Should volumetric, slice-wise or non-linear registration be used for motion correction of fMRI data?
Malte Hoffmann1 and Stephen J Sawiak1,2

1Wolfson Brain Imaging Centre, University of Cambridge, Cambridge, United Kingdom, 2Behavioural and Clinical Neuroscience Institute, University of Cambridge, Cambridge, United Kingdom

Synopsis

Motion corruption leads to major artefacts in fMRI which are damaging in studies. Most processing pipelines employ linear registration of the brain volume at each time point. As slices are acquired individually, however, it is possible for each slice to require a different transformation. Here, we compared the efficacy of linear volume registration, independent slice-based registration and volume non-linear registration for retrospective correction of fMRI acquisitions using data from forty non-cooperative patients with substantial motion.

Purpose

Motion artefacts severely compromise fMRI analysis and most processing pipelines include a linear volume registration step. Since the 2D EPI data are acquired in slices, can correcting each slice within a volume individually lead to better motion correction? We compare the effects of slice-based registration and non-linear registration between volumes for each frame to assess whether these methods are useful. We did this in a set of particularly motion-prone patients with varying degrees of disorders of consciousness whose motion patterns have been previously studied.1

Methods

Patients with various degrees of disorders of consciousness (n=40) were scanned on a Siemens (Erlangen, Germany) 3T system. A total of 59 EPI time series with 300 frames each was acquired (TR/TE 2000/30ms, voxel size 3×3mm2, matrix size 64×64, 32 3mm slices with 0.75mm gap). All patients gave informed consent, and the study was approved by the local research ethics committee.

Slice-based registration

The slices of each volume of a time series were registered individually within the first volume of the series. Initially, each slice was compared to the corresponding slice of the reference volume. The latter was subsequently moved within the reference volume to find an optimum rigid-body transformation. Optimisation was performed minimising the sum of squared slice differences (SSD) with a six-dimensional gradient descent method.

We evaluated the accuracy of the registration by corrupting a low-motion volume with synthetic motion. Extreme motion parameters were drawn from a normal distribution with 1mm/° standard deviation (STD) for each of the three translations and rotations respectively. Each slice of the corrupted volume was generated by applying the resulting transform to the original voxel coordinates of the slice, and interpolating at the new location.

Data quality was evaluated with three different measures: (1) We determined the total SSD between all volumes of a time series and the first. (2) The overlap of CSF with CSF of the first volume was computed by thresholding each image at 60% of the maximum intensity within the time series and counting matching voxels. (3) We evaluated temporal standard deviation (tSTD) by averaging the STD of each voxel time series in a dataset.

We compared the performance of volumetric image registration with spm_realign2, slice-based motion correction and non-linear registration with FNIRT3. All volumes were registered with the first of each time series.

Results

Figure 1 shows a low-motion dataset before and after simulation of between-slice motion. Despite injected motion exceeding five times the motion detected in the worst dataset, slice-based registration successfully restored the dataset (see Figure 1C). The deviation between simulated and detected motion parameters is given in Table 1.

Figure 2B shows the typical slice displacement from a patient dataset, while the slice alignment in the absence of motion is illustrated in Figure 2A. Quality measures for the different registration methods are compared in Table 2. Both slice-based and non-linear correction improved data quality across all categories.

Discussion

Both slice-based and non-linear registration had a significant (p < 0.05, Student’s t-test) improvement over volume methods to tSTD. There was no significant different in the tSTDs between these methods. Motion that occurs during the acquisition of single image slices in EPI (typical acquisition times of ~50ms) is a second-order effect that is often neglected in fMRI analysis pipelines. Here, we showed that data quality can be improved substantially over volume registration if a slice-based method is used. While registration of individual slices models the acquisition scheme, it does not take other second-order effects such as spin history and coil sensitivity profiles into account. This is an explanation why the more general non-linear registration achieves further improvements.

Conclusion

If prospective motion correction is not an option, retrospective non-linear registration of each volume of a time series can be used to significantly improve data quality metrics. The similarity in improvement seen between slice-based and non-linear registration techniques suggests that the majority of non-linear deformations are the result of inter-slice motion. This being the case, incorporating slice-based transforms into the regularisation of the non-linear transforms may be an important contribution to avoid overfitting in non-linear image registration of fMRI time series.

Acknowledgements

The authors would like to thank the Cambridge Research into Impaired Consciousness Group for sharing data.

References

1. Hoffmann M et al. A survey of patient motion in disorders of consciousness and optimization of its retrospective correction. Magn Reson Imaging 2015;33(3):346-50.

2. Friston KJ et al. Spatial registration and normalization of images. Hum Brain Map 1995;2:165-189.

3. Jenkinson M et al. FSL. NeuroImage 2012;62(2):782-90.

Figures

Figure 1 A low-motion data set (A) before and (B) after simulation of between-slice motion. Rigid-body motion parameters were drawn from a normal distribution with 1mm/degree standard deviation (more severe than any motion trajectory we have observed). (C) The data set after slice-based registration showing successful reconstruction of the brain.

Figure 2 (A) Stack of image slices in perfect alignment. Only the central half of all slices is shown, and the slice gap was increased for illustration purposes. (B) Illustration of typical slice displacement from a patient data set (volume-based motion was removed prior to slice registration).

Table 1 Mean and standard deviation of the mean (SEM) for the differences between simulated and detected rigid-body motion parameters across 30 slices.

Table 2 Mean sum of squared differences (SSD), CSF overlap with the first volume of each time series and temporal standard deviation (tSTD) for the different motion correction strategies.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1746