Automated processing pipeline for neonatal diffusion MRI in the developing Human Connectome Project

Abstract

The developing Human Connectome Project is set to create and make available to the scientific community a 4-dimensional map of functional and structural cerebral connectivity from 20 to 44 weeks post-menstrual age, to allow exploration of the genetic and environmental influences on brain development, and the relation between connectivity and neurocognitive function. A large set of multi-modal MRI data from fetuses and newborn infants is currently being acquired, along with genetic, clinical and developmental information. In this overview, we describe the neonatal diffusion MRI (dMRI) image processing pipeline and the structural connectivity aspect of the project. Neonatal dMRI data poses specific challenges, and standard analysis techniques used for adult data are not directly applicable. We have developed a processing pipeline that deals directly with neonatal-specific issues, such as severe motion and motion-related artefacts, small brain sizes, high brain water content and reduced anisotropy. This pipeline allows automated analysis of in-vivo dMRI data, probes tissue microstructure, reconstructs a number of major white matter tracts, and includes an automated quality control framework that identifies processing issues or inconsistencies. We here describe the pipeline and present an exemplar analysis of data from 140 infants imaged at 38-44 weeks post-menstrual age.

Keywords: Brain; Connectome; Diffusion MRI; Newborn; Quality control; Tractography.

Fig. 1

Overview of some challenges posed by neonatal dMRI data processing. A) An image damaged by motion causes slice misalignment due to within-volume motion, severe signal dropouts across multiple slices (red arrows) and signal hyper-intensities due to spin-history effects (green arrows). Moreover, when using fast EPI sequences, susceptibility induced distortions are present along the phase encoding direction. B) Changes in gyrification profiles and tissue contrast throughout development need to be accounted for when analysing neonatal data; moreover, small age differences typically correspond to significant brain volume changes, which can modulate the data SNR. C) The white matter signal attenuation (i.e., the ratio between averaged diffusion weighted and diffusion un-weighted signal) of neonates is higher and shows more contrast than in the adult brain.

Fig. 2

Summary overview of the dHCP neonatal processing pipeline. The “best” (i.e., the least affected by within-volume motion) pairs of b0 volumes for each phase encoding direction are extracted from the raw dMRI data. These are used to estimate the off-resonance field that is then used in the simultaneous correction of motion artefacts, susceptibility-induced and eddy current distortions. Data are super-resolved and, after pre-processing is complete, local diffusion and microstructural models are fitted in every voxel. After registration of dMRI data to the age-matched template automated tractography extracts a number of white matter tracts.

Fig. 3

Raw b0 volumes acquired using four different PE directions (LR: Left - > Right, RL: Right - > Left, AP: Anterior - > Posterior, PA: Posterior - > Anterior). Fifth column shows an example volume after correcting for susceptibility-induced distortions. Last column shows the estimated field map in Hertz.

Fig. 4

Coronal and axial views of raw and pre-processed dMRI images using the outlier identification and replacement extension of EDDY. The data has been corrected for motion, susceptibility and eddy-current-induced distortions, and slices affected by severe signal dropouts have been correctly identified as outliers and replaced with the predicted signal in undistorted space.

Fig. 5

The leftmost column shows sagittal views from two volumes (b = 1000 s/mm²) of one subject. The middle and rightmost columns show pre-processed sagittal views of the same two volumes using a volume-based and a slice-to-volume movement model respectively.

Fig. 6

Registration accuracy using BBR. The fully pre-processed average attenuation profile volume for the b = 1000 s/mm² shell is registered to the high resolution T2-weighted volume in subject's native space (top row). When using BBR the tissue boundaries are better aligned to those extracted from the high-resolution volume, as shown by the two insets (red box). Red arrows point to areas where standard volumetric registration resulted in residual misalignments in the left hemisphere (bottom row). White/grey matter (green) and grey/csf boundaries (light blue) are overlaid on top of the attenuation profile transformed to the T2 anatomical space.

Fig. 7

Comparison between three subjects (age-matched: 41 weeks post-menstrual age) based on average tissue-specific CNR and SNR. SNR was sampled separately from the whole white and grey matter mask, CNR was sampled only from the whole white matter mask. Violin plots show CNR and SNR group distributions for 20 age-matched subjects. The colour-coded stars represent the three example subjects. Axial slices are obtained from the respective averaged b-shell volumes. Contrast between the three main tissue types (white, grey matter and CSF) increases from the bottom to the top together with CNR and SNR estimated indices. The QC framework identifies cases where the pre-processing pipeline or data acquisition were not fully successful (bottom row). The best subject shows clear delineation between white and grey matter over the whole brain volume.

Fig. 8

Output of the dMRI processing pipeline showing average DT-derived microstructural measures from 38 to 44 weeks (20 subjects per-age group). Expected changes with increasing age include: FA increases and MD decreases in the frontal white matter (red arrows) and internal capsules (green arrows); the corpus callosum (yellow arrows) and internal capsule both show high FA values; RD decreases; and MK declines in grey matter and increases in primary projections (green arrows). Diffusivities are expressed in μm²/ms.

Fig. 9

Output of the dMRI processing pipeline showing NODDI metrics from 38 to 44 weeks post-menstrual age (20 subjects per-age group). Extra-neurite volume fraction is defined as 1-(intra-neurite). Periventricular deep white matter regions (red arrows) show a high isotropic compartment volume fraction that decreases with age. Fibre dispersion also decreases in those regions. The posterior limb of the internal capsule (green arrows) shows very high intra-neurite volume fractions and low dispersion values.

Fig. 10

Output of the dMRI processing pipeline from 38 to 44 weeks post-menstrual age (20 subjects per bin) showing changes in fibre orientation estimations. Upper panel shows volume fractions of the 2^nd (red-yellow) and 3^rd (blue-light blue) fibres averaged across bin overlaid on top of the T2-weighted template. Second fibres get increasingly estimated across age in the periventricular deep white matter regions (green arrows). In the highlighted posterior portion of the centrum semiovale region (green box, zoomed insets in the bottom row) up to three fibres are estimated across age and survive the .05 threshold on their average volume fractions (1^st fibre in green, 2^nd in red and 3^rd in blue).

Fig. 11

Automated probabilistic tractography framework. A set of standardized ROIs is defined for week 44. The example shows the masks for virtually dissecting the forceps major: seed mask is in green, target mask is in blue and exclusion mask is in red. The ROIs are then warped first from week 44 to the age-at-scan-matched template and then to the subject's dMRI space. Tractography is run in subject diffusion space and results are then stored in age-at-scan-matched template space.

Fig. 12

Output of the dMRI processing pipeline from 38 to 44 weeks post-menstrual age (20 subjects per bin) showing maximum intensity projections of the automated probabilistic tractography results averaged across 80 dHCP subjects (20 per age bin). All tracts are visualized using the same thresholds (0.005, 0.05). These thresholds extract the main body of the tracts and might remove lateral projections due to the variable and low number of crossings that can be estimated especially at the very young age. A) Sensory-motor projection fibres: corticospinal tract (pink) and superior thalamic radiation (yellow) B) Thalamic projection fibres: anterior thalamic radiation (green), acoustic radiation (red-yellow), posterior thalamic radiation (blue-light blue); C) Callosal and association fibres: forceps minor (pink), forceps major (yellow), superior longitudinal fasciculus (red), inferior longitudinal fasciculus (green), inferior fronto-occipital fasciculus (blue). D) Brainstem tracts and association fibres in the left hemisphere: middle cerebellar peduncle (blue-light blue), medial lemniscus (brown), cingulate gyrus part of the cingulum (yellow), parahippocampal part of the cingulum (green), fornix (pink), uncinate fasciculus (red-yellow).

Fig. 13

Linear regression analysis coefficients for tract-specific microstructural indices. Maturational speed (i.e., slopes) indicated as β_age, predicted microstructural index value at the 41^st week (i.e., intercept) indicated as β₀. Slopes are normalised by the tract-specific scalar value averaged across ages, to facilitate comparison. Values in bar plots are sorted according to the FA magnitudes. Diamonds indicate that microstructural measures and age-at-scan were significantly, i.e., p < 0.05 corrected related for a specific tract. Different colours represent different tracts category. Scatterplots relate microstructural indices obtained from the same analysis technique. Arrows in the scatterplot show predicted changes after one week. High MD values in the fornix can be explained (as in the case of adult MD images) by partial volume effects with CSF regions. MD is expressed in μm²/ms. See Table 2 for abbreviations.