Synopsis

We present a data-driven method for the correction of stripe artefacts in multi-shell diffusion data that might arise for example from spin-history or stimulated echo effects. It relies on a filter, based on a deep neural network trained with simulated data to detect and remove stripe artefacts from single volumes. This is used to destripe signal predictions obtained from a slice to volume reconstruction, which are then projected onto the input data to determine the appropriate modulation field. The corrected input data are then reconstructed again with reduced stripe artefacts. This approach is applied to super-resolution reconstructions of neonatal multi-shell high angular resolution data.

Introduction

Diffusion MRI (dMRI) provides unique information about the microstructural properties of brain tissue. The diffusion weighting is achieved via strong motion sensitisation gradients, which poses major challenges for in-vivo imaging.

Interactions with previous pulses (spin-history effects) in interleaved EPI, variations in slice timing, stimulated echoes, and scanner hardware limitations can all lead to inter-slice inconsistencies, which can particularly destabilise super-resolution reconstruction.

Here, we present a new strategy for destriping dMRI data based on a deep neural network integrated with a slice-to-volume reconstruction (SVR) framework. This approach does not require a model of the source of the artefact, is applicable to super-resolution acquisitions¹, images with slab boundary artefacts^2,3,4 and variable slice-timing artefacts⁵, is comparatively robust to motion⁶.

We use the reconstructed image (Recon) as the basis for destriping because data in this space is corrected for other image artefacts, facilitating estimating slice-wise intensity modulation in the absence of distortions, dropout and misalignment due to subject motion. The intensity modulation is then back-projected to source-space, and the stripe correction applied to the input data (figure 1a).

Methods

Data and reconstruction dMRI data were acquired in 300 volumes at $$$1.5\times1.5\times3mm$$$ resolution in 64 slices with 1.5mm slice-overlap; interleave 3, shift 2; MB=4; TR=3800ms; TE=90ms; diffusion weightings $$$b=[0,\,400,\,1000,\,2600]s/mm^2$$$ with $$$n=[20,\,64,\,88,\,128]$$$ directions, respectively^7–9. Images were reconstructed to native resolution and denoised¹⁰. Field maps and brain masks were estimated with FSL topup¹¹ and bet¹².

The dMRI data was processed using an SVR motion correction and super-resolution (native slice profile) reconstruction algorithm, based on a data-driven multi-shell low-rank data representation and outlier estimation^13,14.

Destripe neural network The destripe-filter consists of a convolutional neural network (figure 1c) that operates on individual $$$3D$$$ volumes. It was trained to remove simulated per-slice scale factors imprinted on reconstructed DW volumes from 6 neonates (100 training epochs) by predicting the original image volume from the corrupted input image via a multiplicative modulation-field applied to the input volume. The network uses pooling to down-sample the in-plane resolution to 16x16 points followed by linear upsampling and Gaussian smoothing ($$$std=7\,$$$voxels) of the field (figure 1c). Hence, the modulation-field is smooth in-plane. It is additionally constrained by high-pass filtering in the slice-direction to remove slowly varying bias fields and by scaling it globally to conserve the average image intensity in the brain mask, to avoid potentially shell-specific biases. This is repeated thrice to iteratively remove stripe-patterns.

Source field estimation The constrained modulation-field is applied to the motion-corrected intensity predictions, produced by projecting the reconstructed multi-shell SH image post SVR (Recon, see figure 1a), which were then converted back to spherical harmonics to obtain the filtered reconstruction (Reconfiltered). The projection of Reconfiltered to the scattered (but dropout-free) source-space yields the filtered signal predictions DW_pred^filtered. These are divided by the unfiltered signal prediction (DW_pred), and median-filtered in-plane (3x3-kernel) to obtain the modulation-field in source-space (Destripe field). The Destripe field is multiplied with the raw DW data (DW_raw) and passed through the SVR reconstruction to obtain the final reconstruction Recon^final (figure 1b).

Results and discussion

An example subject with a clear stripe pattern in the super-resolution SH coefficients is shown in figure 2a). The destripe network operates on the amplitude projection of Recon, back-projected into the multi-shell basis; this shows that destriping in amplitude space has removed the majority of the stripe pattern in the SH coefficients (figure 2b) across all shells and for $$$\ell=0,\,2\,4$$$. The reconstruction of the destriped raw data (Recon final) contains residual stripe patterns (figure 2d) that can be mostly removed by repeating the destripe filtering and reconstruction thrice (figure 2e). This indicates that our approach can suppress the majority of the stripe pattern present in the raw data.

To check whether a projection to the angular domain is necessary, the destripe field was estimated from the much higher SNR shell-average terms ($$$\ell=0$$$) and then applied to all SH terms of the respective shell. This performs comparably for the $$$\ell=0$$$ coefficients, yet in this case, the angular coefficients contain significantly higher stripe artefacts (figure 2c, 2d). Hence, a projection into the angular domain is beneficial and necessary for destriping.

Figure 3 shows example source-space Destripe fields. They do not contain anatomical features and have no discernable low-frequency or global bias. Multiple iterations refine the field with decreasing amplitude, indicating convergence (figure 3b). The destriping significantly removes stripe patterns of multi-shell-based metrics (figure 4).

Conclusion

Acknowledgements

References

1. Scherrer, B., Gholipour, A. & Warfield, S. K. Super-resolution reconstruction to increase the spatial resolution of diffusion weighted images from orthogonal anisotropic acquisitions. Med. Image Anal. 16, 1465–1476 (2012).

2. Setsompop, K., Fan, Q. & Stockmann, J. High‐resolution in vivo diffusion imaging of the human brain with generalized slice dithered enhanced resolution: Simultaneous multislice (gSlider‐SMS). Magnetic resonance (2018).

3. Van, A. T. et al. Slab profile encoding (PEN) for minimizing slab boundary artifact in three-dimensional diffusion-weighted multislab acquisition. Magn. Reson. Med. 73, 605–613 (2015).

4. Wu, W., Koopmans, P. J., Frost, R. & Miller, K. L. Reducing slab boundary artifacts in three-dimensional multislab diffusion MRI using nonlinear inversion for slab profile encoding (NPEN). Magn. Reson. Med. 76, 1183–1195 (2016).

5. Meyer, M., Biber, A. & Koch, M. A. Chasing the Zebra. The Quest for the Origin of a Stripe Artifact in Diffusion-Weighted MRI. in Student Conference Medical Engineering Science 2014: Proceedings 3, 199 (GRIN Verlag, 2014).

6. David, S., Heemskerk, A., Viergever, M. & Leemans, A. Closing the Venetian blinds: A processing strategy for correcting stripe artifacts in diffusion MRI data. in ESMRMB (2017).

7. Hutter, J. et al. Time-efficient and flexible design of optimized multishell HARDI diffusion. Magn. Reson. Med. (2017). doi:10.1002/mrm.26765

8. Hughes, E. J. et al. A dedicated neonatal brain imaging system. Magn. Reson. Med. 78, 794–804 (2017).

9. Tournier, J.-D. et al. Data-driven optimisation of multi-shell HARDI. in Proc. Intl. Soc. Mag. Reson. Med. 23, (2015).

10. Veraart, J. et al. Denoising of diffusion MRI using random matrix theory. Neuroimage 142, 394–406 (2016).

11. Andersson, J. L. R., Skare, S. & Ashburner, J. How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging. Neuroimage 20, 870–888 (2003).

12. Smith, S. M. Fast robust automated brain extraction. Hum. Brain Mapp. 17, 143–155 (2002).

13. Christiaens, D. et al. Multi-shell SHARD reconstruction from scattered slice diffusion MRI data in the neonatal brain. ISMRM (Paris) (2018).

14. Christiaens, D. et al. Learning compact q-space representations for multi-shell diffusion-weighted MRI. IEEE Trans. Med. Imaging (2018). doi:10.1109/TMI.2018.2873736

15. Pietsch, M. et al. A framework for multi-component analysis of diffusion MRI data over the neonatal period. Neuroimage (2018). doi:10.1016/j.neuroimage.2018.10.060

16. Jeurissen, B., Tournier, J.-D. D., Dhollander, T., Connelly, A. & Sijbers, J. Multi-tissue constrained spherical deconvolution for improved analysis of multi-shell diffusion MRI data. Neuroimage 103, 411–426 (2014).