The period of rapid cortical expansion during fetal and early neonatal life is a crucial time over which the brain’s surface transforms from a smooth sheet to a highly convoluted surface. During this time, the cellular foundations of our advanced cognitive abilities are mapped out, as connections start to form between distant regions, myelinating, and later pruning, at different rates. Alongside the development of this neural infrastructure, functional brain activations start to be resolved, reflecting the development of cognition.
Much of what is currently known about the early human connectome has been learnt from models of preterm growth . Whilst invaluable, it is known that early exposure to the extra-uterine environment has long-term implications For this reason, the Developing Human Connectome Project (dHCP) seeks to image emerging brain connectivity for the first time in a large cohort of fetuses and newborn infants.
Within the next few months the dHCP will make a second, more substantial release of healthy term-born, and preterm neonatal images. This comprehensive data sets reflects morphological, functional and micro-structural development, including 0.5mm3 reconstructed T1 and T2 structural images, 15 mins resting state functional MRI (fMRI), and and 300 volumes of diffusion MRI (dMRI) acquired using a multi-shell (b=400, 1000 and 2600 s/mm2) High Angular Resolution Diffusion Imaging (HARDI) protocol. All data has gone through, carefully refined, reconstruction, artefact correction and image processing protocols (see Makropoulos and Robinson et al. 2018, Bastiani et al 2018, Fitzgibbon et al 2017.); surface mesh models of the cortex have been estimated as a target for morphological, micro-structural and functional analysis; and all stages of the processing have faced robust quality control (QC) and manual checking by expert observers.
In delivering this it is important to stress that neonatal data presents unique challenges relative to adult data sets (such as the adult HCP cohort) as our babies are imaged during natural sleep. Data is thus susceptible to high levels of motion (Fig. 1). In addition. scan times are limited as it is unethical to scan babies for comparable lengths of time to adults (the dHCP neonatal total scan time is approximately 1 hour, whereas the adult HCP total scan time is 4 hours). Limited scan times impact image resolution and whilst significant effort has been made to push image resolution through novel acquisition schemes, ultimately babies brains are much smaller than those of adults leading to a lower relative resolution; this results in significant partial volume (blurring of intensities across tissue boundaries – Fig 2). Maturing brains have inverted image contrast on T2’s relative to adults, meaning that standard protocols for cortical surface mesh modelling such as FreeSurfer are not applicable. Finally over the period of dHCP data collection (20-44 weeks gestation), the brain undergoes rapid growth and development – leading to much more dramatic changes in brain morphology, function and microstructure that would be observed in an adult cohort.
For these reason, we (the dHCP consortium) have developed a series of specialised image reconstruction and processing protocols that differ significantly from adult pipelines, such as those described by the HCP.
Specifically, to first address the concerns of motion, images are acquired in stacks of 2D slices which are subsequently combined and reconstructed with motion correction, and dedicated corrections for spin history and diffusion effects for fMRI and dMRI.
Tailored tissue segmentations have been developed that are specially designed to accurately recover tissue boundaries in the face of partial volume and white matter hyperintensities (see Makropoulos 2014, Makropoulos and Robinson et al 2018, ). These form the basis of a surface reconstruction algorithm, which equally takes into account intensity values from the original T2 algorithm. This allows for correction of segmentation errors that are very specific to neonatal data (i.e. mis-segmentation of CSF cortical white matter, as a result of CSF and white matter both being high intensity in neonatal T2 Figs 3, 4).
The high resolution T1 and T2 structural images, and refined surfaces have then formed the basis of advanced volume and surface templating approaches, generating sets of spatio-temporally evolving atlases spanning the age range 36-44 weeks.
In the case of the volumetric template, advanced SVFFD (stationary velocity field free-form deformation) based registration and refined atlasing procedures has led to the generation of arguably the sharpness, and most temporally-consistent neonatal template available at this time (Fig 5). On the other hand, surface templating procedures have incorporated advances made to the MSM multimodal surface matching algorithm (first developed for alignment of multi-modality cortical imaging data from the HCP) to generate biologically plausible, evolving templates reflecting changes in cortical morphology, thickness and T1/T2 myelination over the same time period (Fig 6, 7). Further advances to MSM have allowed for the estimation of smooth and biologically plausible deformation fields for longitudinal cortical growth. These have been used to evaluate evolving trends in cortical expansion for preterm babies between 28 and 36 weeks PMA (Fig 8).
In parallel, dMRI and fMRI pipelines (papers in preparation) have been developed utilising advances in artefact and distortion correction (provided through extensions of FSL’s eddy algorithm) to generate sensitive local models of local diffusion, microstructure and function at each voxel. Residual noise in fMRI data has been reduced through classifying and regressing out nuisance signals from the data using specially trained ICA FIX.
Examples of the scale of advances to the neonatal fMRI are shown in Fig. 9 which demonstrates 17 strongly recognisable functional modes, reconstructed from 242 co-registered neonatal data sets, and decomposed using FSL’s PROFUMO approach.
For dMRI major white matter fasciculi are automatically estimated from the data using virtual in vivo dissection. In this, all data is registered to the spatio-temporal template space (Fig 5) in which seed, target, exclusion and waypoint masks for 30 tract bundles were drawn. These masks are then projected back into each subject’s native diffusion space where probabilistic tractography was run. The results, averaged across subjects, binned across different weeks are shown in Fig. 10 (right). Separate estimates of local tissue microstructure (intra/extra neurites compartments fractions, isotropic compartment fraction and estimates of neurites’ dispersion – all estimated using NODDI) are shown in Fig. 10 (left).
All this is due to be released for a large batch of data, within the next few months. For those interested in using or viewing the data once it is released, please register for the mailing list at https://neurostars.org/tags/developing-hcp . For those interested in applying the structural pipelines to their own data, these are already available here https://github.com/BioMedIA/dhcp-structural-pipeline.