Abstract
Prediction of anatomical development plays a crucial role in the selection and planning of pediatric surgical treatments. However, the rapid growth of young patients and the potential involvement of pathology makes prediction of anatomical changes challenging. We present a novel deep learning architecture to make personalized predictions of normative and pathologic head development using only cross-sectional data. We designed a novel phenotype encoder that uses domain adversarial training to create age- and sex-independent representations of patient phenotypes and growth predictor that learns to generate the head shape of patients given the anatomical effects of age, sex and pathology. These modules cooperate to instantiate patient anatomies to any age, enabling personalized predictions conditioned to the presence of pathology. We trained our model using head segmentations from cross-sectional CT images and 3D photograms and evaluated model performance on an independent longitudinal dataset. The proposed model achieved a head surface growth prediction error of 4.93 ± 2.29 mm and a volumetric error 0.16 ± 0.11 L in patients with cranial pathology, and 4.61 ± 3.28 mm and 0.27 ± 0.19 L for normative subjects, demonstrating high accuracy. Our method is the first to create age- and sex-agnostic patient phenotype representations, and to enable personalized predictions of pathological development without requiring longitudinal data for training.