Development of deep learning methods to optimize patient personalized treatment for craniosynostosis

Project Summary/Abstract Craniosynostosis is the premature fusion of one or more cranial sutures. The growth at the cranial plates normally separated by the fused suture is arrested, resulting in compensatory overgrowth parallel to the fused suture. This abnormal development causes head malformations and can lead to increased intracranial pressure

and developmental complications. Patients with this condition normally undergo surgical treatment to remove the growth constraints and create more aesthetically normative head shapes. However, the traditional evaluation of these patients after treatment has been based on subjective clinical expertise. During the last decade, three-

dimensional (3D) photogrammetry has gained popularity to evaluate craniofacial anomalies, but existing methods to analyze this data have failed in their clinical translation due to the use of inefficient and inaccurate processing techniques. Hence, there is a lack of quantitative evidence to evaluate and compare the continuous

process of head shape normalization between different surgical treatments. This proposal aims to develop new geometric deep learning methods to enable the fully automated, real-time evaluation of head shapes using 3D photogrammetry at the clinic and will address the lack of quantitative evidence in the objective assessment of

surgical outcomes. The first aim of this proposal is to characterize the normalization of head shape after corrective surgery for different treatment approaches. A statistical model of head shape normalization for each surgical approach will be created to quantify novel metrics of head shape anomaly and probabilistic risk of

craniosynostosis. This model will be the first to incorporate the effects of age at surgery, sex, and the pre-surgical severity of head shape anomalies on the normalization of head shape. The second aim of this proposal is to develop a personalized geometric deep learning framework to determine the optimal surgical approach for each

patient with craniosynostosis. This aim will incorporate a novel context-encoding geometric deep learning method to estimate the expected post-surgical head shape normalization for each potential surgical treatment and identify the optimal surgical treatment for every patient based on objective retrospective data. The results

from these aims will enable the data-driven, personalized, and objective assessment of surgical treatment for craniosynostosis that can be used to optimize patient management. This proposal includes a comprehensive training plan consisting of mentored computational training in the development of deep learning methods,

mentored clinical and translational research training at Children’s Hospital Colorado, and didactic coursework in statistical and machine learning methods. This proposal is uniquely positioned to be successful given the multidisciplinary environment at the University of Colorado Anschutz Medical Campus with resources in

translational research at Children’s Hospital Colorado and machine learning expertise within the Medical Image Phenotyping lab through the Department of Biostatistics and Informatics.