spectral interpolation. Patients with a diagnosis of Apert and Crouzon syndrome who had undergone sub- or trans-cranial midfacial procedures utilising rigid external distraction were then interpreted using this model as the point of comparison. A total of 56 patients met our inclusion criteria, 20 with Apert and 36 with Crouzon syndrome. By using linear discriminant analysis to project the high-dimensional vectors derived by SD-VAE onto a 2D space, the shape properties of Apert and Crouzon syndrome can be visualised in relation to the healthy population. In this way, we are able to show how surgery elicits global shape changes in each patient. To assess the regional movements achieved during surgery, we use a novel metric derived from the Malahanobis distance to quantify movements through the latent space. Objective outcome evaluation, which encourages in-depth analysis and enhances decision making, is essential for the progression of surgical practice. We have demonstrated how artificial intelligence has the ability to improve our understanding of surgery and its effect on craniofacial morphology.