Synopsis

Here, we present a novel approach of utilizing Boundary-Based Registration for realigning submillimetre 7T fMRI time series. We collected fMRI data from 6 human participants and processed the data using either standard rigid body realignment using SPM or our BBR realignment method. We compared the two pre-processed datasets with multiple metrics (tSNR, fCNR and percentage of variance explained by the model) and show that realigning using BBR consistently outperforms conventional methods.

Introduction

As fMRI progresses towards higher fields and submillimetre resolutions, accurate realignment of the time series is critical for any form of analysis. In particular, Multi-voxel Pattern Analysis (MVPA) techniques have been shown to be sensitive to poor voxel-wise realignment across time and small realignment inaccuracies can easily result in masking of real effects or increasing false positives^1,2.

Here, we demonstrate a novel application of Boundary-Based Registration (BBR) to realign a set of submillimetre 7T fMRI time series. BBR³ was originally developed to co-register images across different imaging modalities or functional contrast but to the best of our knowledge, has not been used to realign time series data. We compared the BBR realignment results with those of a standard SPM 12 pipeline involving volumetric, rigid-body realignment followed by structural-functional co-registration, and demonstrate a benefit of using BBR for realignment.

Methods

Six participants (two female) were scanned across two sessions for this study. Within each session, we carried out a MP2RAGE scan and eight runs of EPI with a simple visual task. All data were acquired on a Siemens 7T Terra scanner with a Nova Medical 1TX/32RX Head Coil. The EPI parameters were: 0.8mm isotropic voxels, TR=2390ms(2440ms for two participants), TE= 24ms(24.4ms for two participants), FA=80^o, Matrix size= 200*168*84, ToA=~11mins.

The task followed a block design where two stimuli were presented at opposing quadrants and participants were required to make a same/different judgement. The four conditions (Figure 1) were presented in a randomized order over four runs (20 16-second blocks per run).

Freesurfer 6.0.0 was used for surface reconstruction. Images first underwent slice time correction in SPM 12. The entire set of images then underwent either standard rigid body realignment in SPM using sum-of-squares cost function, followed by co-registration using Normalized Mutual Information, or BBR realignment. Note that the BBR realignment provides a one-step correction of both head motion and registration with the structural.

In BBR realignment, the mean fMRI image was first registered to the structural using BBR and the registration matrix is saved. Each volume from every timepoint was then registered to the structural individually using BBR. The mean registration matrix was used as the initial seed to reduce computational time. This process aligns all functional volumes to the structural and thus, also to each other.

For temporal tSNR, analysis was restricted to the grey matter voxels, segmented using Freesurfer. The tSNR for each voxel was obtained by dividing the mean voxel intensity across the entire time course by the standard deviation of the voxel intensity.

fCNR was also calculated for both retinotopic regions of interest (ROIs) (V1, V2, V3) and categorical ROIs (scene-selective transverse occipital sulcus [TOS], parahippocampal place area [PPA], occipital face area [OFA], fusiform face area [FFA]) that were obtained from a previous functional localizer session. The fMRI data were fitted to a general linear model of the experimental design, and the fCNR for each voxel was obtained by dividing the difference between the two beta values that form the contrast of interest by the standard deviation of the residuals from the model fit. The contrast of interest was between the different stimulated quadrants for the retinotopic ROIs and between the different stimulus categories for the categorical ROIs. We also calculated the percentage of variance explained (PVE) by the model fit.

Results and Discussion

The tSNR of the BBR realignment data shows a consistent increase across all six participants (Figure 2, Panel A). Voxel-wise comparisons of tSNR showed that the greatest benefit of BBR realignment occurs close the brain surface (Figure 2, Panel B). This behaviour is expected since BBR is driven by aligning the images along the grey matter boundaries. Similarly, we note a consistently higher average fCNR and PVE across all ROIs when using BBR realignment (Figure 3).

We believe that realignment using BBR generates higher tSNR and fCNR due to a multitude of factors, including: higher contrast of the structural data, higher degrees of freedom and utilizing grey matter boundaries. The relative contribution of these factors will be explored in future work. Realignment using BBR also removes the need for an added registration step if the structural data is utilized for further processing (e.g. laminar layer segmentation or projecting the voxels into surface space).

Conclusion

Acknowledgements

References

1. Huang, P. et al. Prospective motion correction improves the sensitivity of fMRI pattern decoding. Hum. Brain Mapp. 39, 4018–4031 (2018).

2. Field, A. S., Yen, Y. F., Burdette, J. H. & Elster, A. D. False cerebral activation on BOLD functional MR images: Study of low-amplitude motion weakly correlated to stimulus. Am. J. Neuroradiol. 21, 1388–1396 (2000).

3. Greve, D. N. & Fischl, B. Accurate and Robust Brain Image Alignment using Boundary-based Registration. Neuroimage 48, 63–72 (2009).