Title: Principal Component Regression Approach for Functional Connectivity of Neuronal Activation Measured
1Principal Component Regression Approach for
Functional Connectivity of Neuronal Activation
Measured by Functional MRI
Eini I. Niskanen1,, Mika P. Tarvainen1, Mervi
Könönen2, Hilkka Soininen3, and Pasi A.
Karjalainen1
- 1University of Kuopio
- Dept. of Applied Physics
- P.O.Box 1627, FIN-70211 Kuopio
- FINLAND
- E-mail Eini.Niskanen_at_uku.fi
2Kuopio University Hospital Dept. of Clinical
Neurophysiology P.O.Box 1777, FIN-70211
Kuopio FINLAND
3University of Kuopio Dept. of Neuroscience and
Neurology P.O.Box 1627, FIN-70211 Kuopio FINLAND
2functional Magnetic Resonance Imaging (fMRI)
3fMRI signal
- Each fMRI study contains a huge number of voxel
time series (70 000 100 000 or more) depending
on the imaging parameters - Typical interscan interval is 1-3 seconds ? low
sampling frequency - A lot of noise from head motion, cardiac and
respiratory cycles, and hardware-related signal
drifts
4Blood Oxygenation Level Dependent (BOLD) response
- Paramagnetic deoxyhemoglobin causes local
inhomogeneities in transversal magnetization - ? signal decrease in T2-weighted images
- Stimulus increases the need of oxygen in active
cortical areas - Blood flow and blood volume increase
- concentration of oxygenated hemoglobin increases
- relative concentration of deoxygenated
hemoglobin decreases - in T2-weighted images this is seen as a signal
increase BOLD response
5BOLD response
- BOLD response is slow time to peak 3-5 s, total
duration over 10 s - The signal change due to functional activation is
small 0.5 5 - The shape of the BOLD response varies across
subjects and also within subject depending on the
type of the stimulus and active cortical area - The summation of the consecutive responses for
short interstimulus intervals is highly nonlinear
6Balloon model
volume v '
Inflow f '
Stimulus u
signal s'
deoxyHb q'
Buxton et al. 1998, MRM 39855-864 Obata et al.
2004, NeuroImage 21144-153 Friston et al. 2000,
NeuroImage 12466-477
7Functional connectivity
the temporal correlations among
neurophysiological events between spatially
remote cortical areas
Area 1
Area 2
How to detect the functional connectivity from
the fMRI data
?
Primary visual cortex, Brodmann area 17
Primary motor cortex, Brodmann area 4
8Principal Component Regression (PCR)
- The data is presented as a weighted sum of
orthogonal basis functions - The basis functions are selected to be the
eigenvectors of either covariance or correlation
matrix of the data - The eigenvectors are obtained from eigenvalue
decomposition - The first eigenvector is the best mean square fit
to the ensemble of the data, thus, often similar
to the mean. - The significance of each eigenvector is described
by the corresponding eigenvalue
9Simulations
- A young healthy volunteer was scanned in the
Department of Clinical Radiology in the Kuopio
University Hospital with a Siemens Magnetom
Vision 1.5 T MRI scanner - 700 T2-weighted gradient-echo echo-planar (EP)
images were acquired with interscan interval of
2.5 seconds - Each EP image comprised of 16 slices, slice
thickness 5 mm, in-plane resolution 44 mm - A voxel from primary visual cortex (area 1) and
primary motor cortex (area2) were selected for
analysis and 70 artificial BOLD-responses were
added to both voxel time series - Two data sets were created one set where the
response in area 2 was independent on the
neuronal delay in area 1, and the other where the
response in area 2 was dependent on the neuronal
delay in area 1
10Artificial activations
- The artificial BOLD responses were generated
using the Balloon model - Response amplitude was scaled 5 above the fMRI
time series baseline
11Artificial activations
- The artificial BOLD responses were generated
using the Balloon model - Response amplitude was scaled 5 above the fMRI
time series baseline - Sampling interval was 2.5 seconds used
interscan interval
12Artificial activations
- The artificial BOLD responses were generated
using the Balloon model - Response amplitude was scaled 5 above the fMRI
time series baseline - Sampling interval was 2.5 seconds used
interscan interval - 70 artificial BOLD responses with variable delay
were added to both time series
13Artificial activations
- A delay on response onset time effects on the
sampled activation time series
14Artificial activations
- A delay on response onset time effects on the
sampled activation time series - Small delays are seen as change on amplitude in
sampled response - Larger delays may change the shape of the sampled
response
15Simulated data sets
- The neuronal delays were assumed to be ?2
distributed in both areas - Two data sets were created in the dependent case
the delay in area 1 was a part of the total delay
in area 2, and in the independent case the delay
in area 2 did not depend on the delay in area 1 - A constant delay of 300 ms between the responses
in area 1 and area 2 was assumed in both data sets
16Results
- The voxel time series were divided into adequate
BOLD responses and an augmented data matrix Z was
formed
- Data correlation matrix was estimated
and its eigenvectors and corresponding
eigenvalues were solved RZV V ?
17Results
Independent data set
Dependent data set
?i1 0.5968 ?i2 0.1220 ?i3 0.0850
?d1 0.6055 ?d2 0.1390 ?d3 0.0711
18Discussion and conclusions
- A PCR based method for studying functional
connectivity in fMRI data was presented - Using the method the dependency between two
cortical areas can be determined from the second
and the third eigenvectors - In case of independent responses, the second and
third eigenvectors are required to cover the time
variations of the BOLD responses - In case of dependent responses, this time
variation can be mainly covered by one
eigenvector - The second and third eigenvalues in the
independent case are somewhat closer to each
other than in the dependent case - (??i23 0.0370 vs. ??d23 0.0679) ? the
third eigenvector is not so significant in the
dependent case as in the independent case - In the future the method will be tested with real
fMRI data and the trial-to-trial information of
the BOLD responses is further estimated from the
principal components