Xiaodi Zhang1, Eric Maltbie1, and Shella Dawn Keilholz1
1BME, Emory University/Georgia Tech, Atlanta, GA, United States
Synopsis
We trained a novel convolutional
variational autoencoder to extract intrinsic spatial temporal patterns from
short segments of resting-state fMRI data. The network was trained in an
unsupervised manner using data from the Human Connectome Project. The extracted
latent dimensions not only show clear clusters in the spatial domain that were
in agreement with DMN/TPN anticorrelations and principal gradients, but also
provide temporal information as well. The method provides a way to extract
orthogonal spatial temporal patterns within fMRI data in a short time window,
among which many patterns were not previously discovered and are worth
investigating in the future.
Introductions
Recent studies have shown that the human
brain consists of several networks with distinct functions, and their
connectivity is dynamically evolving over time. Many methods have been proposed
to study the dynamic aspect of functional connectivity among these major
networks, including independent component analysis (ICA), sliding window
correlation (SWC), quasiperiodic patterns (QPPs), co-activation patterns (CAPs)
and hidden-Markov models (HMM) [1-7]. However, Most of these methods consider
spatial and temporal information separately, when in reality the temporal and
spatial aspects of brain activity are intricately related. Here we proposed to
train a convolutional variational autoencoder using resting-state fMRI data to
identify common spatiotemporal trajectories that describe the flow of activity
across the brain.Methods
FMRI data preprocessing:
The minimally processed resting state and
fMRI data was downloaded from the Human Connectome Project (HCP) S500 release [8].
The first 5 frames were removed to minimize the transient effects. Gray matter,
white matter and cerebrospinal fluid signals, 12 motion parameters (all provided
by HCP), linear and quadratic trends were regressed out altogether
pixel-wisely. The regressed BOLD signals were then band pass filtered using a
0.01-0.1Hz 6-order Butterworth filter, and spatially smoothed using a Gaussian
kernel (FWHM=2pixel). Finally, the BOLD signals were parcellated using
Brainnetome atlas [9] and z-scored. The final parcellated BOLD signal has 412
subjects by 1195 time points (TR=0.72s) by 246 parcels. For better
visualization, the 246 parcels were then sorted into 7 functional networks
using Yeo’s 7-network model [10] provided by Brainnetome website, namely the default
mode (DMN), visual (VIS), somatomotor (SM), dorsal attention (DA), ventral
attention (VA), frontal parietal (FP), limbic (LIM) networks and subcortical
regions (SC).
Convolutional Variational
Autoencoder:
A variational autoencoder (VAE) [11] with 1-D
convolutional layers applying to the temporal dimension was implemented using
pytorch [12], based on the assumption that the rules governing the network
dynamics are shift-invariant across time. This neural network (architecture
shown in figure 1) was selected instead of a plain autoencoder because both the
random sampling in variational autoencoder and the parameter sharing in the
convolutional layer improve generalizability. The 412 subjects were randomly
split into training set (n=248), validation set (n=82) and testing set (n=82). Each
fMRI scan (1195 TR) was divided into 36 segments that are 33-TR long (23.76sec),
with 50% overlapping. The 33-TR segment length was chosen based on prior work
identifying a strong spatiotemporal pattern with a duration of ~20s [5]. Then
the segments were shuffled, resulting a training set with size of [248x36,246,33],
a validation set and a testing set both with size of [82x36,246,33]. The loss
function is the sum of reconstruction loss (root mean square error between
input and output) and the Kullback-Leibler (KL) divergence loss. The networks
were trained on a Nvidia GTX2080Ti GPU using Adam optimizer [13] with a
learning rate of 0.001 for 90 epochs.Results and Discussions
Since the latent variables should be
multidimensional standard Gaussian distribution (all components are
independent, zero-mean, unit-variance), which is encouraged by penalizing the
KL divergence, the features extracted by the latent variables are almost
orthogonal to each other. Thus the spatial temporal feature can be visualized
by propagating change along individual latent dimension through the trained
decoder, which is shown in Figure 2. Note that many latent variables are spatially
similar to each other, thus they were further organized into several groups
based on the spatial profiles at the time points when the variance across the spatial
dimension reaches its maximum (shown as black cursors in figure 2). The latent
variables were grouped into 6 spatially similar clusters using K-means
clustering method, and were sorted by their variance explained, as illustrated
in figure 3.
It can be seen that within the primary
cluster, whose mean variance is the highest, every latent dimension has the DMN
and FPN on one end, whereas the task positive network (TPN), which includes VIS,
SM, DA, VA, is on the opposite end. This finding agrees with many previous
studies, including the DMN/TPN anticorrelation found in [14] and QPPs [5]. The
secondary cluster further separates different networks within the TPN, by
having VIS and DA on one end and SM and VA on the other end. This together with
the primary cluster share remarkably high resemblance with principal gradient 1
and 2 [15]. In addition, the feature learned by the first latent dimension is
very similar to the primary QPP as well (shown in figure 4). But VAE also
offers many other spatial temporal features that have different frequencies,
which were not previously discovered by QPP, whose role in brain function is
worth further investigating. Figure 5 shows the reconstruction of the signal
and the corresponding weights of latent variables, which provide a compact
representation of brain activity.Conclusion
In this abstract we proposed a novel convolutional
variational autoencoder to extract intrinsic spatial temporal patterns from short
segments of resting-state fMRI data. The extracted latent dimensions show clear
clusters in the spatial domain that were in agreement with previous findings,
but also provide temporal information as well. Some spatial temporal features
were similar to QPPs, but there are others with smaller variances that were not
previously discovered, which is worth investigating in the future.Acknowledgements
Funding sources: NIH 1 R01NS078095-01,
BRAIN initiative R01 MH 111416 and NSF INSPIRE. The authors would like to thank
the Washington University–University of Minnesota Consortium of the Human
Connectome Project (WU-Minn HCP) for generating and making publicly available
the HCP data. The authors would like to thank Chinese Scholarship Council (CSC)
for financial support.References
1. Beckmann CF,
Smith SM. Probabilistic independent component analysis for functional magnetic
resonance imaging. IEEE transactions on medical imaging. 2004 Feb
6;23(2):137-52.
2. Calhoun VD,
Adali T, Pearlson GD, Pekar JJ. A method for making group inferences from
functional MRI data using independent component analysis. Human brain mapping.
2001 Nov;14(3):140-51.
3. Allen EA,
Damaraju E, Plis SM, Erhardt EB, Eichele T, Calhoun VD. Tracking whole-brain
connectivity dynamics in the resting state. Cerebral cortex. 2014 Mar
1;24(3):663-76.
4. Chang C, Glover
GH. Time–frequency dynamics of resting-state brain connectivity measured with
fMRI. Neuroimage. 2010 Mar 1;50(1):81-98.
5. Majeed W,
Magnuson M, Hasenkamp W, Schwarb H, Schumacher EH, Barsalou L, Keilholz SD.
Spatiotemporal dynamics of low frequency BOLD fluctuations in rats and humans.
Neuroimage. 2011 Jan 15;54(2):1140-50.
6. Liu X, Duyn JH.
Time-varying functional network information extracted from brief instances of
spontaneous brain activity. Proceedings of the National Academy of Sciences.
2013 Mar 12;110(11):4392-7.
7. Vidaurre D,
Smith SM, Woolrich MW. Brain network dynamics are hierarchically organized in
time. Proceedings of the National Academy of Sciences. 2017 Nov
28;114(48):12827-32.
8. Glasser MF,
Sotiropoulos SN, Wilson JA, Coalson TS, Fischl B, Andersson JL, Xu J, Jbabdi S,
Webster M, Polimeni JR, Van Essen DC. The minimal preprocessing pipelines for
the Human Connectome Project. Neuroimage. 2013 Oct 15;80:105-24.
9. Fan L, Li H,
Zhuo J, Zhang Y, Wang J, Chen L, Yang Z, Chu C, Xie S, Laird AR, Fox PT. The
human brainnetome atlas: a new brain atlas based on connectional architecture.
Cerebral cortex. 2016 Aug 1;26(8):3508-26.
10. Yeo BT, Krienen
FM, Sepulcre J, Sabuncu MR, Lashkari D, Hollinshead M, Roffman JL, Smoller JW,
Zöllei L, Polimeni JR, Fischl B. The organization of the human cerebral cortex
estimated by intrinsic functional connectivity. Journal of neurophysiology.
2011 Sep 1.
11. Kingma DP,
Welling M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114.
2013 Dec 20.
12. Paszke A, Gross
S, Chintala S, Chanan G, Yang E, DeVito Z, Lin Z, Desmaison A, Antiga L, Lerer
A. Automatic differentiation in pytorch.
13. Kingma DP, Ba
J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
2014 Dec 22.
Glasser MF,
Sotiropoulos SN, Wilson JA, Coalson TS, Fischl B, Andersson JL, Xu J, Jbabdi S,
Webster M, Polimeni JR, Van Essen DC. The minimal preprocessing pipelines for
the Human Connectome Project. Neuroimage. 2013 Oct 15;80:105-24.
14. Fox MD, Snyder
AZ, Vincent JL, Corbetta M, Van Essen DC, Raichle ME. The human brain is
intrinsically organized into dynamic, anticorrelated functional networks.
Proceedings of the National Academy of Sciences. 2005 Jul 5;102(27):9673-8.
15. Margulies DS, Ghosh
SS, Goulas A, Falkiewicz M, Huntenburg JM, Langs G, Bezgin G, Eickhoff SB,
Castellanos FX, Petrides M, Jefferies E. Situating the default-mode network
along a principal gradient of macroscale cortical organization. Proceedings of
the National Academy of Sciences. 2016 Nov 1;113(44):12574-9.