Olivia Viessmann1, Jingyuan Chen1, Kawin Setsompop1,2, Lawrence L Wald1,2, and Jonathan R Polimeni1,2
1Athinoula A. Martinos Center for Biomedical Imaging, Harvard Medical School, Massachusetts General Hospital, Charlestown, MA, United States, 2Massachusetts Institute of Technology, Division of Health Sciences and Technology, Cambridge, MA, United States
Synopsis
BOLD fMRI signals vary with the orientation of the cortex to the
B0-field as extravascular susceptibility effects vary with the orientation of
the cortical pial vasculature. This creates regional tSNR biases with cortical folding.
Certain cortical folds are more homogenous across the population than others
and orientation variability across subjects should introduce tSNR variability
at the group-level. Here, we use HCP 3T rs-fMRI data to show that B0-orientation
contributes to within-subject tSNR bias and orientation variability contributes to tSNR
variance across subjects. We found that functional connectivity networks with more
perpendicular orientation exhibit higher tSNR and networks with high
orientation consistency have lower tSNR variability.
Introduction
The BOLD fMRI signal has been shown
to be strongly dependent on the angle$$$\,\theta\,$$$between the cerebral cortex
surface normal and the main magnetic field B0-axis1,2, as extravascular susceptibility effects scale with the orientation angle of the blood vessels relative to the B0-axis
by approximately cosine-squared. This results in regional differences in BOLD
amplitudes that vary with cortical orientation (Fig.1a). Recently, in a pilot 7T resting-state fMRI (rs-fMRI) study, a
70% difference in tSNR was reported between parallelly and perpendicularly B0-orientated cortical locations3. Here we characterise this effect in data from the Human Connectome Project (HCP)4 that represents a modern 3T fMRI acquisition across a large subject cohort. We test whether specific cortical regions exhibit consistent B0-orientations and therefore tSNR biases across
the population. We confirm a detectable tSNR bias with orientation in a single-subject rs-fMRI
data (Fig.1b), and the bias becomes more evident when averaged over 118 HCP
subjects (Fig.1c). This effect
therefore raises the following questions to be addressed below:
- Within-subject tSNR bias: We created surface-based
average orientation (<θ>) and tSNR (<tSNR>) atlases over 118 subjects
and show how <tSNR> increases with more perpendicular B0-orientation,
with a bias of 20% between parallel and perpendicular orientations.
- Across-subject tSNR variability: Certain
cortical folds vary more than others in geometry across the population5, thus orientation consistency is expected to vary spatially. We found that tSNR variability σ(tSNR) increases with orientation
inconsistency across subjects by up
to 30%.
- Functional connectivity (FC) networks:
We evaluated the effects in common FC-networks. Networks with more perpendicular orientation exhibited higher tSNR
across subjects, and networks with inconsistent orientation exhibited higher across-subject
tSNR variability.
Methods
118 subjects from the HCP 3T dataset (S1200, release 03/2017) were randomly chosen. The unprocessed T1w-MPRAGE and four rs-fMRI scans (2mm isotropic
resolution,TR=720 ms,TE=33ms,SMS=8, no in-plane acceleration) were
downloaded. The HCP minimal processing pipeline6 (including motion and distortion correction, no spatial or temporal smoothing), was followed except for spatial normalisation to
MNI space, to leave the rs-fMRI data in native
subject space. We calculated tSNR maps from these preprocessed rs-fMRI time-series and normalised by the global mean and divided out the provided B1−estimate (to account for bias from the receive coil sensitivities that are proportional to tSNR for thermal-noise-dominated data). Cortical
surfaces were reconstructed using FreeSurfer and registered
to the subject's head position during the rs-fMRI run. For every run vertex-wise orientations were calculated3,7. The tSNR volume maps were then projected onto the surfaces. Each subject's orientation
and tSNR surface-maps were then projected into the “fsaverage” surface-based
atlas space to generate group-level <tSNR> and σ(tSNR) maps.
Orientation mean <θ> and orientation dispersion D were calculated using directional statistics8:$$$\langle\theta\rangle=arctan2[\sum_i sin(\theta_i),\sum_i cos(\theta_i)]\,\text{and}\,D=1-\sqrt{\sum_i sin(\theta_i)^2+\sum_i cos(\theta_i)^2}$$$. We used the Yeo2011
Atlas seven-network FC-based parcellation9 to calculate network averages $$$\overline{<tSNR>},\overline{<\theta>},\overline{\sigma(tSNR)}\,$$$and$$$\,\overline{D}$$$. Here$$$\,$$$we$$$\,$$$use$$$\,\langle\cdot\rangle\,$$$denote$$$\,$$$pooling$$$\,$$$across$$$\,$$$subjects$$$\,$$$and$$$\,^\overline{\,\cdot\,}$$$across$$$\,$$$space/network-ROI.Results
The
orientation atlas (Fig.2a) shows how orientation varies smoothly with cortical
folding. Fig.2b displays where orientations are more consistent $$$(D\approx0)$$$ or random $$$(D\approx1)$$$ across subjects. Fig.3 displays the <tSNR> atlas where
areas with high orientation consistency are bright, and areas that are more orientationally random across subjects are darkened out. We indeed found an
approximately linear relationship (r=0.11,p<0.01) between across-subject
tSNR variability σ(tSNR) and orientation variability D, plotted in Fig.4b. The across-subject mean <tSNR>
shows a non-linear dependence on mean orientation (Fig.4a), with a tSNR bias of 20% between parallel and perpendicular orientations. Low <θ> values
are noisier due to the smaller counts number (see histogram in Fig.2a). The
tSNR variability σ(tSNR) varies by approximately 30%,
with a lower σ(tSNR) of about 0.21 for vertices that are highly orientation consistent $$$(D\approx0)$$$ and a higher σ(tSNR) of about 0.27
for vertices that are highly inconsistent $$$(D\approx1)$$$ across subjects (Fig.4b). We also find that FC-networks follow this tSNR bias trend. Networks with more perpendicular orientation exhibit higher
tSNR (Fig.5b) and networks with greater orientation consistency
have lower tSNR variability (except for the limbic network, which may be an
outlier due to its distortion- and low sensitivity-prone location). Discussion
Cortical orientation contributes to tSNR bias and variability within and across subjects. Here, we accounted for coil sensitivities that alter tSNR and also change with head position. Orientation effects are expected to increase with field strength and in physiologically noise dominated data.
Some FC-networks exhibit consistent cortical orientations across the population,
and more perpendicularly orientated networks will have enhanced detectability. For
group-level analysis orientation variability is another nuisance that should be accountable for, e.g. via regression. On a positive outlook orientation biases represent vascular information and could test the sensitivity of fMRI methods (gradient-echo vs.
spin-echo). Orientation bias in the signal reflects the signal contribution of
the vascular hierarchies which have distinct
orientations with respect to the cortex (pials=parallel, divers=perpendicular, capillaries=pseudo-random).Acknowledgements
We
thank F. Isik Karahanoglu for insightful discussion on fc-network
parcellations. This work was supported in part by
the NIH NIBIB (grants P41-EB015896 and R01-EB019437), by the BRAIN Initiative (NIH NIMH grant
R01-MH111419), and by the MGH/HST Athinoula A. Martinos Center for Biomedical
Imaging; and was made possible by the resources provided by NIH Shared
Instrumentation Grants S10-RR023043 and S10-RR019371.References
[1] Gagnon et al., Journal of Neuroscience; 35(8):3663-3675, 2015.
[2] Baez-Yanez et al., NeuroImage; 163:13-23,2017.
[3] Viessmann et al., Proceedings of the ISMRM-ESMRMB 2018.
[4] Van Essen et al., NeuroImage; 80:62-97, 2013.
[5] Fischl et al., Cerebral Cortex; 18:1973-1980, 2008.
[6] Glasser et al., NeuroImage; 80:105-124, 2013.
[7] Cohen-Adad et al., NeuroImage; 60:1006-1014, 2012.
[8] Ley and Verdebout, Modern Directional Statistics; New York: Chapman and Hall/CRC, 2017.
[9] Yeo et al. Journal of Neurophysiology; 106:1125-1165, 2011.