Introducing edge-wise graph signal processing: application to connectome fingerprinting
Poster No:
1937
Submission Type:
Abstract Submission
Authors:
Thomas Bolton1, Jagruti Patel2, Mikkel Schöttner3, Patric Hagmann4
Institutions:
1CHUV, Lausanne, Vaud, 2CHUV, Zurich, Not Specified, 3Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL), Lausanne, Not Specified, 4Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL), Lausanne, Switzerland
First Author:
Co-Author(s):
Mikkel Schöttner
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Not Specified
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Not Specified
Patric Hagmann
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland
Introduction:
In magnetic resonance imaging (MRI), graph signal processing (GSP) treats brain structural connectivity (SC) as a graph on which functional activity is studied [5]. Given the potential of edge-centric analyses to reveal novel aspects of brain function [3], we assessed whether edge-centric GSP (eGSP) could outperform classical regional GSP (rGSP) in subject fingerprinting, an endeavour tackled by both research lines [4,6].
Methods:
We considered S=100 subjects from the Human Connectome Project [10] with T1-weighted (T1W), diffusion MRI (dMRI) and resting-state functional MRI (RS fMRI; four 15-minute scans acquired in pairs over two days) data.
Connectome Mapper 3 [9] was used to compute SC for each subject. Tractograms were generated from the dMRI data (deterministic tractography, white matter-seeded, order 8 constrained spherical deconvolution, 10 million output streamlines), and SC reflective of normalized fiber density (NFD, normalized by the sizes of both regions at hand) between R=126 brain regions [2] was obtained.
RS fMRI data (with ICA-FIX denoising) were detrended, high-pass filtered (0.01 Hz), and regressed for 6 head motion parameters, their derivatives, and average white matter and cerebrospinal fluid signals. Following averaging into regions of interest [2], output time courses were temporally z-scored.
For rGSP analysis [5] (Fig. 1A), we computed the normalized Laplacian from average SC across subjects, and its eigendecomposition yielded connectome harmonics (CHs, the eigenvectors) and their associated frequencies (the eigenvalues). A smaller-eigenvalue CH is smoother with respect to SC. Functional signals were expressed as a weighted sum of CHs and filtered to retain their subpart aligned with the underlying SC (i.e., linked to the h smallest-eigenvalue CHs) [8]. The L2-norm was computed over time to yield one feature vector per subject (f, size R). We probed h = {5,10,15,20,⌊0.3R⌋,⌊0.4R⌋,⌊0.5R⌋}.
For eGSP analysis, we built a matrix quantifying cross-subject correlations between the NFD of connection pairs. To focus on direct associations, we set to 0 the entries for which both involved connections did not share a common region. We split the resulting matrix in two (with only positive or sign-flipped negative values), and extracted associated CHs. Regional functional time courses were converted into co-fluctuation time courses [3] (one per connection pair), and alignment was quantified (at h = {R, ⌊10E/R⌋,⌊0.1E⌋,⌊0.2E⌋,⌊0.3E⌋,⌊0.4E⌋, ⌊0.5E⌋}) to yield f+ and f- (size E=7264 each). We also computed log(f+/f-).
Fingerprinting was assessed on pairs of sessions acquired over separate days. The cross-subject similarity matrix between full feature vectors was used for one-to-one matching [7]. Accuracy was taken as the correct percentage of assignments. We also quantified differential identifiability [1] (DI, average within-subject similarity - average between-subject similarity).
Connectome Mapper 3 [9] was used to compute SC for each subject. Tractograms were generated from the dMRI data (deterministic tractography, white matter-seeded, order 8 constrained spherical deconvolution, 10 million output streamlines), and SC reflective of normalized fiber density (NFD, normalized by the sizes of both regions at hand) between R=126 brain regions [2] was obtained.
RS fMRI data (with ICA-FIX denoising) were detrended, high-pass filtered (0.01 Hz), and regressed for 6 head motion parameters, their derivatives, and average white matter and cerebrospinal fluid signals. Following averaging into regions of interest [2], output time courses were temporally z-scored.
For rGSP analysis [5] (Fig. 1A), we computed the normalized Laplacian from average SC across subjects, and its eigendecomposition yielded connectome harmonics (CHs, the eigenvectors) and their associated frequencies (the eigenvalues). A smaller-eigenvalue CH is smoother with respect to SC. Functional signals were expressed as a weighted sum of CHs and filtered to retain their subpart aligned with the underlying SC (i.e., linked to the h smallest-eigenvalue CHs) [8]. The L2-norm was computed over time to yield one feature vector per subject (f, size R). We probed h = {5,10,15,20,⌊0.3R⌋,⌊0.4R⌋,⌊0.5R⌋}.
For eGSP analysis, we built a matrix quantifying cross-subject correlations between the NFD of connection pairs. To focus on direct associations, we set to 0 the entries for which both involved connections did not share a common region. We split the resulting matrix in two (with only positive or sign-flipped negative values), and extracted associated CHs. Regional functional time courses were converted into co-fluctuation time courses [3] (one per connection pair), and alignment was quantified (at h = {R, ⌊10E/R⌋,⌊0.1E⌋,⌊0.2E⌋,⌊0.3E⌋,⌊0.4E⌋, ⌊0.5E⌋}) to yield f+ and f- (size E=7264 each). We also computed log(f+/f-).
Fingerprinting was assessed on pairs of sessions acquired over separate days. The cross-subject similarity matrix between full feature vectors was used for one-to-one matching [7]. Accuracy was taken as the correct percentage of assignments. We also quantified differential identifiability [1] (DI, average within-subject similarity - average between-subject similarity).
Results:
For eGSP, smoothness sharply decreased from eigenvalue R onwards, paralleled by a switch from CHs contrasting large-scale connection communities to more localized attributes (Fig. 1B).
For rGSP and eGSP, fingerprinting accuracy increased to comparable levels (around 60%) as more CHs were retained (Fig. 2A). DI was consistently better for eGSP than for rGSP and reached the largest values for f- (around 10%).
For rGSP and eGSP, fingerprinting accuracy increased to comparable levels (around 60%) as more CHs were retained (Fig. 2A). DI was consistently better for eGSP than for rGSP and reached the largest values for f- (around 10%).
Conclusions:
Similar accuracy but better DI for eGSP schemes suggests greater robustness to increased data noise. EGSP graphs likely capture neurodevelopmental or experience-induced molecular processes that reshape SC. Connections may be jointly strengthened (as reflected in f+; e.g., if functionally overlapping), or may compete for enhancement (as captured in f-; e.g., due to limited resources to distribute). As f- was the best scheme overall, the uniqueness of one's brain may be more strongly related to competition between structural connections for resources than to their joint strengthening.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 2
Methods Development 1
Task-Independent and Resting-State Analysis
Other Methods
Keywords:
Computational Neuroscience
Data analysis
FUNCTIONAL MRI
Multivariate
STRUCTURAL MRI
Other - Graph signal processing
1|2Indicates the priority used for review
Provide references using author date format
[2] Cammoun, L. et al. (2012), 'Mapping the human connectome at multiple scales with diffusion spectrum MRI', Journal of neuroscience methods, vol. 203, no. 2, pp. 386-397
[3] Faskowitz, J. et al. (2020), 'Edge-centric functional network representations of human cerebral cortex reveal overlapping system-level architecture', Nature neuroscience, vol. 23, no. 12, pp. 1644-1654
[4] Griffa, A. et al. (2022), 'Brain structure-function coupling provides signatures for task decoding and individual fingerprinting', NeuroImage, vol. 250, no. 118970
[5] Huang, W. et al. (2018), 'A graph signal processing perspective on functional brain imaging', Proceedings of the IEEE, vol. 106, no. 5, pp. 868-885
[6] Jo, Y. et al. (2021), 'Subject identification using edge-centric functional connectivity', Neuroimage, vol. 238, no. 118204
[7] Kuhn, H.W. (1955), 'The Hungarian method for the assignment problem', Naval research logistics quarterly, vol. 2, no. 1, pp. 83-97
[8] Medaglia, J.D. et al. (2018), 'Functional alignment with anatomical networks is associated with cognitive flexibility', Nature human behaviour, vol. 2, no. 2, pp. 156-164
[9] Tourbier, S. et al. (2022), 'Connectome mapper 3: a flexible and open-source pipeline software for multiscale multimodal human connectome mapping', Journal of Open Source Software, vol. 7, no. 74, p. 4248
[10] Van Essen, D.C. et al. (2013), 'The WU-Minn human connectome project: an overview', Neuroimage, vol. 80, pp. 62-79