3274

Grasping the noise – applying brain vasculature as a priori for sLFO reduction in the BOLD-fMRI signal and the impact on subject identifiability

Andrew Xie¹, Rémi Dagenais², Mary Miedema², Emad Askarinejad², and Georgios D. Mitsis¹
¹Bioengineering, McGill University, Montreal, QC, Canada, ²Biological and Biomedical Engineering, McGill University, Montreal, QC, Canada

Synopsis

Keywords: fMRI Analysis, fMRI, Physiological Denoising, Subject Identifiability

Motivation: Denoising systemic low-frequency oscillations (sLFOs) using global signal regression (GSR) can possibly impact the neural component of the BOLD fMRI signal.

Goal(s): Our goal was to use the spatial relationship between the sLFO component of the BOLD signal and brain vasculature to perform a less aggressive form of GSR.

Approach: We collected structural and functional images at 3T in ten subjects. We then used to temporal and spatial characteristics of the sLFOs to denoise the BOLD signal.

Results: The spatial correlation between the sLFOs and venograms confirmed their underlying vascular origin. The performance of our novel denoising technique still needs to be evaluated.

Impact: We propose a novel sLFO denoising method that uses the temporal and spatial patterns of physiological noise to preserve a larger fraction of the neural activity.

Introduction

A significant source of noise in the BOLD-fMRI signal inherently stems from systemic physiological signals, such as systemic low-frequency oscillations (sLFOs)^1,2. Although global signal regression (GSR) does reduce sLFOs, it may affect the neural component of the BOLD signal^3,4; a consensus of its usage has not been reached⁵. Moreover, the spatial extent of sLFO effect on the BOLD signal is likely associated with the underlying brain vasculature. As such, we propose the use of susceptibility weighted imaging (SWI) venograms as prior knowledge for the spatially specific denoising of sLFO BOLD effects. Based on previous reports that correcting for physiological noise increases subject identifiability, we also examined whether this would result in an increase in subject identifiability compared to GSR⁶.

Methods

Functional and structural images were acquired in 10 subjects at the Montreal General Hospital on a Siemens 3T Prisma (Erlangen, Germany) using a 64-channel head/neck coil and peripheral physiological recordings. The functional scan consisted of a 309 second resting-state whole brain accelerated EPI sequence. The structural images were acquired with a T1-MPRAGE sequence and a SWI sequence (see Fig. 1).
Extraction of the brains was performed using BET⁷. The BOLD data was minimally preprocessed using FSL⁸ (5mm gaussian kernel spatial smoothing, 100sec high-pass filtering, slice-timing correction, motion correction, EPI distortion correction, and registration to standard space). SWI images were preprocessed using CLEAR-SWI⁹, with the vasculature extracted using a pipeline based on Braincharter Vasculature Extraction¹⁰.
The sLFOs were extracted by fitting the respiratory flow (RF), heart rate variability (HRV), and arterial blood pressure (ABP) fluctuations on the global signal (GS) using subject-specific physiological response functions for all three signals, which were estimated using double gamma functions (see ¹for a detailed explanation). The sLFOs were then correlated voxel wise for each subject to obtain the sLFO maps.
Co-registration of the sLFO maps and the SWI onto the structural space was performed using FSL⁸ and in-house code. A box filter of 15x15x15 voxels was applied onto the vasculature maps to blur the venograms to account for BOLD signal leakage around the large veins¹¹.
To measure identifiability, parcellated BOLD data (Schaefer 300 17 networks¹²) was processed in four different ways (baseline, GSR, pGSR, and pGSR-seg. See Fig. 2 for details). Subject identification through static functional connectivity (FC) was performed 2000 times per subject, with a calculation using 30 randomly selected consecutive volumes. Identification was deemed to be successful if the correlation between the subject’s FC matrix and a specific connectome FC matrix belonging to the same participant was maximized¹².

Results

Spearman’s correlation analysis between the physio maps and venograms showed moderate positive average correlation (Fig. 3).
Our results showed an increase in mean identifiability between GSR and pGSR-seg (Fig. 4). However, the increase was not statistically significant. Moreover, repeating pGSR-seg with phase shuffled physiological data (pGSR-seg-surrogate) resulted in a similar mean identifiability compared to pGSR-seg. The identifiability when using the isolated sLFO was significantly higher than the baseline and pGSR identifiability.

Discussion

The BOLD signal is known to leak around the vessels¹¹; therefore, the box smoothing filter was employed to simulate this leakage. The correlation between sLFO maps and venograms overall confirms the vascular origin of sLFO effects on the BOLD signal. We believe that this may have important implications for the correction of physiological noise in the BOLD-fMRI, especially at higher field strengths where physiological noise is more important¹³, and whereby more accurate venograms can be obtained.
The increase in identifiability does suggest that using the pGSR-seg protocol may improve denoising performance⁶. However, with the surrogate and sLFO data showing such high identifiability, it also suggests that the regression may be also artificially creating fingerprints on the data. Replication of this experiment on a dataset with multiday scans (e.g. HCP dataset) and a sample size considerably higher than ten would likely provide more conclusive outcomes regarding the impact of pGSR-seg and whether these effects are subject-specific or scan-specific. Furthermore, our results confirms that identifiability has strong associations with not only neural signals, but also physiological noise⁶.

Conclusion

Developing a better understanding of the spatial link between sLFOs, brain vasculature, and the BOLD-fMRI can improve the interpretability of fMRI experiments. Fine tuning methodologies such as the proposed pGSR-seg may lead to improved processing pipelines that reduce their impact on neural signals. We aim to apply the pGSR-seg technique to larger, multiday datasets.

Acknowledgements

The authors would like to thank NSERC and FRQNT for funding Andrew and Rémi’s work on the project, respectively. Furthermore, the authors would like to acknowledge the Biosignals and Systems Analysis group for their insight in the project.

References

1. Kassinopoulos, M. & Mitsis, G. D. Physiological noise modeling in fMRI based on the pulsatile component of photoplethysmograph. NeuroImage 242, 118467 (2021).

2. Tong, Y., Hocke, L. M. & Frederick, B. B. Low Frequency Systemic Hemodynamic “Noise” in Resting State BOLD fMRI: Characteristics, Causes, Implications, Mitigation Strategies, and Applications. Front. Neurosci. 13, 787 (2019).

3. Li, J. et al. Global Signal Regression Strengthens Association between Resting-State Functional Connectivity and Behavior. NeuroImage 196, 126–141 (2019).

4. Hahamy, A. et al. Save the Global: Global Signal Connectivity as a Tool for Studying Clinical Populations with Functional Magnetic Resonance Imaging. Brain Connect. 4, 395–403 (2014).

5. Murphy, K. & Fox, M. D. Towards a consensus regarding global signal regression for resting state functional connectivity MRI. NeuroImage 154, 169–173 (2017).

6. Xifra-Porxas, A., Kassinopoulos, M. & Mitsis, G. D. Physiological and motion signatures in static and time-varying functional connectivity and their subject identifiability. eLife 10, e62324 (2021).

7. Smith, S. M. Fast robust automated brain extraction. Hum. Brain Mapp. 17, 143–155 (2002).

8. Jenkinson, M., Beckmann, C. F., Behrens, T. E. J., Woolrich, M. W. & Smith, S. M. FSL. NeuroImage 62, 782–790 (2012).

9. Eckstein, K. et al. Improved susceptibility weighted imaging at ultra-high field using bipolar multi-echo acquisition and optimized image processing: CLEAR-SWI. NeuroImage 237, 118175 (2021).

10. Bernier, M., Cunnane, S. C. & Whittingstall, K. The morphology of the human cerebrovascular system. Hum. Brain Mapp. 39, 4962–4975 (2018).

11. Turner, R. How Much Cortex Can a Vein Drain? Downstream Dilution of Activation-Related Cerebral Blood Oxygenation Changes. NeuroImage 16, 1062–1067 (2002).

12. Schaefer, A. et al. Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI. Cereb. Cortex 28, 3095–3114 (2018).

13. Hutton, C. et al. The impact of physiological noise correction on fMRI at 7T. NeuroImage 57, 101–112 (2011).

Figures

Figure 1: Functional and structural MRI acquisition parameters.

Figure 2: For each subject, four processing pipelines are used to build an FC connectome from 30 consecutive volumes randomly sampled for subject identifiability. The baseline identifiability accuracy is obtained from the minimally preprocessed BOLD signal. The GSR pipeline first regresses out the GS from each voxel. The pGSR pipeline removes the component of the GS linearly correlated with RF, HRV, and ABP from each voxel. Finally, the pGSR-seg applies pGSR independently for each tissue label (white matter (WM), gray matter (GM), cerebrospinal fluid (CSF), and veins).

Figure 3: Subject identifiability using 30 consecutive data points randomly sampled for each of the 2000 iterations. The identifiability for each subject is obtained as the average of the 2000 iterations for each processing pipeline tested. The resulting identifiability matrix (1 row per subject and 1 column for each processing pipeline tested) containing the identifiability accuracy is then used for the statistical analysis.

Figure 4: a), b) Representative subject-level venogram and sLFO maps, respectively. c) Overlay between subject-level venogram and sLFO map. d) Rank correlation results averaged over all subjects.

Figure 5: One-way ANOVA result (p=0.00036) with multiple comparisons (* p<0.05; ** p<0.01; *** p<0.001) for the identifiability accuracy (accuracy = #Identified/#identification). The pGSR-seg-surrogate pipeline is the same as pGSR-seg, but the phase of the physiological data was shuffled to create surrogate data. The sLFOs pipeline consisted of subtracting the minimally preprocessed data with the pGSR corrected data to isolate the contribution of the sLFOs.

Proc. Intl. Soc. Mag. Reson. Med. 32 (2024)

3274

DOI: https://doi.org/10.58530/2024/3274