MultiPiles
Visualizing States in Dynamic Functional Connectivity
Benjamin Bach
Microsoft Research-Inria Joint Research Center,
France
Tim Dwyer
Monash University, Melbourne, Australia
Tara Madhyastha
Radiology, University of Washington, WA
Thomas Grabowski
Radiology and Neurology,
University of Washington, WA
Nathalie Henry Riche
Microsoft Research, Redmond, WA
Jean-Daniel Fekete
Inria, France
During different cognitive or perceptual tasks, the
brain exhibits different patterns of functional
connectivity. Functional connectivity can be represented as a network with transitions between states as time progresses [1]. However, methods for identifying temporal states from networks extracted from fMRI are just
being developed [1,2,4].
Advances in Human Computer Interaction research have yielded tools to help
people visualize and interact with their data at an early phase of analysis. Such tools
can help generate new hypotheses by allowing users to explore states identified by
an analytical method, judge the quality of the states, compare the results of multiple
methods, clean the data, and refine individual states. Moreover, exploring visual representations of connectivity patterns and their temporal states allows researchers to
form hypotheses which can then be tested by statistical methods.
www.visualizingbrainconnectivity.org/multipiles
d = distance (ti,tj)
T= threshold
2 Pile
1 Overview
Each matrix represents a sliding window (16 seconds) on
the network, off-set by 2 seconds. Time runs row-wise
from left to right. Black cells in each matrix show high
correlation, light cells indicate low correlation between
the respective regions. This network contains 6 regions and 96 time steps, 2 seconds each.
Some periods in the scan show high correlation (black matrice), some show one region uncorrelated, others seem more “noisy”.
Matrices (time steps) are “piled”,
i.e. clustered if they are similar below a threshold (distance(ti,tj) < T). Temporal
order of time steps is preserved during piling and distance is calculated as Euclidian
distance between matrices. Users can set
the threshold T interactively and MultiPiles
shows the new piling as indicated below:
Motivation
1
For each time step, we show an adjacency matrix.
2
If two matrices are similar (below a threashold) we combine those time
steps into a “pile” of matrices.
3
Piles provide a quick summary of the connectivity patterns in the network
over time and allow to compare time periods.
4
Different information can be shown for each pile, e.g. mean connectivity,
variation, trend in correlation strength, etc..
5
6
Automatic matrix reordering techniques make clusters and sub-networks
more visible.
{
Visualization
MultiPiles is an exploratory visualization
that allows for the interactive exploration of
brain connectivity over time.
Pile
Matrices within pile
Cover Matrix
Results
Methods:
• 24 Parkinson disease (PD), 21 control
• 29 regions from default mode and attention networks
• Piling threshold: 5.5 (25% change) for all subjects
One region un-correlated (white cross)
Unclear state
{
3 Compare
Each pile is represented by a matrix (“Cover Matrix”) with
a stack of horizontal bars above, indicating the matrices
(i.e. time steps) within this pile. The cover matrix shows
the mean correlation for each pair of regions, for the period indicated by the pile. This examples shows states (piles) of different length (height of stack
above matrix), and each period with a different connectivity pattern (white cross, where
one region is un-correlated to the rest, and black, where all regions are correlated.of
stack above matrix), and each period with a different connectivity pattern (white cross,
where one region is un-correlated to the rest, and black, where all regions are correlated.
Some matrices belong to neither state.This example alternates between both states.
5 Reorder
The order of
rows and columns inside
matrices and
piles can be optimized to better show topological patterns (clusters and sub-networks)
4 Explore
Finally, matrices within piles can be explored by flipping though the piles.
We used MultiPiles in exploratory analysis to
both confirm and develop hypotheses. The results described here involved overviewing matrices in MultiPiles, trying different piling thresholds, and finally testing our hypotheses based on statistical evidence.
{
All regions correlated
Control participants:
MultiPiles provides several interactive operations to explore
piles. The above example shows connections with increasing
weight (strength) in red, connections with decreasing weight
in blue. The data set is the same as above.
PD participants:
6 Flip Though
Matrices in each piles can be retrieved
by flipping though the pile like a book:
Conclusions
• Multipiles visualizes dynamic brain connectivity as adjacency matrices, one for every time steps.
• Piling reduces the visual complexity of the data set
• This allows to spot topological and termporal states
• Piling can be automatic, but MultiPiles is highly interactive, allowing
for individual creation and exploration of piles
• We could formulate, test and reject several hypothesis based on
the visual representation
Hypotheses:
• PD participants have smaller number of states (piles)
• PD participants have more homogenous states (piles)
• Specific patterns appears more often in PD participants
Findings:
• Number of states (piles) is almost identical in the two groups (controls:
11.15 [SD=2.35], PD: 11.35 [SD=2.48])
• PD had greater standard deviation in the number of states (1.71 versus 1.44, p < .001)
• PD had larger maximum pile sizes (9.29 versus 8.25, p =.002)
• PD had more of the states marked in blue (t(40)=2.28,p=.027); rater
ICC=.78.)
Try Online
www.visualizingbrainconnectivity.org/multipiles
MultiPiles is a web application where everyone can upload and
visualize his/her data for free. Data are uploaded in Nifty format,
will be kept absolutely private and not circulated to any third party.
References
This research was supported by grants from the National
Institutes of Health
1RC4NS073008-01 and P50NS062684.
[1] Calhoun, V. (2014), ‘The Chronnectome: TimeVarying Connectivity Networks as the Next
Frontier in fMRI Data Discovery’, Neuron, Vol. 84, No. 2, pp. 262–274
[2] Hutchison, R. M., (2013), ‘Dynamic functional connectivity: promise, issues, and interpretations’, Neuroimage, vol. 80, pp. 360 378.
[3] Madhyastha, T. M. (2014), ‘Dynamic Connectivity at Rest predicts Attention Task Performance’, Brain Connectivity
[4] Smith, S. M., et al. (2012), ‘Temporallyindependent functional modes of spontaneous brain
activity’, Proceedings of National Academy of Science USA, vol. 109, no. 8, pp. 31313136
© Copyright 2026 Paperzz