Figure 2: The network for thelinks returning the strongestintersubject correlations. Thebilateral symmetry is strong.

Figure 3: The connectivity mapcorrelated with singing in the movie.Lateralization over right temporalareas for melody/pitch processing.

Figure 4: Network for head motionin the film. Strong parietal regions,bits of fusiform face area, frontalareas and amygdala.

Figure 5: Network for hand motionin the film. Strong parietal regionsand parieto–occipital connections.

Figure 6: The same network as inFig. 2, but calculated by the wavelet

method by Chang and Glover [1].

Con

tact

deta

ils: a

rno.

solin

@aa

lto.fi

,web

addr

ess:

http

://ar

no.s

olin

.fiN

o.1877,OH

BM

Conference

Poster 2013, S

eattle, US

.

Time–Frequency Dynamics of Brain Connectivityby Stochastic Oscillator Models and Kalman Filtering

Arno Solin∗ Enrico Glerean Simo SarkkaDepartment of Biomedical Engineering and Computational Science (BECS)

Aalto University, FinlandINTRODUCTION

I Functional connectivity in functionalmagnetic resonance imaging (fMRI)is often estimated as the averagepairwise similarity between thetemporal dynamics of two regions ofinterests (ROIs) [5].

I In time-varying functionalconnectivity, where the stimulus iscontinuously changing, the functionalconnectivity is modulated in time[1–2].

I Here, we study the dynamicfunctional connectivity between pairsgray matter of voxels by consideringthe time–frequency dynamicsbetween them.

I Each pair of time series isdecomposed into oscillatorycomponents described by astochastic oscillator model [4].

I The oscillators are inferred from thedata using optimal Bayesian filteringmethods.

I The coherence between the differentfrequency components is convertedinto a time-varying functionalnetwork.

Figure 1: The time–frequency cross-power between two simulated time series.

METHODS

I We use a stochastic oscillator model,that is described by a second orderstochastic differential equation (SDE)for each voxel i and frequency fj :

dx i(t)dt

=

(0 2πfj

−2πfj 0

)x i(t)+

(01

)ξi(t),

where ξi(t) is a random white noisecomponent with spectral density qi .

I This is a linear SDE, and it can besolved for all the observations timepoints tk , k = 1,2, . . .

I The discrete model can then be fittedto the fMRI data using Kalmanfiltering and smoothing.

I The model is based on DRIFTER —a method for removing physiologicalnoise from fMRI data [4] — butevaluated over a set of fixedfrequencies.

I Summing the cross-coherence powersurface over frequencies defines thefunctional activation between theregions (see [1]).

I We compare the pairwise activationtime series to ground truth stimuli inorder to estimate connectivity mapsdescribing the types of connections.

MATERIAL

I Functional brain images together withanatomical data were acquired for 14healthy native speakers of Finnish.

I The subjects watched 22 min 58 s ofa Finnish language film in an MRIscanner (3 T GE Signa Excite,8-channel head coil).

I The film was a shortened version ofthe feature film “Match Factory Girl”(dir. Aki Kaurismaki, 1990, originallength 68 min).

I 16 visual and auditory features fromthe film were extracted as groundtruth references (see [3] for details).

I Sequence parameters: EPI slices:29, TR: 2 s, TE: 32 ms, matrix size:64×64, FA: 90◦, voxel size:3.4×3.4 mm, slice thickness: 4 mm,gap size: 1 mm. Data downsampledto 6 mm isotropic voxels, and 6235gray matter voxels chosen for study.

RESULTS

I Figure 1 shows how the oscillatormodel captures the coupling of twosimulated signals.

I The strongest network of intersubjectconsistency of connectivity timeseries is shown in Figure 2.

I Figures 3–5 show networks of linearregression between stimulus features(singing, head motion, hand motion)[3] and connectivity time series.

I Comparisons of intersubjectconsistency connectionscaptured bythe wavelet approach by Chang andGlover [1]: inter-hemisphericconnections between the occipitallobes appear to be less consistentacross subjects when using wavelets.

CONCLUSIONS

I We have proposed a method forestimating coupling of oscillatoryphenomena in fMRI data.

I The outcome can be used forestimating time-varying functionalconnectivity networks.

I This approach is related to that of [1],where a wavelet basis was used formodeling the cross-coherence.

ACKNOWLEDGEMENTSWe thank Max Hurme and Sirius Vuorikoski for help. We aregrateful for the support provided by Fa-Hsuan Lin and RistoIlmoniemi.

REFERENCES[1] Chang, C. and Glover, G.H. (2010), ‘Time–frequency

dynamics of resting-state brain connectivity measuredwith fMRI’, NeuroImage, vol. 50, no. 1, pp. 81–98.

[2] Glerean, E., et al. (2012), ‘Functional magnetic res-onance imaging phase synchronization as a measureof dynamic functional connectivity’, Brain Connectivity,vol. 2, no. 2, pp. 91–101.

[3] Lahnakoski, J.M., et al. (2012), ‘Stimulus-related inde-pendent component and voxel-wise analysis of humanbrain activity during free viewing of a feature film’, PloSone, vol. 7, no. 4, p. e35215.

[4] Sarkka, S., et al. (2012), ‘Dynamical retrospective fil-tering of physiological noise in BOLD fMRI: DRIFTER’,NeuroImage, vol. 60, no. 2, pp. 1517–1527.

[5] Smith, S.M., et al. (2011), ‘Network modelling meth-ods for fMRI’, NeuroImage. vol. 54, no. 2, pp. 875–891.