1st Supervisor: Jonathan O’Muircheartaigh, King’s College London
2nd Supervisor: Emma Robinson, King’s College London
Clinical Champion: David Edwards, King’s College London
Aim of the PhD Project:
- Use novel deep generative modelling to describe the range and variation of cortical surface anatomy in infants and children
- Investigate the sensitivity of these methods to known perinatal and childhood neurological disease
Project Description / Background:
Common neurodevelopmental disorders such as epilepsies or autism, have diverse causes, ranging from genetics to environmental stressors such as premature birth. However, what they all have in common, is that they alter the complex processes of typical brain development.
Subtle changes to patterns of brain myelination, cortical folding, and cortical thickness can have profound implications. However, detecting these changes can be complex due to the heterogeneity of typical brain development, and the extent of natural variability in brain morphology and organisational structure. As a result, there is currently little sensitivity to predict which infants will develop neurological impairment, nor understand the morphological and functional mechanisms which mediate later neurodevelopmental problems.
In previous work, we built a local model of brain growth and maturation using parametric growth curves [1] and more recently non-parametric Gaussian Processes (GP) [2]. This sought to predict what volumetric neonatal brain images should look like for a given age, degree or prematurity and sex, bounded by estimates of error (or expected variance) for the population. From this, it has been possible to detect babies whose brains (or areas of their brains) are outside of the normal range for their age. Nevertheless, while useful in white matter, this existing model lacks precision in cortex (brain surface), where variability in cortical morphology is greatest.
Currently, the most accurate way to compare different cortices is to perform analysis directly on the models of the brain’s surface [3,4,5]. Performing analysis in this way has been shown to improve understanding of cortical organisation [4], cortical development [6], cognition and behaviour [7]. For this reason, we seek to develop new generative models of cortical development using, geometric deep learning [8]. Specifically, techniques such as Variational Autoencoders (VAEs) or General Adversarial Networks (GANs), will be used to learn a spatio-temporal model of cortical growth and tissue maturation for healthy and diseased cohorts. This will be used to build models of healthy and pathological neurodevelopment in Downs Syndrome, Epilepsy and cognitive impairment following preterm birth and cardiac disease. The most suitable candidate for this project will have expertise in programming (python/c++), an interest in neuroscience and neonatal imaging, and preferably a background in machine and/or deep learning.
References:
- Dean DC, O’Muircheartaigh J, Dirks H, Waskiewicz N, Walker L, Doernberg E, et al. Characterizing longitudinal white matter development during early childhood. Brain Struct Funct 2015; 220: 1921–1933.
- O’Muircheartaigh J, Robinson EC, et al Modelling brain development to detect white matter injury in term and preterm born neonates. Brain (In press).
- Robinson EC, et al. Neuroimage. 2014 Oct 15;100:414-26.
- Glasser MF, Coalson TS, Robinson EC, et al. Nature. 2016 Aug;536(7615):171.
- Makropoulos A, Robinson EC, et al Neuroimage. 2018 Jun 1;173:88-112.
- Garcia KE, Robinson EC, et al Proceedings of the National Academy of Sciences. 2018 Mar 20;115(12):3156-61.
- Bijsterbosch JD, et al. Elife. 2018 Feb 16;7:e32992.[8] Bronstein MM, et al. IEEE Signal Processing Magazine. 2017 Jul 11;34(4):18-42.
- Bronstein MM, et al. IEEE Signal Processing Magazine. 2017 Jul 11;34(4):18-42.