AbstractDevelopmental processes taking place during the third trimester and the neonatal period lay the foundation for a functioning human brain. In the course of these months, neuronal migration, cellular organisation, cortical development and myelination shape the form and function of our arguably most complex and outstanding organ.
Di˙usion weighted MRI (dMRI) has been extensively used to study the rapid changes in microstructural properties of white and grey matter non-invasively and provides contrast that is complementary to other imaging modalities [Yoshida et al., 2013]. The sensitivity to processes on the cellular level has made di˙usion imaging a tool for studying white matter development and the early detection of injury [Hüppi, Dubois, 2006].
Linking the measured signal to changes in the cellular composition and organisation of brain tissue poses data processing challenges unique to the pediatric population. In particular, movement during the acquisition corrupts di˙usion images beyond repair and requires manual data cleaning. We developed a neural network classifier that can perform this task automatically, allowing large-scale automated processing and analysis of di˙usion data.
Also, inferring cellular tissue properties from the signal is diÿcult as the brain simul-taneously undergoes a number processes that could alter the contrast in various ways. In simulations, I investigate the validity of often implicitly assumed relations between quantities derived from Di˙usion Tensor Imaging (DTI) and myelination in the context of changing tissue compartment volume fractions, showing that the interpretation of DTI parameters is flawed in the absence of a-priori knowledge about tissue microstructure.
In recent years, progress in acquisition and reconstruction techniques have facilitated acquiring quantitatively and qualitatively richer di˙usion images. Currently, High Angu-lar Resolution Di˙usion Imaging (HARDI) and higher order di˙usion models are uniquely positioned to capture and characterise developmental and maturation processes. The De-veloping Human Connectome Project (dHCP) is a group e˙ort to advance the field of pediatric MRI and has made possible much of the work in this thesis. The HARDI data acquired as part of the dHCP captures microstructural properties of the developing brain with an unprecedented quality and information content.
Characterising tissue properties requires a model that allows inferring processes on the cellular level from HARDI data. To build this model, it is necessary to incorporate domain knowledge about physical and biological properties of brain tissue. Even for adult populations, where brain tissue properties are comparatively static, developing higher order di˙usion models that provide microstructure-specific markers is an open research question [Novikov, Kiselev, Jespersen, 2018]. For these reasons, this thesis investigates the use of data-driven techniques for the study of brain development, which do not require explicit a priori models of tissue microstructure, but rather attempt to decompose the observed signal into interpretable components.
In chapter 8, we develop tools to produce an unbiased group template of tissue prop-erties at term, using a method that makes few assumptions about the microstructual properties of neonatal brain tissue. However, rapid brain maturation entails changes in tissue properties that require taking the temporal component into account. This term-time template is extended to the longitudinal domain in chapter 9, capturing tissue maturation patterns from 33 to 44 weeks gestational age in the dHCP cohort.
Together, these developments pave the way for detailed investigations into the devel-opment of the human brain. These techniques will form the basis for more advanced analyses, and will hopefully provide useful insights not available using existing methods.
Parts of this thesis and work related to experiments performed in this thesis have been presented at conferences under the titles "E˙ect of demyelination on di˙usion ten-sor indices: A Monte Carlo simulation study" [Pietsch, Tournier, 2015], "Multi-contrast di˙eomorphic non-linear registration of orientation density functions" [Pietsch et al., 2017a], "Transfer learning and convolutional neural net fusion for motion artefact de-tection" [Kelly et al., 2017], "Multi-shell neonatal brain HARDI template" [Pietsch et al., 2017b], and "Longitudinal multi-component HARDI atlas of neonatal white matter" [Pietsch et al., 2018].
A manuscript with the title "A framework for multi-component analysis of di˙usion MRI data over the neonatal period" based on chapter 9 is currently under review in NeuroImage.
|Date of Award||2018|
|Supervisor||Jacques-Donald Tournier (Supervisor), Jonathan O'Muircheartaigh (Supervisor), Paul Aljabar (Supervisor) & Serena Counsell (Supervisor)|