2867

Pre-processing of high-resolution gradient-echo images for laminar fMRI applications
Patricia Pais-Roldan1, Seong Dae Yun1, and Jon N Shah1
1Forschungszentrum Juelich, Juelich, Germany

Synopsis

Due to their rapid acquisition and high contrast-to-noise-ratio, MRI sequences based on gradient-echo (GE) schemes, e.g., gradient-echo echo-planar-imaging (GE-EPI), are the most commonly used method in human functional MRI. However, their enhanced sensitivity to veins restricts their use in laminar fMRI due to poor signal localization, i.e., venous bias. Here, we investigated the spatial specificity of high-resolution GE signals, pre-processed with ten different approaches. Removal of motion parameters, physiological signals and non-GM tissue contributions, as well as regression of the pre-processed phase image from the magnitude image, significantly increased the spatial specificity of GE-fMRI in resting-state and task paradigms.

Introduction

With the advent of sub-millimetre resolution functional magnetic resonance imaging (fMRI), the function of the cerebral cortex is increasingly being assessed at different depths. Although several studies have investigated human laminar responses to particular tasks (e.g., [1-3]); whole-brain resting-state fMRI (rs-fMRI) remains challenging due to the limited field of view afforded by most high-resolution sequences. Gradient-echo (GE) fMRI schemes offer the possibility of achieving very high BOLD signal sensitivity, which facilitates investigations of the cortical thickness in the whole brain. However, GE sequences are sensitive to magnetic field inhomogeneities and introduce bias towards the superficial layers of the cortex due to the influence of ascending and pial veins. Task-fMRI obtained with GE sequences can be partially corrected by applying existing methods [4-9]; in contrast, no specific method has been optimized to analyse laminar rs-fMRI. Here, we applied ten different pre-processing pipelines to whole-brain, sub-millimetre GE fMRI (0.63×0.63×0.63mm) to study their effect on the spatial localization of GE-signals.

Methods

One rs-fMRI scan (~10 min) and one task-fMRI scan (~8 min) were obtained from 13 volunteers using a Siemens Magnetom Terra 7T scanner with a 1-channel Tx / 32-channel Rx Nova Medical head coil. The task consisted of finger tapping following the protocol: 21s-ON, 21s-OFF x 12. One volunteer performed an additional task consisting of touching the thumb with the index finger, i.e., adding somato-sensation. fMRI was obtained with GE EPIK (GE EPI with keyhole) [10]. TR/TE = 3500/22 ms, FA = 85°, partial Fourier = 5/8, 3-fold in-plane/3-fold inter-plane (multi-band) acceleration, matrix = 336 × 336 × 123 slices, voxel size = 0.63 × 0.63 × 0.63 mm3. Pre-processing was performed with ten different pipelines (Fig. 1a). The most complete pipeline included slice-timing correction, realignment, regression of motion parameters, regression of the pulse and respiration (RETROCOR [11]), regression of the mean signal of the cerebrospinal fluid (CSF) and white matter (WM), and regression of the pre-processed phase image [12]. Smoothing was also added to one pipeline for comparison. The effects of pre-processing were evaluated in the resting-state data in terms of mean correlation with noise, power of low-frequency fluctuations in the grey matter (GM) and signal homogeneity (a joined measure of temporal correlation and coherence in line crossing the cortical ribbon) (Fig. 2). The specificity of the corrected evoked responses was assessed by computing line-activation profiles and the t-statistic GM/CSF ratio (Fig. 3).

Results

High-resolution GE-data which was only subjected to realignment resulted in high levels of noise contamination (Fig. 1b). The addition of a band-pass filter, but not motion regression, significantly decreased the influence of non-neuronal signals. In contrast, voxel-wise regression of the phase signal, previously pre-processed in the same way as the magnitude image (here, slice-timing correction, realignment and filtering) resulted in a significant decrease in signal contamination (pre-processing pipeline #6 vs. #2 in Fig. 1b). Application of the phase-based regression at later stages did not significantly reduce noise. Regression of the physiological recordings and added CSF/WM regression achieved maximum signal cleaning. More complete pipelines also resulted in higher signal power in the cortical GM (Fig. 2b). The addition of smoothing increased the power within the GM, presumably due to averaging with CSF voxels, but led to the highest signal homogeneity (blurring) (see #9 vs. #10 in Fig. 2c). Signal individuality, e.g., low similarity between neighbour voxels, was maximized by regressing out motion, physiological and CSF/WM tissue signals (see #5 and #9 in Fig. 2c). Signals evoked in the pre-central gyrus by two different tasks were best distinguished, based on laminar profiles, with pre-processing pipelines #4, 5, 8 and 9, i.e., those including more complete regression steps which identified an activation peak in deep and intermediate depths in response to a motor task and motor + sensory task, respectively (Fig. 3a). The application of smoothing eradicated the definition of deeper peaks, leaving only the typical GE-bias in superficial layers, near the CSF. Group analysis demonstrated the significant effect the phase-regression step had on the localization of evoked signals in the GM, i.e., a higher average t-statistic in the GM with respect to CSF (Fig. 3c).

Discussion

Our results demonstrate that, when subjected to certain pre-processing steps, high-resolution GE-fMRI can be used to track laminar responses. In general, we observed higher signal independence when more pre-processing steps were applied to the data; however, the smoothing step deteriorated the voxel heterogeneity. The use of phase-regression, a method introduced by Menon [12], produced a significant improvement in terms of signal cleaning, especially when applied in little-pre-processed data; however, its addition to already-highly-pre-processed data did not significantly modify signal individuality or contamination, although it did improve the localization of evoked responses in the GM. Interestingly, the correction of partial volumes by removing CSF/WM did not significantly improve the specificity of resting-state signals, i.e. did not alter the homogeneity levels, possibly due to lower partial volume effects in high-resolution images. The fMRI sequence employed here is known to have a robust performance against field distortions, therefore, common EPI sequences may require additional field correction steps [13].

Conclusion

Laminar-fMRI, especially resting-state applications requiring whole-brain coverage, could benefit from an optimal pre-processing (e.g. #8-#9), to exploit the benefits of GE-sequences.

Acknowledgements

We thank Elke Bechholz for technical support, Claire Rick for proof-reading and the fMRI volunteers for their excellent cooperation.

References

1. Kashyap, S., et al., Resolving laminar activation in human V1 using ultra-high spatial resolution fMRI at 7T. Sci Rep, 2018. 8(1): p. 17063.

2. Huber, L., et al., High-Resolution CBV-fMRI Allows Mapping of Laminar Activity and Connectivity of Cortical Input and Output in Human M1. Neuron, 2017. 96(6): p. 1253-1263 e7.

3. Polimeni, J.R., et al., Laminar analysis of 7T BOLD using an imposed spatial activation pattern in human V1. Neuroimage, 2010. 52(4): p. 1334-46.

4. Havlicek, M. and K. Uludag, A dynamical model of the laminar BOLD response. Neuroimage, 2020. 204: p. 116209.

5. Kay, K., et al., A temporal decomposition method for identifying venous effects in task-based fMRI. Nat Methods, 2020. 17(10): p. 1033-1039.

6. Fracasso, A., et al., Laminar imaging of positive and negative BOLD in human visual cortex at 7T. Neuroimage, 2018. 164: p. 100-111.

7. Kashyap, S., et al., Impact of acquisition and analysis strategies on cortical depth-dependent fMRI. Neuroimage, 2018. 168: p. 332-344.

8. Heinzle, J., et al., A hemodynamic model for layered BOLD signals. Neuroimage, 2016. 125: p. 556-570.

9. Markuerkiaga, I., M. Barth, and D.G. Norris, A cortical vascular model for examining the specificity of the laminar BOLD signal. Neuroimage, 2016. 132: p. 491-498.

10. Yun, S., P. Pais-Roldán, and N.J. Shah. Detection of Cortical Depth-dependent Functional Activation using Whole-brain, Half-millimetre Resolution EPIK at 7T. in ISMRM. 2020. Paris, France.

11. Glover, G.H., T.Q. Li, and D. Ress, Image-based method for retrospective correction of physiological motion effects in fMRI: RETROICOR. Magn Reson Med, 2000. 44(1): p. 162-7.

12. Menon, R.S., Postacquisition suppression of large-vessel BOLD signals in high-resolution fMRI. Magn Reson Med, 2002. 47(1): p. 1-9.

13. Schallmo, M.P., Weldon, K.B., Sponheim, S.R., Olman, C.A., Assessing methods for geometric distortion compensation in 7T gradient echo fMRI data. bioRxiv, 2020.

Figures

Figure 1. Pre-processing pipelines and their influence on signal noise. a) Pre-processing steps and pipelines (numbered in blue). b) Average correlation between voxels in the fMRI images pre-processed with ten different approaches, the time course of the noise sources, i.e., motion parameters, physiological parameters, and mean time courses of CSF and WM. Error bars represent standard error of the mean. Different letters indicate significantly different groups in terms of their correlation with noise.

Figure 2. Signal amplitude and independence in resting-state data. a) The image shows the location of nine voxels crossing the cortical ribbon (“cross-cortex line”) that were used in the successive analysis. b) Power of low-frequency fluctuations at different points of the cross-cortex line calculated for ten different pre-processing pipelines. c) Mean homogeneity, assessed as correlation and coherence between pairs of voxels in the cross-cortex-line. Error bars represent standard error of the mean. Different letters indicate significantly different groups.

Figure 3. Identification of evoked responses. a) Line activation profiles, calculated as the mean beta-value across 20 lines (see inset on the bottom left), for two task-fMRI scans involving either motor-only (M) or motor and sensory processing induced by touch (M+T), computed after pre-processing with ten different pipelines. The dashed lines indicate the intensity of the mean functional image. b) ROIs selected for group analysis. c) The GM to CSF ratio of the t-statistic. Asterisks indicate significant differences between pre-processing pipelines (p<0.05).

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
2867