Sydney Bailes1 and Laura D. Lewis1
1Biomedical Engineering, Boston University, Boston, MA, United States
Synopsis
Given
the substantial variations in the shape and timing of the hemodynamic response
function (HRF) across the brain, it is critical to develop methods to
characterize these variations for proper interpretation of the BOLD fMRI signal.
Here, we identified significant differences in spectral properties of resting
state fMRI signals between voxels with fast and slow hemodynamics. We found
that these spectral properties can be used to classify fast and slow voxels,
suggesting that information from the resting state can provide a way to understand
and predict the temporal dynamics of the HRF across the brain.
Introduction
Both the shape and timing
of the hemodynamic response function (HRF) vary substantially across the brain1,2, and proper
characterization of these variations is critical for the interpretation of the
BOLD signal obtained from fMRI3. Additionally, as
techniques for fast acquisition of fMRI have become more prevalent4, the ability to
assess differences in the HRF across the brain becomes even more critical5 in order to
determine whether fMRI timing differences result from neural or vascular
origins. Previous work has implemented breath hold6 or visual
stimulus tasks7 to estimate
voxel-wise differences in vascular latencies; however, these methods require
additional scanning which can be difficult for certain studies or populations.
We aimed to test if we can use information from the spectrum of resting state
fMRI to characterize vascular delays across the brain. Since resting state
scans are acquired in most studies and require no task performance on the part
of the subject, using information from the resting state spectrum would provide
a simple and effective way of understanding the temporal dynamics of the HRF
across different regions of the brain. Methods
First, we performed
simulations to determine if different temporal properties of the HRF would
result in different properties of the spectrum of a voxel’s response to an identical
stimulus. Six different HRFs (Fig. 1a)
were defined each with a different time-to-peak (TTP), full width at half
maximum (FWHM), and peak percent signal change. These properties were drawn
from previous work characterizing varying HRF temporal dynamics7. We calculated the
predicted fMRI response to neural activity at various frequencies by convolving
oscillating stimuli (ranging 0.1-0.5 Hz) with the six defined HRFs. We also
calculated the spectrum of the HRF itself.
Three subjects were
scanned on a 7T Siemens scanner across eight sessions, each consisting of resting
state runs and runs with a flickering checkerboard visual stimulus with
oscillating luminance contrast (frequency ranging 0.025-0.2 Hz, Fig. 2a-b). Functional runs consisted
of 15 oblique slices positioned to include primary visual cortex (V1) acquired
as single-shot gradient-echo blipped-CAIPI SMS-EPI8 with 2 mm isotropic
resolution (R=2 acceleration,
MultiBand factor=3, matrix=120×120, CAIPI shift=FOV/3, TR=227 ms, TE=24 ms,
echo-spacing=0.59 ms, flip angle=30°). We identified voxels in V1 that were
significantly driven by the stimulus using a functional localizer run and then
calculated the phase delay of the response for each voxel using the arctangent
of the sine and cosine regressor estimates. Then, we examined the resting state
spectra of these voxels. In particular, we looked at the slope of the spectrum
under 0.2 Hz, the average power under 0.1 Hz, and the exponent of the aperiodic fit (Fig. 3a). We trained k-nearest neighbors (KNN) models (N=200, 5-fold cross
validation) within and across subjects to classify voxels based on these
spectral properties. For the across subject model, the predictors were
normalized by the mean of all voxels. Results
Examining the simulated resting state
spectra, we found that HRFs with faster dynamics exhibited less power in the
low frequency bands compared to slower HRFs, but a shallow decline at higher
frequencies (Fig. 1b). The slower
HRFs had larger response amplitudes to lower stimulus frequencies and steeper
drop-off as the stimulus frequency increased. This difference is most prominent
under 0.2 Hz, after which the faster HRFs tend to have larger response
amplitudes (Fig. 1c).
We
next identified V1 voxels as fast or slow based on their responses to the
oscillating visual stimulus (Fig. 2).
Each feature of the resting state spectra showed significant differences
between fast and slow voxels (Wilcoxon rank sum test, ): the slopes for all 8/8
sessions and both power and aperiodic exponent for 7/8 sessions (one session
had for power and for exponent) (Fig. 3b-d). We then tested whether these spectral features could
predict local hemodynamic delays. We trained a KNN classifier with N=200 to identify
fast and slow voxels within each subject and achieved accuracies of 78.5%,
67.1%, and 55.96%. After normalizing each metric across subjects and retraining
a KNN classifier we were able to achieve a classification accuracy of 71.5%. Discussion
Our simulation results clearly
show that differences in the temporal dynamics of the HRF will change the
frequency content of the fMRI response. Specifically, slower HRFs dynamics had
larger response amplitudes in the low frequency band but the response quickly
attenuates at higher frequencies. Conversely, faster HRFs had smaller response
amplitudes in the low frequency band and a shallower slope. Notably, these
differences predicted by simulations are also reflected in the resting state data:
resting-state spectral features were found to be significantly different for
voxels with fast versus slow hemodynamics. This allowed us to classify voxels
as fast or slow both within and across subjects. Conclusion
Our
results demonstrate that temporal properties of the HRF affect the spectral
features of spontaneous fMRI signals. We find that this insight enables us to
distinguish voxels that exhibit faster or slower hemodynamic responses using
only resting state information. This finding can allow researchers to better
understand the temporal properties of the HRF across voxels, which is crucial
for accurate fMRI analyses, especially as we investigate faster neural
dynamics. Acknowledgements
This work was funded by the NIH Grant R00-MH111748 and the NIGMS Training Program 5 T32 GM008764-20. References
- Aguirre, G. K., Zarahn, E. & D’Esposito, M.
The Variability of Human, BOLD Hemodynamic Responses. NeuroImage 8,
360–369 (1998).
- de Zwart,
J. A. et al. Temporal dynamics of the BOLD fMRI impulse response. NeuroImage
24, 667–677 (2005).
- Handwerker, D. A., Ollinger, J. M. &
D’Esposito, M. Variation of BOLD hemodynamic responses across subjects and
brain regions and their effects on statistical analyses. NeuroImage 21,
1639–1651 (2004).
- Feinberg, D. A. &
Setsompop, K. Ultra-fast MRI of the human brain with simultaneous multi-slice
imaging. J. Magn. Reson. San Diego Calif 1997 229, 90–100 (2013).
- Lewis, L. D., Setsompop, K., Rosen, B. R. &
Polimeni, J. R. Stimulus-dependent hemodynamic response timing across the human
subcortical-cortical visual pathway identified through high spatiotemporal
resolution 7T fMRI. NeuroImage 181, 279–291 (2018).
- Chang, C.,
Thomason, M. E. & Glover, G. H. Mapping and correction of vascular
hemodynamic latency in the BOLD signal. NeuroImage 43, 90–102
(2008).
- Siero, J. C. W., Petridou, N., Hoogduin, H.,
Luijten, P. R. & Ramsey, N. F. Cortical depth-dependent temporal dynamics
of the BOLD response in the human brain. J. Cereb. Blood Flow Metab. Off. J.
Int. Soc. Cereb. Blood Flow Metab. 31, 1999–2008 (2011).
- Setsompop,
K. et al. Blipped-controlled aliasing in parallel imaging for simultaneous
multislice echo planar imaging with reduced g-factor penalty. Magn. Reson.
Med. 67, 1210–1224 (2012).