Efficient Macroscopic Motion Correction for Multi-shot DTI
zhongbiao xu1, Yanqiu Feng1, Wufan Chen1, Zhigang Wu2, Ha-kyu Jeong3, Wenxing Fang2, Yingjie Mei1,4, Li Guo1, and Feng Huang2

1School of Biomedical Engineering and Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou, China, People's Republic of, 2Philips Healthcare (Suzhou), Suzhou, China, People's Republic of, 3Philips Korea, Seoul, Korea, Republic of, 4Philips Healthcare, Guangzhou, China, People's Republic of


Though cMUSE proposed by our group tackles the pixels mismatch of macroscopic motion in multi-shot EPI by clustering and registration, it neglects altered gradient directions. In this work, we treat motion induced variations in gradient direction as addtional diffusion direction(s). The proposed method simply and effectively solves the gradient direction alternation due to macroscopic motion in multi-shot DTI.


Multi-shot EPI has been an alternative method of single-shot EPI in DTI acquisition due to reduced distortion and improved spatial resolution [1]. However, subjects’ involuntary motion often challenges this technique. Namely, miniscule motion during diffusion gradient will induce shot-to-shot phase variations, leading to ghosting artifact. Furthermore, macroscopic motion will introduce pixels mismatch, causing images blurring, and altered diffusion directions, resulting in inaccurate tensor estimations. The famous MUSE [2] and IRIS [3] well resolve the miniscule motion induced shot-to-shot phase variations, but not account for macroscopic motion. Though cMUSE [4] proposed by our group tackles the pixels mismatch of macroscopic motion, it neglects altered gradient directions and has the limitation on net acceleration factor due to the dependence on SENSE reconstruction. The goal of this work is to address the altered diffusion direction issue from macroscopic motion in multi-shot DTI while avoiding the drawback of limited acceleration factor in cMUSE.


IRIS extracts phase information from low resolution and high SNR navigator images, thus avoids the limitation on acceleration factor. Here, we propose a novel IRIS-based method, named clustered IRIS (CIRIS) based on the idea of cMUSE and the advantage of IRIS. CIRIS firstly classifies the navigator images into clusters which have no macroscopic intra-cluster motion using the method suggested in cMUSE. After that, intermediate images are reconstructed with the clustered navigator data and corresponding image data by using IRIS. For DWI, these intermediate images are combined using weighted average after registration to generate the final reconstruction. For DTI, they are registered to the non-DW image and the rotation matrix from registration is used to calculate the actual gradient direction to compensate motion’s effect on diffusion encoding gradient directions. Next, the registered images and corrected gradient directions are used to calculate the diffusion tensor. That is to say, we treat motion induced variations in gradient direction as updated diffusion direction(s), instead of correcting these directions as in AMUSE [5]. The performance of CIRIS was evaluated on multi-shot EPI brain data acquired on a Philips Multiva 1.5T scanner (Philips healthcare, Suzhou, China) using an 8-channel head coil. The acquisition parameters include: 1) DWI scan: number of shots (NS) =3, SENSE factor (SF) =2, number of signal averages (NSA) =6; 2) DTI scan: NS=2, SF=2, NSA=12, number of diffusion gradients=10. For each scan, two datasets were acquired: a stationary dataset used as gold-standard and a motion dataset where the volunteer was asked to rotate his head 4 times during the scan. Two quantitative measures were used to assess tensor accuracy: 1) angular deviation (A) between the major eigenvectors of the gold-standard image and the reconstructed images; 2) roots mean square error (RMSE) in fractional anisotropy (FAerr) compared to the golden standard. Two regions of interest (ROIs) were selected in the genu and splenium of the corpus callosum (GCC and SCC). All processing was performed using Matlab (2.33GHZ, 4GM RAM).


The total computation time except registration is 5.60s per direction for DTI. Fig.1 compares the reconstructions with IRIS, cMUSE and CIRIS on DWI motion dataset with an acceleration factor 6. Compared to IRIS (Fig.1a) and cMUSE (Fig.1b), CIRIS (Fig.1c) gave a clear improvement in image quality. Fig.2 shows the reconstructed images with different methods and the corresponding angular deviation maps in GCC and SCC. For macroscopic motion dataset, IRIS (Fig.2a) produced significant image blurring, causing large angular deviation. Using the proposed method, the artifact was removed and thus the angular deviation was significantly reduced (Fig.2b and Fig.2c). When accounting for altered diffusion direction, the lower angular deviation was achieved (Fig.2c vs. Fig.2b). Table 1 further demonstrates that the tensor measurements were more accurate using CIRIS with gradient correction.


The proposed method corrects the pixels mismatch by clustering and registration, and treats motion induced gradient direction variations as updated diffusion direction(s). Because motion is estimated from navigator data, the proposed method overcomes the limitation of acceleration factor in cMUSE (Fig.1c vs. Fig.1b). Compared to AMUSE [5], CIRIS has lower computational cost due to the clustering, and avoids diffusion direction correction.


The proposed method simply and effectively solves the gradient direction alternation due to macroscopic motion in multi-shot DTI while avoiding the shortcomings of cMUSE


No acknowledgement found.


[1] Bammer, R. European Journal of Radiology 2003, 40:169-184 [2] Chen, N-k. et.al. NeuroImage 2013;72:41-47 [3] Jeong, H-k. et.al. MRM 2013; 69:793-802 [4] Xu, Z-b. et.al. ISMRM 2015, p961 [5] Guhaniyogi, S. et.al. MRM DOI 10.1002/mrm.25624


Fig.1 Reconstruction with different methods on DWI motion data with acceleration factor 6: (a) IRIS, (b) cMUSE, (c) CIRIS

Fig.2 Reconstruction with different methods on DTI motion data: (a) IRIS, (b) CIRIS without gradient correction, (c) CIRIS without gradient correction. The top row is the reconstructed images and the bottom row is angular deviation of the major eigenvectors.

Table 1 In vivo tensor correction

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)