Distortion correction in diffusion weighted imaging of the brain: a quantitative comparison of four correction approaches
Ileana Hancu1, Ek Tsoon Tan1, Luca Marinelli1, Nathan White2, Dominic Holland2, Tim Sprenger3, and Jonathan Sperl3

1GE Global Research Center, Niskayuna, NY, United States, 2University of California San Diego, San Diego, CA, United States, 3GE Global Research Center, Munich, Germany


The performance of four distortion correction algorithms was investigated in a cohort of normal volunteers. While all approaches reduced distortion, it was found that the reversed polarity gradient methods were inherently better than registration or B0-mapping approaches. It was likely that the limited degrees of freedom of the registration approach could not account for localized magnetic field inhomogeneity. The extrapolation of B0 maps in the distorted EPI space introduced errors that decreased the overall performance of the B0-mapping method.


Correction of susceptibility-induced distortion in EPI images has become an integral part of brain DTI processing pipelines in the research community. There are three main avenues for distortion correction, and many implementations of similar approaches. Although pairwise comparisons of distortion correction performance were previously documented [1-3], a more comprehensive comparison of multiple leading correction algorithms remains elusive; such work would enable one to choose the optimal strategy for a given study. The goals of this work are to evaluate four of most commonly used distortion correction approaches, and document their performance in a cohort of normal volunteers.


Five normal volunteers underwent scanning sessions on a 3T, GE MR750 scanner. The acquisition series included anatomical T1-weighted 3D GRE imaging, T2-weighted FSE and fat-suppressed B0 mapping using 2D GRE (TE=2.3/3.1ms). Two B0 mapping acquisitions were performed, one at 128x128x6mm/2min acquisition time and the second one at 128x32x12mm/15sec acquisition time. Two DTI volumes (b=0 and 1000 s/m2) were also acquired, one at R=2 (48 slices; 128x128x3mm resolution), and the second at R=1 (72 slices; 128x128x2mm). The acquisition of the b=0 volumes of the DTI scans was repeated using reversed polarity gradient (RPG) EPI. The four separate approaches tested for correcting susceptibility induced distortion in the 10 sets of DTI images were:

a. Separately-acquired B0 maps: These were used as input for the correction algorithm described in [4]. Fat suppression and a bandwidth of 600Hz ensured the absence of phase jumps and phase wraps; fitting to cosine basis functions was used to extend the undistorted B0 maps in the distorted EPI space.

b. Higher-order image registration: Similar to [5], following rigid registration of the distorted EPI images to the anatomical, T1-weighted volumes, refinement of alignment was performed using a 20-parameter cubic-polynomial basis-function for only the phase-encode direction ((x',y',z')=(x,f(x,y,z),z)). As compared to affine or nonlinear free-form deformation image registration, this approach limits the nonlinearity to the primary direction where distortion is expected.

c. RPG-1: The two sets of spin-echo EPI images (acquired with forward and RPG’s) were used as input for the algorithm of [6] to generate the field map, hence pixel displacement.

d. RPG-2: The same 2 sets of EPI images described above were used as input to the minimization algorithm employed in FSL-TOPUP [7]. The quantitative metric of success was the cross-correlation coefficient (CCC) between the T2-weighted FSE images and the uncorrected/corrected b=0 volume of the DTI scans. This measure was computed using the entire imaging volume, as well as using only the brain volume obtained after skull-stripping (implemented using ROBEX [8]).

Results and discussion

An example of the performance of all corrections methods on one volunteer is presented in Figure 1. Table 1 presents the average CCC for the 5 volunteers, displayed before correction and after the four correction operations, over the head and brain areas, in the R=1 and R=2 acquisitions. Note a few important results:

- The majority of the distortion was outside of the brain area (note the higher CCC’s over the brain compared to over the head)

- B0 based methods performed as well as registration based methods

- The spatial resolution of the B0 maps did not impact results significantly, enabling one to proceed with brief B0 mapping protocols

- While all correction methods reduced distortion, the RPG methods performed better than B0 or registration-based correction Moreover, while the two RPG methods were comparable on average, RPG-1 performed more consistently than RPG-2 in the high distortion case (R=1); the difference between the highest and lowest CCC in our 5 volunteers were 0.04/0.09 for RPG-1/RPG-2 respectively. Moreover, our implementation of RPG-1 resulted in an average 50 second processing time, shorter than the ~4min processing speed of RPG-2.


The performance of four distortion correction methods was assessed in a cohort of normal volunteers. While all approaches reduced distortion, it was found that the RPG methods were inherently better than registration or B0-based approaches. It is likely the limited degrees of freedom of the registration approach could not account for the localized magnetic field inhomogeneities of the human brain. Similarly, the need to extrapolate the B0 maps in the distorted EPI space also led to errors/decreased performance of the B0-mapping approach. It is hence suggested for the prospective DTI studies to acquire the additional RPG images; at one TR additional scan time, they add little burden to the study duration and can significantly improve image quality.


No acknowledgement found.


[1] Zeng et al, Magn Reson Med. 2002,48(1):137-46.

[2] Gholipur et al, 33 Annual Int Conf IEEE EMBS, 2011, 6997-7000

[3] Wu et al, Med Image Comput Comput Assist Interv. 2008;11(Pt 2):321-9.

[4] Jezzard P, et al, Magn Reson Med. 1995 Jul;34(1):65-73.

[5] Jenkinson et al, NeuroImage 2002, 17(2), 825-841.

[6] Holland D et al, NeuroImage 2010, 50, 175-183

[7] Andersson et al, NeuroImage 2003, 20, 870-888. [8] Iglesias et al, IEEE TMI 2011, 30(9), 1617-1637.


Figure 1: Exemplary performance (R=1) of the four correction algorithms in one volunteer/ two separate slices

Figure 2: Average cross correlation coefficients for 5 volunteers for all acquisitions and correction methods

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