Maxime Chamberland1,2, Gabriel Girard2, Michaël Bernier1, Michael Paquette2, David Fortin3, Maxime Descoteaux2, and Kevin Whittingstall1,4
1Nuclear Medicine and Radiobiology, Université de Sherbrooke, Sherbrooke, QC, Canada, 2Computer science, Université de Sherbrooke, Sherbrooke, QC, Canada, 3Division of Neurosurgery and Neuro-Oncology, Université de Sherbrooke, Sherbrooke, QC, Canada, 4Department of Diagnostic Radiology, Université de Sherbrooke, Sherbrooke, QC, Canada
Synopsis
"Your brain is unique" is an
unequivocal sentence that has spanned many research topics in the recent years.
For example, functional connectivity (FC) based on resting-state fMRI is highly
variable from one subject to the next, yet the source of this variability is
unclear. Understanding the source of FC
variability is important as it is often used in clinical studies. Here, we
explore how this might be explained by variability of white-matter structural
connectivity (SC) derived from diffusion MRI tractography connectivity matrices.
Our results show that, across multiple brain areas, motor and visual networks
show the lowest inter-subject variability. This suggests that, at least in these
areas, SC might explain a portion of FC variability.Purpose
Part of FC variability of
the human brain can be explained by evolutionary cortical expansion, cortical thickness and connection
range
1. In this abstract, we hypothesise that SC variability may also
explain a significant portion of functional variability. A method was proposed
by Mueller
et al. 2013 to estimate
inter-subject functional variance by taking into account intra-subject
variability. The result is a variability
index which can be mapped to cortical areas and indicates the variance of a
specific region. We applied the proposed method on structural connectivity
matrices derived from diffusion MRI and compared our results with functional
variability maps to determine if those variances are predictive of one another.
Methods
Diffusion datasets were obtained from 9 subjects (S) over 3
sessions (T) sessions separated by a 1-month interval. Functional datasets encompassed 1
additional scan per subject within the last session.
Structural
connectivity: Diffusion-weighted images
were acquired along 64 uniformly distributed directions using a b-value of 1000
s/mm2, a single-shot echo-planar imaging (EPI) sequence on a 1.5
Tesla SIEMENS Magnetom (128 × 128 matrix, 2 mm isotropic resolution, TR/TE
11000/98 ms) and a GRAPPA factor of 2. An anatomical T1-weighted 1 mm isotropic
MPRAGE image was also acquired. Fiber Orientation
Distribution Functions from spherical deconvolution2 were used for
tractography. Partial volume estimation maps from the T1-weighted image were
obtained using FSL/Fast3 and used in the tracking process.
Streamline tractography was done with particle filtering tractography4, seeding from the white
matter (WM) and grey matter (GM) interface (1 × 1 × 1 mm3, ≈200k
voxels, 10 seeds/voxel). Freesurfer5
was used to generate a macroscale parcellation of each individual brain
into 150 regions6. Structural connectivity (SC)
matrices7 (Msc) were finally normalized (i.e. ∑ij M = 1; i,
j ∈ lower diagonal)8.
Functional
connectivity: Continuous functional
recordings were carried out using a standard EPI sequence (eyes closed). For
each run, 108 functional volumes consisting of 35 axial slices were obtained
with a 64 × 64 matrix, field of view 220 mm, TR/TE 2730/40 ms, for a voxel size
of 3.4 × 3.4 × 4.2 mm3. Data was
motion-corrected, smoothed, skull-stripped, and filtered using non-local means
denoising9, 10. Signals were averaged within each of the previously
described 150 regions and correlated between each other. Functional connectivity
(FC) matrices (Mfc) were finally group-normalized by applying Fisher’s r-to-z
transform.
Inter-subject variability: Variability indices were computed as described by Mueller et al. 2013. The outline of
the method is described as follows: the similarity (VTsim)
between any 2 connectivity vectors (i.e. rows of M) is
given by the correlation (corr) of
those 2 vectors across subjects and for each session T. In addition, the
intra-subject variability (VSintra) was estimated by first
computing corr – 1, (dissimilarity
between 2 connectivity vectors across sessions for each subject S). The
intra-subject variability was then averaged across all subjects, resulting in a
single value (N) for each GM region. Linear regression (VT = [1 - VTsim] - βN - c) allowed the
estimation of VTinter, the inter-subject
variability for each session. Inter-subject variability (Vinter) was
finally obtained for each region by averaging VTinter over the 3 sessions. The Vinter indices
were then back-projected to a brain surface for qualitative assessment. The
mean variability was also computed for both SC and FC variability maps.
Finally, values located in the top 10 percentiles (two-tailed) of Vinter
were computed and the intersected between both SC and FC.
Results
Qualitative whole-brain structural and functional variability maps (
Vinter)
are shown on Fig. 1 (normalized scale). Fig. 2 show quantitative variability
index for each of the 150 GM regions. Fig. 3 shows regions where the
intersection between SC and FC variability maps was the lowest (a, left and
right sulcus central, right post-central sulcus and gyrus, cuneus and calcarine
sulcus) and highest (b, right BA47). No significant
correlation was found between SC and FC variability at the whole brain level (Fig.
4,
r = 0.1).
Discussion and Conclusion
We generated structural and
functional variability maps for 9 subjects by taking into account the
intra-subject variability. SC and FC were smallest in motor and visual cortex,
yet were not correlated at the whole-brain level. Our
results are in line with previous studies that compared functional variability
with the evolutionary cortex expansion
1. We also showed that the inter-subject
structural variability is lower than the inter-subject functional variability
(0.07 vs 0.25, respectively). Finally, BA47 revealed to be the highest
structural and functional variability, a brain structure associated with
language and memory processing.
Acknowledgements
The authors would like to thank David Provencher for useful discussions and acknowledge the funding agencies which have supported this research: NSERC Discovery Grants, QBIN (Quebec Bio-Imaging Network) and the FMSS graduate scholarship program. Maxime Chamberland is supported by the Alexander Graham Bell Canada Graduate Scholarships-Doctoral Program (CGS-D3) from the Natural Sciences and Engineering Research Council of Canada (NSERC). Michaël Bernier, Gabriel Girard and Michael Paquette are supported by the Postgraduate Scholarships-Doctoral Program (PGS D).References
1Mueller et al. (2013) Neuron, 2Tournier et
al. (2007) NeuroImage, 3Zhang et al.
(2001) IEEE Trans. Med. Imaging, 4Girard et al.
(2014) NeuroImage, 5Fischl et al. (2004) Cereb.
cortex, 6Destrieux et al. (2009)
NeuroImage, 7Hagmann
et al. (2007) PLoS ONE, 8Girard et al. (2015) ISMRM Toronto, 9Bernier et al. (2014) Frontiers in Human
Neuroscience, 10Chamberland et
al. (2015) Frontiers in Neuroscience.