A Deep Generative Model of Neonatal Cortical Surface Development

Abstract

The neonatal cortical surface is known to be affected by preterm birth, and the subsequent changes to cortical organisation have been associated with poorer neurodevelopmental outcomes. Deep Generative models have the potential to lead to clinically interpretable models of disease, but developing these on the cortical surface is challenging since established techniques for learning convolutional filters are inappropriate on non-flat topologies. To close this gap, we implement a surfacebased CycleGAN using mixture model CNNs (MoNet) to translate sphericalised neonatal cortical surface features (curvature and T1w/T2w cortical myelin) between different stages of cortical maturity. Results show our method is able to reliably predict changes in individual patterns of cortical organisation at later stages of gestation, validated by comparison to longitudinal data; and translate appearance between preterm and term gestation (> 37 weeks gestation), validated through comparison with a trained term/preterm classifier. Simulated differences in cortical maturation are consistent with observations in the literature.

Surface Generative Modelling of Neurodevelopmental Trajectories

“…Prior to running all experiments we ran ablation analyses, on the task of simulating healthy cortical neurodevelopment. Experimental design was validated against a baseline CycleGAN, trained to translate examples between two discrete classes: term (PMA ≥ 37 weeks) and preterm (PMA < 37 weeks) [40]. Model parameters were then evaluated by investigating the impact of changing training cycle length (from n = [2, 3, 4, 5]), and evaluating how well the model performed when trained to learn images directly instead of age difference maps - to test the hypothesis that learning difference maps improves subject specificity (run for the n=3 cycle only).…”

“…In this paper we propose an extension to our preliminary work [40], which trained a CycleGAN to translate cortical appearance of preterm scans to appear like healthy term controls. As the model was conditioned purely on discrete classes, it was was not able to continuously age brains.…”