Introduction

Diffusion MRI (dMRI) is commonly used in many clinical and neuroscientific studies. For fast dMRI, echo planar imaging (EPI) is widely used, however geometric distortion due to field inhomogeneity often degrades the image quality of EPI. To address this problem, FSL TOPUP¹ is often used, where inverted phase-encodings, i.e blip-up and blip-down acquisitions are acquired separately and used to estimate the field map and perform distortion correction. Recently the BUDA-EPI approach² was proposed, which performs blip-up and blip-down acquisitions with complementary k-space sampling (Figure 1A) along with joint parallel imaging reconstruction of these two acquisitions to boost SNR and mitigate artifacts at high accelerations.

Motion artifacts pose a challenge in dMRI, as acquisitions are typically long and across a large number of diffusion directions³. In particular for BUDA-EPI, motion corruptions between the blip-up and –down shots result in poor B0 map estimation and image artifacts in the joint reconstruction (Figure 1C).

In this work, we introduce a motion-correction approach for BUDA-EPI reconstruction to achieve motion-robust distortion-free dMRI.

Method

Figure 1B shows the flowchart of the BUDA-EPI reconstruction approach. First, the blip-up EPI data and the blip-down EPI data were separately reconstructed using SENSE⁴ to obtain two images that were input to FSL TOPUP to estimate field maps. Along with the undersampled Fourier operator F_t and ESPIRiT⁵ sensitivity maps C, the estimated field maps E_t were incorporated into the Hankel structured low-rank constrained ($\parallel H(b)\parallel_{*}$) joint reconstruction^6-8 for both blip-up and -down data (d_t), to obtain distortion-free images x_t (for t^th shots).

To mitigate the motion artifacts shown in Figure 1C, we proposed a motion-correction reconstruction framework as shown in Figure 1D which utilizes the initial SENSE reconstructed images to estimate motion parameters and the field map (E₀), as follows:
i.We assumed at the beginning of scan, the first pair of blip-up/down shots were acquired without motion, and regard this as our base reference position (POS₀) from which a base field-map E₀ was estimated using FSL TOPUP. (other blip-up/down pairs with no motion could also be used as reference)
ii.The method shown in Figure 2A was used to estimate the motion parameters of each EPI-shot separately compared to the base position using FSL MCFLIRT⁹. By iteratively using MCFLIRT incorporated with field-map (E₀) information, a more accurate motion estimation was obtained with the distortion-corrected images.
iii.Motion parameters for each blip-up and each blip-down acquisition relative to the reference position were then incorporated into the motion transformation operator (R_t) to jointly reconstruct each blip-up and blip-down acquisition pair, to obtain a motion-corrected distortion-free image (Figure 1D).

The proposed method assumes that when motion occurs, the underlying B0-field inhomogeneity would approximately rotate with the motion. Figure 2B shows this to be a good approximation, where the rotated field map mimics closely that of the reacquired actual field map.

To test our proposed motion correction method, we designed static and dynamic experiments. For the static experiment, the volunteers first were required to keep still, this was regarded as a static reference. Then, the volunteers would move to another position and keep still during acquisition of all blip-up shots, and then move to the next position while acquiring the blip-down shots. For the dynamic experiment, the volunteers were asked to move indiscriminately after about 1 minute from the start of scan.

All studies were performed on a 3 Tesla MAGNETOM Prisma scanner with a 64-channel head receiver coil. For high-resolution in-plane correction experiments, the data were collected using: FOV=198×198×120mm³, 0.8×0.8×5mm³ resolution, R_inplane=4, partial-Fourier 6/8, TE/TR=67/3400ms, b-value=1000s/mm² with 24 diffusion directions. For the isotropic 3D image correction experiments, imaging parameters were: FOV=204×204×90mm³, 1.5mm isotropic resolution, R_inplane×MB=4x2, partial-Fourier 6/8, TE/TR=55/3500ms.

Results

The performance of the proposed method at different motion levels was tested with the static experiment (as shown in Figure 3). BUDA-EPI reconstruction without motion-correction exhibited increasing levels of blurring and artifacts with increasing motion level (A to C), while the proposed motion-corrected reconstruction produced high-quality results in all cases, comparable to the reference images.

Figure 4 shows the result of the dynamic experiment with high in-plane resolution and thick-slice, where 2D motion-correction was utilized. Figure 4A shows the time-varying motion estimates, with large translation and rotation up to +/-7mm and +/-15^o. Corresponding reconstructed images of 4 timeframes with/without motion-correction are shown in Figure 4B respectively. Figure 4C shows the averaged DWI of the still motion-free period (first minute of scan) and the moving period (from first minute onward). The proposed method produced much improved averaged DWI, with some residual image blurring in this extreme motion case, likely due to significant spin-history related-artifacts.

Figure 5 shows the result of the static experiment of images in three orthonormal planes with 1.5-mm isotropic resolution and 3D motion-correction, with the estimated motion level shown at the top of the figure. The proposed method presented similar image quality compared to the reference images.

Discussion and Conclusion

We proposed a motion-correction reconstruction method for brain dMRI acquired with BUDA-EPI. In vivo results demonstrated the efficacy of our proposed method. Future work includes refinement of the motion estimation and incorporation of constrained reconstruction to mitigate the spin-history issue, which will further improve the performance of our technique.

Acknowledgements

No acknowledgement found.