Introduction

Myelin water imaging techniques based on T₂ relaxometry proved to deliver quantitative measurements specific to myelin alterations^1–4. Recently, image resolution and acquisition time (TA) to obtain 3D whole-brain multi-echo T₂ data could be improved by combining parallel imaging with the multi-echo gradient and spin echo (GRASE) sequence⁵. However, using standard encoding and reconstruction techniques resulted in residual artifacts at high undersampling rates, and precluded further TA reduction without compromising image quality.
Following recent advances in the reconstruction of MR acquisitions with multiple echoes/phase-cycles/contrasts^6,7, this work proposes to exploit controlled aliasing, complementary encoding and redundant anatomical information across the different GRASE echoes in a joint-CAIPI (J-CAIPI) reconstruction to improve image quality and possibly further reduce the TA.

Methods

Sampling schemes
To investigate joint reconstruction of multi-echo GRASE data while preserving the contrast information in each echo, autocalibration signals (ACS) need to be acquired in the k-space of each echo. Therefore, in contrast to the previously proposed CAIPI undersampling⁵ (Figure 1A), a sampling pattern comprising an integrated ACS region and CAIPI undersampling of the outer k-space on an elliptical Cartesian grid was implemented (Figure 1B). Additionally, the implementation allows shifting the sampling patterns across echoes in both phase encoding directions, providing a complementary k-space coverage across the echoes which proved to enhance the output of joint reconstructions⁶ (Figure 1C). We also propose and investigate a further possibility to accelerate the scan time by retrospectively undersampling the ACS region (Figure 1D-E).
Reconstruction methods
The reference CAIPI acquisitions were reconstructed after calibrating 3D convolution kernels on an external scan (i.e. a fast ACS scan of ~3.5$$$\,$$$seconds performed before the actual GRASE acquisition).
Undersampled data with integrated ACS were reconstructed both with CAIPI, calibrating the 3D convolution kernels on the ACS region, and with a J-CAIPI algorithm that extends previous work on joint-GRAPPA⁶. In J-CAIPI, k-space data are first concatenated along the echo dimension to create additional channels as extra coils and then reconstructed using 3D CAIPI convolutional kernels. The number of available coils was compressed⁸ from 52 to eight for faster reconstruction in all experiments.
A significant amount of acquisition time is spent to sample the ACS region in integrated scans. Hence, to explore further acceleration possibilities, ACS regions were retrospectively undersampled and a two-step reconstruction was investigated: first, the integrated ACS regions were reconstructed using kernels trained on an external ACS scan; subsequently, both CAIPI and J-CAIPI algorithms were used to reconstruct the echo images.
Dataset
Three healthy volunteers (all male, [27-35]$$$\,$$$y/o) were scanned using a prototype 3D multi-echo GRASE sequence (whole-brain, 1.6$$$\,$$$mm³ isotropic) using three different sampling schemes: 1) external reference scan (N_y$$$\times$$$N_z$$$\,$$$=$$$\,$$$32$$$\times$$$32) and CAIPI R_y$$$\times$$$R_z^Δk_z$$$\,$$$=$$$\,$$$3$$$\times$$$2⁽¹⁾ (TA$$$\,$$$=$$$\,$$$10:30$$$\,$$$minutes); 2) integrated ACS (N_y$$$\times$$$N_z$$$\,$$$=$$$\,$$$32$$$\times$$$32) and CAIPI R_y$$$\times$$$R_z^Δk_z$$$\,$$$=$$$\,$$$3$$$\times$$$3⁽¹⁾, without and 3) with complementary k-space shifts across echoes (TA$$$\,$$$=$$$\,$$$10:34$$$\,$$$minutes). A 3D MPRAGE image was also acquired as anatomical reference. All sequence parameters are detailed in Table 1. The two acquisitions with integrated ACS were retrospectively undersampled following a CAIPI R_y$$$\times$$$R_z^Δk_z$$$\,$$$=$$$\,$$$2$$$\times$$$2⁽¹⁾ scheme, which would correspond to an effective TA of 6:18$$$\,$$$minutes.
Myelin water fraction comparison
Myelin water fraction (MWF) maps were computed from the multi-echo T₂ volumes using a non-negative least-square optimization with second-order Tikhonov regularization⁹. Averaged MWF values in white matter regions were extracted from segmentation masks automatically computed on the MPRAGE contrast¹⁰, and compared via Bland-Altman analysis.

Results

Residual aliasing artifacts in the T₂-weighted images were found to be more prominent in the data with higher acceleration rates and to be located at different positions in the shifted k-space acquisitions (Figure 2). Thanks to this complementary sampling, artifacts in shifted k-space data were drastically mitigated in the J-CAIPI reconstructions (Figure 2).
The residual artifacts in CAIPI reconstructions with higher acceleration rate in the outer k-space were found to severely compromise the image quality of the retrieved MWF maps (Figure 3). Conversely, such artifacts were cancelled out in MWF maps obtained from the J-CAIPI reconstruction of shifted k-space data (Figure 3). These maps also showed a more homogenous contrast over brain regions, as shown for instance by the MWF standard deviation in the corpus callosum, which was found to decrease from 5.97% in conventional CAIPI to 3.46% in J-CAIPI (Figure 3). Average MWF values across the enrolled volunteers are shown in Figure 4A. The Bland-Altman analysis comparing the MWF values obtained from J-CAIPI reconstructed k-space with fully sampled and retrospectively undersampled ACS regions revealed a mean bias of 0.38% with the limits of agreement spanning from -1.80% to 0.99% (Figure 4C).

Discussion and Conclusion

In this work, the J-CAIPI reconstruction of multi-echo GRASE data with integrated ACS and k-space shifts across echoes was proposed as an alternative to recently proposed CAIPI schemes⁵. While preserving comparable MWF values, J-CAIPI reconstructed data were found to be free from aliasing artifacts. Additionally, we showed that the J-CAIPI reconstruction can work robustly from undersampled integrated ACS regions reconstructed from an external ACS scan, allowing to further reduce the TA from 10:30$$$\,$$$minutes to 6:18$$$\,$$$minutes for a 1.6$$$\,$$$mm³ isotropic whole-brain scan.
Future work is planned to implement the prospective undersampling of the ACS region and the random sampling of the outer k-space to incorporate compressed sensing priors into J-CAIPI reconstruction, which showed good promise in single-echo GRASE imaging¹¹.

Acknowledgements

No acknowledgement found.