Sean Fitzgibbon1, Jesper Andersson1, Samuel Harrison1, Emma Robinson2, Jelena Bozek3, Antonios Makropoulos4, Matteo Bastiani1, Ludovica Griffanti1, Robert Wright4, Andreas Schuh4, Emer Hughes5, Jonathan O'Muircheartaigh5, Tomoki Arichi5,6, Judit Ciarrusta5,7, Ana Dos Santos Gomes5, Joanna Allsop5, Johannes Steinweg5, Nora Tusor5, Julia Wurie5, Suresh Victor5, Anthony Price5, Lucillio Cordero Grande5, Jana Hutter5, Christian Beckmann8, Joseph Hajnal5, Daniel Rueckert4, David Edwards5, Stephen Smith1, Mark Jenkinson1, Eugene Duff1,9
1FMRIB, Wellcome Centre for Integrative Neuroimaging, University of Oxford, Oxford, United Kingdom, Oxford, United Kingdom, 2Department of Biomedical Engineering, King's College London, London, United Kingdom, 3Faculty of Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia, 4Biomedical Image Analysis Group, Imperial College London, London, United Kingdom, 5Centre for the Developing Brain, King's College London, London, United Kingdom, 6Department of Biomedical Engineering, King’s College London, London, United Kingdom, 7Institute of Psychiatry, King’s College London, London, United Kingdom, 8Donders Institute for Brain, Cognition and Behaviour, Nijmegen, Netherlands, 9Department of Paediatrics, University of Oxford, Oxford, United Kingdom
The developing Human Connectome Project (dHCP) is a large-scale imaging project that aims to create a detailed 4-dimensional connectome of the developmental period spanning 20 to 44 weeks post-menstrual age (PMA), recording structural, diffusion and functional MRI measures in over 1000 in- and ex-utero subjects. Neonates present significant challenges to data processing due to low and variable contrast and high levels of head motion. This abstract presents the latest features to be incorporated into the fMRI processing framework for neonates. Specifically, dynamic distortion correction, multimodal registration to standard space, neonatal HRF prior for estimating RSNs, improved FIX training, and a new automated QC framework.
Imaging was acquired at St. Thomas Hospital, London, on a Philips 3T scanner. High temporal resolution fMRI optimized for neonates [Price et al., 2015] used MB9 accelerated echo-planar imaging and was collected for 15 minutes, TE/TR=38/392ms, 2300 volumes, with an acquired resolution of 2.15mm isotropic.
Distortion and motion correction now incorporates slice-to-volume (S2V; Fig. 1a) correction to mitigate intra-volume movement (EDDY [Andersson et al., 2016]) and dynamic distortion correction (DDC) to correct time-varying susceptibility-induced distortions due to subject movement [Andersson et al., 2001].
The standard space is defined as the 40-week template from the Gousias atlas [Gousias et al., 2012], which contains T1/T2 volumetric templates per week from 28-44 weeks PMA. We have augmented it with week-to-week nonlinear transforms estimated using a diffeomorphic multi-modal (T1/T2) registration (ANTs SyN [Avants et al., 2008]).
Functional-to-structural registration is performed in 2-stages: (1) linear registration (6-dof) of the multiband EPI to a single-band EPI (SBref), and (2) boundary-based registration of the SBref to the T2 structural.
Standard-to-structural registration is performed with a multimodal non-linear registration (ANTs SyN) of the age-appropriate T1/T2 template to the subjects T1/T2 structural, which is then combined with the appropriate atlas week-to-week warps to yield a (40wk) standard-to-structural warp.
ICA denoising is performed using FIX [Salimi-Khorshidi et al., 2014], pre-trained on manually-labelled data from 25 subjects, to identify artefactual ICs (median TPR=100%, median TNR=95.4%).
Group-average RSNs are identified with PROFUMO[Harrison et al., 2015], a Bayesian framework that identifies probabilistic functional modes using constraints associated with the neonatal hemodynamic response function [Arichi et al., 2012] and inter-subject variability.
The pipeline incorporates a new automated QC which compares numerous individual subject quality metrics reflecting different stages of the pipeline against the population distribution and flags outliers for manual inspection (Fig. 1b).
We assessed the pipeline on a subset of 40 subjects. S2V correction significantly (p<0.001) improves SNR compared to traditional rigid-body motion correction (Fig. 1c), and DDC further improves SNR in anterior and posterior areas where susceptibility distortions are expected. There was significant additional improvement (p<0.001) in SNR after ICA denoising (Fig. 1d).
PROFUMO was performed on 267 subjects (aged 37-44 weeks PMA), and 13 RSNs that correspond qualitatively to known adult [Smith et al., 2009] and neonate [Doria et al., 2010] RSNs (Fig. 2) were resolved with fine spatial detail. Previous revisions of the framework could only resolve 9 reliable RSNs.
Processing refinements integrated into the dHCP fMRI framework provide substantial reduction in movement related distortions, resulting in significant improvements in SNR, and detection of high quality RSNs from neonates. Ongoing analyses are probing the fine structure of these networks, and their variability across subjects and age, with the aim of defining a multi-modal time-varying map of the neonatal connectome.
BOLD fMRI 2
Normal Brain Development: Fetus to Adolescence 1
Modeling and Analysis Methods:
Motion Correction and Preprocessing
1|2Indicates the priority used for review