The diffusion SHARD pipeline

Inputs

From reconstruction pipeline: rawdata/sub-{subid}/ses-{sesid}

Description Filename
Multi-band dMRI EPI (denoised reconstruction) dwi/sub-{subid}_ses-{sesid}_run-{seqnum}_rec-denoised_dwi.nii
Multi-band dMRI EPI (Chi2-maps) dwi/sub-{subid}_ses-{sesid}_run-{seqnum}_rec-chi2_dwi.nii

Outputs

The primary outputs of the full SHARD preprocessing pipeline are:

Description Filename
Preprocessed DWI data (denoising, motion correction, and destriping) dwi/sub-{subid}_ses-{sesid}_desc-preproc_dwi.nii.gz
List of b-values in FSL format dwi/sub-{subid}_ses-{sesid}_desc-preproc_dwi.bval
List of gradient directions in FSL format dwi/sub-{subid}_ses-{sesid}_desc-preproc_dwi.bvec

(located in derivatives/dhcp_dmri_shard_pipeline/sub-{subid}/ses-{sesid})

The complete list of outputs and QC reports generated by the diffusion SHARD pipeline is listed in the Diffusion SHARD pipeline section of the directory structure summary.

Pipeline

The third data release includes an alternative dMRI processing pipeline based on denoising1 and SHARD motion correction2. This pipeline correponds more closely to the fetal processing pipeline3, and is therefore better suited for studies that include both the fetal and the neonatal data.

The pipeline consists of the following processing steps:

  1. Image denoising using random matrix theory with optimal shrinkage in the complex domain (i.e., using magnitude and phase images)1. This denoising approach also accounts for spatial noise correlations introduced by SENSE and Partial Fourier subsampling. After denoising, the magnitude images are extracted.

  2. The denoised images are then corrected for Gibbs ringing4.

  3. Fat shift suppression is achieved through local outlier reweighting in slice-to-volume reconstruction (step 5)5. The local outlier weights are computed independently based on the residuals of the SENSE reconstruction (Chi2-maps). This is achieved using a 2-class Gaussian Mixture Model with a Markov Random Field, fitted using Expectation-Maximization.

  4. The B0 field map, used to model susceptibility-induced distortion, is estimated using FSL Topup6. As input for topup, we selected the two “best” b=0 volumes for each of the 4 phase encoding (PE) directions based on a edge-detection filter in the slice direction. The indices of the selected 8 volumes are stored and used as reference; the ordering of the image volumes and the corresponding diffusion encoding is thus never changed.

  5. Motion correction using SHARD slice-to-volume reconstruction2. The inputs are the denoised and degibbsed multiband dMRI images (stage 2), the voxel weights (step 3), and the field map (step 4). The output are estimated subject motion traces, slice weights used to correct dropouts, and the SHARD representation fitted to the scattered slice data. The SHARD reconstruction also modelles the slice profile to recover the images at isotropic resolution. The motion- and distortion-corrected image is stored as a series of Spherical Harmonics (SH) coefficients for each shell.

  6. The 5D SHARD output image is projected onto the original diffusion encoding used during acquisition, to provide the dMRI output in a format compatible with conventional software. This is a one-to-one mapping representing identical image information. Note that motion-induced gradient reorientation is modelled during SHARD slice-to-volume reconstruction in step 5; the gradient table hence remains unchanged.

  7. Inter-slice intensity inhomogeneities were subsequently estimated and corrected on the motion-corrected projected dMRI volumes from step 6 using dStripe7. Note that these corrections were applied to the reconstructed dMRI data, not to the 5D SHARD output image.

  8. Rigid alignement to high-resolution structural (T2-weighted) space using normalised mutual information (NMI) based registration with FSL Flirt 8on the mean b=1000 s/mm2 shell. This transformation is combined with a non-linear registration9 of the T2w volume to the 40 weeks template10 to allow transformations between diffusion and atlas spaces.

Diffusion MRI QC

Automated quality control metrics are calculated for several key steps in the pipeline. Specifically, the data release includes estimates of:

In addition, overview screenshots of the motion traces and destriped output data are generated. Based on these screenshots, visual pass/fail Quality Control identified a small subset of cases to be discarded from analysis.

All QC metrics are available in the combined.tsv spreadsheet in the supplementary.

References

1. Cordero-Grande, L., Christiaens, D., Hutter, J., Price, A.N., Hajnal, J.V. Complex diffusion-weighted image estimation via matrix recovery under general noise models Neuroimage (2019), 200: 391-404. DOI: 10.1016/j.neuroimage.2019.06.039

2. Christiaens, D., Cordero-Grande, L., Pietsch, M., Hutter, J., Price, A.N., Hughes, E.J., Vecchiato, K., Deprez, M., Edwards, A.D., Hajnal, V., Tournier, J-D. Scattered slice SHARD reconstruction for motion correction in multi-shell diffusion MRI NeuroImage (2021), 225: 117437. DOI: 10.1016/j.neuroimage.2020.117437

3. Christiaens, D., Cordero-Grande, L., Price, A.N., Hutter, J., Hughes, E.J., ounsell, S.J., Tournier, J-D., Hajnal, J.V. Fetal diffusion MRI acquisition and analysis in the developing Human Connectome Project ISMRM 2020, O629.

4. Kellner, E., Dhital, B., Kiselev, V.G., Reisert, M. Gibbs-ringing artifact removal based on local subvoxel-shifts. Magnetic Resonance in Medicine (2016) 76: 1574–1581. DOI: 10.1002/mrm.26054

5. Christiaens, D., Cordero-Grande, L., Hutter, J., Price, A.N., O’Murchearthaigh, J., Vecchiato, K., Hajnal, J.V., Tournier, J-D. Fat-shift suppression in diffusion MRI using rotating phase encoding and localised outlier weighting ISMRM 2020, O981.

6. Andersson, J.L., Skare, S., and Ashburner, J. How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging Neuroimage (2003), 20(2): 870-888. DOI: 10.1016/S1053-8119(03)00336-7

7. Pietsch, M. and Christiaens, D. and Hajnal, J.V. & Tournier, J-D. dStripe: slice artefact correction in diffusion MRI via constrained neural network biorxiv (2020) DOI: 10.1101/2020.10.20.347518

8. Jenkinson, M., Bannister, P., Brady, M., and Smith, S. Improved optimization for the robust and accurate linear registration and motion correction of brain image Neuroimage (2002), 17(2): 825-841. DOI: 10.1006/nimg.2002.1132

9. Andersson, J.L.R., Jenkinson, M., and Smith, S. Non-linear registration, aka spatial normalisation FMRIB technical report TR07JA2 (2010).

10. Schuh, A., Makropoulos, A., Robinson, E.C., Cordero-Grande, L., Hughes, E., Hutter, J., Price, A.N., Murgasova, M., Teixeira, R.P.A.G., Tusor, N., Steinweg, J.K., Victor, S., Rutherford, M.A., Hajnal, J.V., Edwards, A.D., and Rueckert, D. Unbiased construction of a temporally consistent morphological atlas of neonatal brain development bioRxiv (2018), 251512. DOI: 10.1101/251512