Nuisance Regression of High-frequency FMRI Data: De-noising Can Be Noisy
Jingyuan E. Chen1,2, Hesamoddin Jahanian2, and Gary H. Glover1,2

1Electrical Engineering, Stanford University, Stanford, CA, United States, 2Radiology, Stanford University, Stanford, CA, United States

Synopsis

A growing number of studies using fast sampling have demonstrated the persistence of functional connectivity (FC) in resting state (RS) networks beyond the conventional 0.1 Hz. However, some RS studies have reported frequencies (e.g., up to 5 Hz) not easily supported by canonical hemodynamic response functions. Here, we investigated the influence of a common preprocessing step – whole-band (the entire frequency band resolved by a short TR) linear nuisance regression (LNR) – on RSFC. We demonstrated via both simulation and real data that LNR can introduce network structures in HF bands, which may largely account for the observations of HF-RSFC.

Introduction

A growing number of studies using fast sampling have demonstrated the persistence of functional connectivity (FC) in resting state (RS) networks beyond the conventional 0.1 Hz[1-3]. However, some RS studies have reported frequencies (e.g., up to 5 Hz[4]) not easily supported by canonical models of brain hemodynamic responses (the upper limit of which is about 0.3 Hz[5]). It is thus questionable whether the observed high-frequency (HF) RSFC originates from BOLD mechanisms identical to the usual low-frequency (LF) components[3], or could be caused by another mechanism such as flow[6] or proton density changes[7], or is artificially introduced by certain preprocessing steps. Here, we investigated the influence of a common preprocessing step – whole-band (the entire frequency band resolved by a short TR) linear nuisance regression (LNR) – on RSFC. We demonstrated via both simulation and real data that LNR can introduce network structures in HF bands, which may largely account for the observations of HF-RSFC.

Potential concerns with whole-band LNR

RS studies (< 0.1 Hz) commonly assume that various nuisance fluctuations (e.g., motion & physiological artifacts) are linearly superimposed on the spontaneous neural-related activities, and remove them from the raw dataset via linear regression (Fig. 1(a)). However, if both the observations and regressors are heavily dominated by LF fluctuations, fitting parameters in the regression model will be driven by the slow fluctuations in the observations/regressors. Therefore, any HF components present in the regressors will be introduced to observations as additional variance (Fig. 1(b)). Furthermore, it has recently been demonstrated that the regressed noise (fluctuations projected out of the original data) in the conventional LF band contains prominent network structures[8]. The fitting parameters, which carry the network structure information in the LF band, may thus introduce spurious variance with network structure into the HF band of the fMRI time series (Fig. 1(b), red).

Real dataset

Ten-minute RS scans from 10 healthy subjects aged 36 ± 12 yrs (4 females) were collected at 3T (GE Signa 750, 32 channel coil, simultaneous multi-slice EPI with blipped CAIPI sequence[9], TR/TE = 350/30 ms, multiband acceleration factor of 6, CAIPI FOV shift factor of 3, flip angle = 40o, 30 slices, voxel size 3.14 × 3.14 × 4 mm3). After correcting for susceptibility-induced distortions with FSL TOPUP toolbox (using an additional 7s scan with reversed phase encoding directions), further basic preprocessing steps included slice timing correction, removal of scanner drifts and physiological fluctuations synchronized with cardiac/respiratory cycles using RETROICOR[10]. The subjects’ data were normalized to MNI template for the ensuing analysis.

Simulation

The simulation goal was to examine whether LNR could yield significant HF-RSFC in a ‘dummy’ dataset, which contained no network structures in the HF bands. The ‘dummy’ dataset was created by eliminating any possible HF correlations in the real dataset (post basic preprocessing). Specifically, we took the Fourier transform of each voxel’s time series for each subject, scrambled the phases of components above 0.2 Hz, and then inversely Fourier transformed back to the temporal domain. RSFC of the created ‘dummy’ dataset was identical with the real dataset below 0.2 Hz, but contained no structured patterns above 0.2 Hz. Then, LNR was performed on the ‘dummy’ dataset using regressors estimated from the real dataset. RSFC results with respect to a seed in the posterior cingulate cortex (PCC) and visual cortex of the ‘cleaned’ dataset (post LNR) are shown in Fig. 2. Contrasting the network patterns of the ‘cleaned’ dataset and that of the ‘dummy’ dataset, we observed that prominent network structures were artificially introduced by LNR, even if only a small fraction of regressors were included (e.g., ‘white matter+CSF’).

Real dataset analysis

The influence of LNR on real data was also evaluated. Instead of the desired de-noising (reducing noisy fluctuations) of fMRI time series in the HF bands, LNR added additional variance from the HF components of the nuisance regressors to the real dataset (Fig. 3, > 0.4/0.8 Hz). The introduced fluctuations, although small, contained network structures closely resembling the LF-RSFC (Fig. 4(a-c)). The similarity between the network patterns of the ‘cleaned dataset’ (Fig. 4(d-e)) and those introduced by nuisance regression (Fig. 4(c)) implies that the observed HF-RSFC is more likely to be artificially introduced by LNR.

Conclusion

We have demonstrated that whole-band LNR can introduce artificial RSFC in the HF bands of fMRI signals. As LNR has been widely employed in studies of HF-RSFC, one should be cautious in drawing conclusions regarding neural bases of HF-RSFC in reported results.

Acknowledgements

This work is supported by NIH grants P41EB15891, R01NS066506, R01NS047607, R01DK092241, and GE Healthcare.

References

[1] Lee et al., Tracking dynamic resting-state networks at higher frequencies using MR-encephalography. NeuroImage (2013), 65: 216–222.

[2] Gohel and Biswal, Functional integration between brain regions at rest occurs in multiple-frequency bands. Brain Connectivity (2015), 5: 23-24.

[3] Chen and Glover, BOLD fractional contribution to resting-state functional connectivity above 0.1 Hz. NeuroImage (2015), 107: 207-218.

[4] Lin et al., Significant feed-forward connectivity revealed by high frequency components of BOLD fMRI signals. NeuroImage (2015), 121: 69-77.

[5] Glover GH, Deconvolution of impulse response in event-related BOLD fMRI. NeuroImage (2015), 9: 416-429.

[6] Gao and Liu, Inflow effects on functional MRI. NeuroImage (2012), 62: 1035-1039.

[7] Figley et al., In contrast to BOLD: signal enhancement by extravascular water protons as an alternative mechanism of endogenous fMRI signal change. Magnetic Resonance Imaging (2010), 28: 1234-1243.

[8] Bright and Murphy, Is fMRI "noise" really noise? Resting state nuisance regressors remove variance with network structure. NeuroImage (2015), 114: 158-169.

[9] Setsompop et al., Blipped-controlled aliasing in parallel imaging for simultaneous multislice echo planar imaging with reduced g-factor penalty. Magnetic Resonance in Medicine (2012), 67: 1210-1224.

[10] Glover et al., Image-based method for retrospective correction of physiological motion effects in fMRI: RETROICOR. Magnetic Resonance in Medicine (2000), 44: 162-167.

[11] Chang et al., Influence of heart rate on the BOLD signal: the cardiac response function. NeuroImage (2009), 44: 857-869.

Figures

Fig. 1 (a) illustration of the LNR; (b) potential concerns with whole-band LNR.

Fig. 2 (a) RSFC with respect to a PCC seed; (b) RSFC with respect to a visual seed. (Group t-map, 10 subjects, ‘<0.2 Hz’: 0~0.2 Hz; ‘>0.4 Hz’: 0.4~1.4 Hz)

Fig. 3 The ratio between the variance of fMRI time series before and after LNR, averaged across 90 functional EOIs (mean and standard deviations across 10 subjects, '<0.2 Hz': 0~0.2 Hz; '>0.4 Hz': 0.4~1.4 Hz; '>0.8 Hz': 0.8~1.4 Hz).

Fig. 4 Illustration of RSFC (with respect to the PCC seed) introduced by LNR in real dataset



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
0062