The Use of Machine Learning to Enable Objective Assessment of Surgical Outcomes in Sagittal Craniosynostosis.

Background: Sagittal craniosynostosis (SC) is the most common type of isolated craniosynostosis, resulting in premature fusion of the sagittal suture and a scaphocephalic head shape. Surgical interventions, such as spring-assisted cranioplasty (SAC) and open calvarial vault remodeling (CVR), are often required to normalize skull morphology and improve functional outcomes. The Swap Disentangled Variational Autoencoder (SD-VAE) was used for the objective assessment of the SC phenotype and quantification of surgical outcomes after SAC and CVR, focusing on global and regional head shape relative to a healthy population. Methods: A dataset of computed tomography, magnetic resonance imaging, and 3-dimensional photographs was used to train the SD-VAE on preoperative SC patients and healthy controls. Paired pre- and postoperative scans were analyzed to evaluate surgical outcomes. Data augmentation addressed class imbalance, and linear and quadratic discriminant analyses were used to visualize, classify, and quantify head shape morphology relative to the healthy norm. Results: For training, scans of 108 healthy children and 239 preoperative SC patients were included, with 21 CVR and 37 SAC pre- and postoperative pairs for evaluating surgical outcomes. The SD-VAE accurately distinguished SC patients from healthy controls based on global and regional morphology. Both SAC and CVR resulted in head shape normalization, with region-specific differences in effectiveness. Conclusions: By providing objective, global, and region-specific analysis of head shape and surgical outcomes, the SD-VAE enhances understanding of the SC phenotype and supports standardized outcome assessment. Future work will explore predictive applications and integration into clinical workflows for patient counseling and surgical planning.