From structure to function: leveraging embedded multimodal imaging to explore brain connectivity

smoia
@SteMoia
s.moia.research@gmail.com

AIRMM 2025, Lecco, 26.03.2025

Stefano Moia, 2024

Faculty of Psychology and Neuroscience, Maastricht University, Maastricht, The Netherlands;    Open Science Special Interest Group (OHBM);    physiopy (https://github.com/physiopy)

I have nothing to disclose.

This is a new chapter

This is a new chapter

Take home #0

This is a take home message

Connectivity across modalities

BOLD signal: "Functional Networks"

Biswal et al., 1995 (Magn. Reson. Med.)

r(X, Y) = \frac{\text{cov}(X, Y)}{\sigma_X \sigma_Y}

Pearson correlation:

Whole brain functional connectivity

r(X, Y) = \frac{\text{cov}(X, Y)}{\sigma_X \sigma_Y}

Images adapted from Moghimi et al., 2022 (J. Neurophysiol.) and Faskowitz et al., 2020 (Nat. Neurosci.)

Whole brain structural connectivity

Huang et al., 2018 (Proc. IEEE)

(i)EEG connectivity

Images adapted from Sinha et al., 2023 (Neurol.)

(i)EEG connectivity

Images adapted from Sinha et al., 2023 (Neurol.)

Comparing Structural and Functional Connectivity

Baum et al., 2020 (Proc. Natl. Acad. Sci.)

Easy-peasy Structure-Function Coupling (SFC)

Baum et al., 2020 (Proc. Natl. Acad. Sci.)

SFC over the life span

727 participants (420 F), 8 to 23 y.o., BOLD FC & dMRI

Sinha et al., 2023 (Neurol.)

SFC and post-surgery seizure outcome

15 SF (7F) 24 nSF (16)

iEEG-FC (1h) & dMRI

SF = Seizure Free    nSF = non-Seizure Free

Structure-function coupling in EEG/dMRI

Caveat

Images adapted from Sinha et al., 2023 (Neurol.)

Sparse

\( [0,+∞) \)

Highly skewed distribution

Non-Euclidean

Dense

\( [-1,+1] \)

Somewhat normal distribution

Euclidean

Take home #1

Observing Multimodal Connectivity can be
as easy as computing a correlation.

HOWEVER,
there are a few caveats you should be aware of when doing so.

Graph Signal Processing

From timeseries to graphs

x1

x2

x3

x8

...

t
y
t=1
x1=y1

From timeseries to graphs

x1

x2

x3

x8

...

t
y
t=2
x2=y1

From timeseries to graphs

x1

x2

x3

x8

...

t
y
t=3
x3=y1

From timeseries to graphs

x1

x2

x3

x8

...

x1

x2

x3

x8

...

x1

x2

x3

x8

...

From timeseries to graphs

x1

x2

x3

x8

...

x1

x2

x3

x8

...

x1

x2

x3

...

x8

t=1
x1=y1

From timeseries to graphs

x1

x2

x3

x8

...

x1

x2

x3

x8

...

x1

x2

x3

...

x8

t=2
x2=y1=x1_{t1}

From timeseries to graphs

x1

x2

x3

x8

...

x1

x2

x3

x8

...

x1

x2

x3

...

x8

t=3
x3=y1=x2_{t2}
x_{k_{t}}=\sum x_{j_{t-1}}

From timeseries to graphs

x1

x2

x3

x8

...

x1

x2

x3

x8

...

x1

x2

x3

...

x8

t=3
x3=y1=x2_{t2}

From timeseries to graphs

x1

x2

x3

x8

...

x1

x2

x3

x8

...

x1

x2

x3

...

x8

t=4
x_{k_{t}}=\sum \frac {x_{j_{t-1}}} {l_{x_j}}

Graph Signal Processing

Ortega 2022 (Introduction to Graph signal Processing)

White dots = nodes
Black dots = timeseries value

Embedding brain function
into the brain structure

Image courtesy of Maria Giulia Preti;     Preti et al., 2019 (Nat. Commun.)

Graph Signal Processing: Decomposition

Decomposition
to access
"graph" spectral domain

(e.g. Principal Components)

Huang et al., 2018 (Proc. IEEE)

Graph Signal Processing: Filtering

Preti et al., 2019 (Nat. Commun.)

Graph Signal Processing: Structural decoupling index (SDI)

1. Cioli et al., 2014 (PLoS One), Mesmoudi et al. 2013 (PLoS One);    Preti et al., 2019 (Nat. Commun.)

Graph Signal Processing: Structural decoupling index (SDI)

Gene expression¹

SDI²

Rapid responses       Slow (sustained) responses

  High functions                    Low functions

Cioli et al., 2014 (PLoS One);    Preti et al., 2019 (Nat. Commun.)

Graph Signal Processing: Subject-Task Identification

SDI in 22q11.2 deletion syndrome

Forrer et al., 2024 (Biol. Psychiatry)

233 subjects (118F), 108 22q11.2DS (50F), 7-34 y.o.

BOLD fMRI & DWI

SDI in Parkinson

Time-resolved SDI in movies and rest¹

1. Subramani et al., 2025 (Imaging Neurosci.);    2. Langer et al., 2017 (Sci. Data)

Healthy Brain
Network² data

EEG & DWI

SDI in temporal lobe epilepsy (TLE)

Rigoni et al., 2023 (Clin. Neurophysiol.)

17 patients (5F),
8 left TLE,
15–56 y.o.

EEG & DWI

pip install nigsp[all]

nigsp -f timeseries.nii.gz -s sc_mtx.tsv \
-a atlas.nii.gz --informed-surrogates -n 1000

Moia et al. (2023) Zenodo; based on Preti, Van De Ville (2019) Nat. Commun. and Griffa et al. (2022) Neuroimage

Applying Graph Signal Processing: NiGSP

Take home #2

We can leverage Graph Signal Processing
to better explore more complex spatio-temporal features of timeseries embedded in graphs.*

BOLD fMRI
M/EEG
fNIRS

Tractography
EEG caps
Anatomy

*And it's still as easy as installing a python package

Thanks to...

That's all folks!

smoia
@SteMoia
s.moia.research@gmail.com

...you for the (sustained) attention!

...the MR-methods group @UM

...the MIP:Lab @ EPFL

Stefano Moia, 2025

Find the presentation at:

slides.com/smoia/
airmm2025/scroll

Any question [/opinions/objections/...]?

Stefano Moia, 2025

Find the presentation at:

slides.com/smoia/
airmm2025/scroll

1. Multimodal Connectivity analysis is as easy as a correlation of two matrices, with important caveats.

2. Graph Signal Processing is a better approach to model multimodal relationships.

Take home messages