1st supervisor: Serena Counsell, King’s College London
2nd supervisor: J.-Donald Tournier, King’s College London
During gestation and the first few years of life, the human brain undergoes rapid changes, particularly in brain size, morphology and tissue microstructure. This large variability, induced by the normal process of development, masks more subtle changes due to normal variation, abnormal development, or pathology. This hinders the study of longitudinal processes of healthy and abnormal brain development, making it difficult to diagnose brain abnormalities, make prognoses about long-term outcomes, or monitor the efficacy of potential therapies.
Leveraging large existing databases of imaging and behavioural data, this project aims to build a longitudinal mapping of brain tissue microstructure in neurodevelopment. However, the rapid brain growth, and also the changing environment before and after birth, lead to a need for different imaging protocols at different ages. There is therefore a need for data harmonization techniques that can map across the fetal, neonatal, and infant data collected in several studies, whilst capturing the longitudinal transition in the developing brain.
The KCL Perinatal Imaging Department leads several large-scale imaging studies into early brain development, most notably the Developing Human Connectome Project (dHCP; http://www.developingconnectome.org/), which comprises high-quality fetal and neonatal MRI images of the developing brain and associated genetic, clinical, environmental and behavioural data, acquired between 20 and 44 weeks gestational age. In this project, we will focus particularly on diffusion MRI data, which provides information regarding tissue microstructure, and link it to non-imaging based data including environmental, clinical, genetic and neurodevelopmental outcome data.
The in-utero and ex-utero environments are clearly very different, and the differences between fetal and neonatal imaging imposes a core challenge on any longitudinal analysis that intends to span the joint age range, as the age-dependent effects of interest can be difficult to disentangle from protocol-related image effects. We wish to overcome this challenge by creating a cross-protocol mapping in the overlapping age group that accounts for cross-subject variability. Specifically, we aim to use machine-learning methods to build a high-dimensional representation (manifold) of normal and abnormal brain images, defined based on differences in anatomical morphology and tissue microstructure. In this framework, the data harmonization problem is cast as a graph-matching problem where a mapping is learned between the overlapping age groups in the fetal and neonatal manifolds.
Furthermore, this framework would enable images corresponding to abnormal development or pathology to be identified as outliers from the manifold, or as lying close to the part of the manifold corresponding to known abnormalities. The outcome information present in the database can then be propagated across the fetal and neonatal population graphs and also between both populations thanks to the learned cross-protocol mapping. This, in turn, can be used to yield accurate subject-specific outcome predictions without the need to explicitly define criteria of healthy and abnormal development.
While the primary focus of the project will be on harmonizing the fetal and neonatal dHCP data, the methods developed in this project will also be applicable to other large-scale imaging studies of brain development, thus extending the age range to infant development and enriching the available data in the overlapping time period.
A pictorial representation of the framework that will form the basis for this project. Each data point represents microstructural information for each subject (as depicted in the inset) in the fetal (dark blue) and neonatal (light blue) cohorts. These are embedded in a nearest-neighbour graph to allow optimal information flow within cohorts, across cohorts, and time (left to right axis). Cross-cohort information flow will be enabled by establishing the optimal mapping between cohort-specific representations via non-parametric deep-learning neural network methods.