Introduction

A retrospective motion correction technique for brain MRI images has been proposed by Gallichan et al.¹ to detect and correct non-deliberate motion during high resolution imaging, that uses the natural sparsity of fat-images to apply the GRAPPA² parallel imaging technique to acquire highly accelerated navigator fat-volumes (FatNavs). However, GRAPPA reconstruction is not expected to perform well on FatNav volumes in the presence of strong motion due to the mismatched calibration data acquired once at the start of the scan. The compromised GRAPPA reconstruction is then expected to lead to motion-parameter misestimation. This study aims to assess FatNav accuracy in the presence of large changes in head position, analysing the relationship between the extent of the motion and the expected degradation in image quality.

Methods

Three datasets of 3D MPRAGE³ brain images have been acquired without deliberate motion on three different subjects using a Prisma scanner (Siemens Healthcare, Erlangen, Germany). During each acquisition, 3D FatNav volumes at 4mm isotropic resolution have been acquired as navigators. Autocalibration lines (ACS) for the FatNavs, acquired as a reference at the beginning of the scan, have been used for reconstruction of each FatNav. 225 rigid motion parameters evenly spaced between 0-20deg and 0-40mm have been applied to the first FatNav volume in the image domain using SPM (Statistical Parametric Mapping) realign tool. Negative displacement has been applied along the z-axis only, considering symmetry for the rotations and displacement along the other directions. FatNav volumes were then undersampled using an acceleration factor R=16 (4x4) and re-reconstructed using GRAPPA, to simulate the final corrupted FatNav volume. By re-registering the corrupted and the reference FatNav volume using SPM, the mis-estimation of motion parameters was found and averaged between datasets for the full range of parameters considered. Applying linear interpolation, it is possible to use this data to estimate the ‘residual motion’ that would result from any given ‘true motion’ curves, where the residual motion curves represent the apparent head motion after FatNav correction. Smoother and rougher motion curves with root-mean-square (RMS) values between 0-20deg and 0-40mm were randomly generated 100 times for each dataset. After the interpolation step, the corresponding residual motion curves were applied to the original brain images without deliberate motion using RetroMoCoBox (https://github.com/dgallichan/retroMoCoBox), to simulate the expected degradation in image quality that would be expected if FatNavs were used. Two observers evaluated each image with a 1-4 rating scale, from 4=no visible motion artefacts to 1=severe motion artefacts. Image quality was calculated using the gradient entropy, which has been found by McGee et al.⁴ as the best metric to measure the quality of shoulder MRI images. A linear regression model was fit between the RMS motion and the resulting image gradient entropy. Multinomial Logistic Regression between image evaluations and gradient entropy was performed, revealing how the probability of falling in an evaluation category changes based on entropy values. Predicted FatNav correction-quality was assessed by calculating the amount of motion that corresponds to each category boundary, combining logit and linear regression information. The level of motion corresponding to each category boundary has been evaluated twice: once assuming that the simulated motion was the true motion (i.e no FatNav correction applied) and once assuming that the simulated motion was the residual motion following FatNav correction. This allows a comparison of the expected increase in motion that can be tolerated when FatNav correction is used.

Results

Figure 1 compares FatNav volumes before and after the GRAPPA ‘re-reconstruction’ step for different amounts of motion, showing that parallel imaging artifacts increase for larger changes in the head position. The multinomial logistic regression shows how the probability distribution changes based on the gradient entropy values. It was found that the relationship between rotational and translational motion and the gradient entropy is well described by a linear regression model, where rotational and translational motion are averaged separately along each direction. Figures 2 shows an MPRAGE image, where an example is shown falling into each image evaluation category. Related real and ‘residual motion’ are reported in Figure 3 and 4. Motion level at each category boundary is reported in Figure 5: FatNavs show a high tolerance to motion, whereas the image quality decreases faster for images without motion correction.

Discussion

Although FatNavs have been demonstrated to correct for a large scale of motion, in the presence of strong head position changes the GRAPPA reconstruction of the FatNavs themselves will be compromised, leading to misestimation of the motion parameters. Nonetheless, our data suggest that motion that would be sufficient to lead to a category-2 rating (strong artifacts) can typically be corrected with FatNavs to a level corresponding to category-4 (no noticeable artifact). If a particular subject group are likely to move more than this, it may be necessary to adapt the FatNav acquisition to make it more robust to large motion, which is expected to be possible if the GRAPPA weights are dynamically updated during the scan.

Conclusion

FatNavs have been shown to be able to correct for a wide range of motion levels, both for smooth and rough kinds of motion. Even greater robustness is expected by updating ACS lines throughout the scan.

Acknowledgements

No acknowledgement found.