MRN fMRI Course
Lecture 3.4 (1h): Group ICA
Vince D. Calhoun, Ph.D.
Chief Technology Officer &
Director, Image Analysis & MR Research
The Mind Research Network
Associate Professor, Electrical and Computer Engineering,
Neurosciences, and Computer Science
The University of New Mexico
2010 MRN fMRI Course
2
• Using ICA to analyze fMRI data of multiple subjects
raises some questions:
• How are components to be combined across subjects?
• How should the final results be thresholded and/or presented?
2010 MRN fMRI Course
3
Group ICA
Sub N
Sub 1
ICA
ICA
?
2010 MRN fMRI Course
4
Group ICA Approaches
Subject N
Time
Time
}
Common Spatial
Common Temporal
Subject 1
Subject N
Subject
(avg)
:
Subject N
Back
reconstruction
}
Tensor2,7
Common Spatial
Common Temporal
Subject Parameter
Voxels
Voxels
Voxels
Subject 1
d Pre-Averaging5 e
Unique Spatial
Common Temporal
Common Spatial
Unique Temporal
Correlate/Cluster
Subject 1
Spatial
Concatenation6,5
:
Unique Spatial
Unique Temporal
c
Time
b Temporal
Concatenation3,7,5
Su
bs
a Combine Single
Subject ICA’s1,4
Subject 1
Subject 1
GIFT
Single subject maps
Single subject components*
MELODIC
Brain
Voyager
1) Calhoun VD, Adali T, McGinty V, Pekar JJ, Watson T, Pearlson GD. (2001): fMRI Activation In A Visual-Perception Task: Network Of Areas Detected Using
The General Linear Model And Independent Component Analysis. NeuroImage 14(5):1080-1088.
2) Beckmann CF, Smith SM. (2005): Tensorial extensions of independent component analysis for multisubject FMRI analysis. NeuroImage 25(1):294-311.
3) Calhoun VD, Adali T, Pearlson GD, Pekar JJ. (2001): A Method for Making Group Inferences from Functional MRI Data Using Independent Component
Analysis. Hum.Brain Map. 14(3):140-151.
4) Esposito F, Scarabino T, Hyvarinen A, Himberg J, Formisano E, Comani S, Tedeschi G, Goebel R, Seifritz E, Di SF. (2005): Independent component analysis
of fMRI group studies by self-organizing clustering. Neuroimage. 25(1):193-205.
5) Schmithorst VJ, Holland SK. (2004): Comparison of three methods for generating group statistical inferences from independent component analysis of
functional magnetic resonance imaging data. J.Magn Reson.Imaging 19(3):365-368.
6) Svensen M, Kruggel F, Benali H. (2002): ICA of fMRI Group Study Data. NeuroImage 16:551-563.
7) Guo Y, Giuseppe P. (In Press): A unified framework for group independent component analysis for multi-subject fMRI data. NeuroImage.
2010 MRN fMRI Course
5
Approach 1
• Separate ICA analysis for each subject [V. D. Calhoun, T.
Adali, V. McGinty, J. J. Pekar, T. Watson, and G. D. Pearlson, "FMRI Activation In A VisualPerception Task: Network Of Areas Detected Using The General Linear Model And Independent
Components Analysis," NeuroImage, vol. 14, pp. 1080-1088, 2001.]
• Must select which components to compare between the
individuals
Sub N
Sub 1
ICA
ICA
?
2010 MRN fMRI Course
6
Example
Press buttons (1-4) to
indicate choice
1
2
3
4
15 “events”
…
0
15.4
31.5
47.0
300
Time (seconds)
2010 MRN fMRI Course
7
SPM Results
N=10
P<0.05 corrected
SPM revealed a large network of
areas including:
•frontal eye fields
•supplementary motor areas
•primary visual
•visual association
•basal ganglia
•thalamic, and an
•(unexpectedly) large cerebellar
activation
•bilateral inferior parietal
regions were deactivated (not
shown)
2010 MRN fMRI Course
8
ICA Results
N=10
Z>3.1
ICA revealed a large network of
similar areas including:
•frontal eye fields (blue)
•supplementary motor areas (green
w/ outline)
•primary visual (red)
•visual association (red)
•thalamic (red)
•basal ganglia (green w/ outline)
•a large cerebellar activation (red)
•bilateral inferior parietal
deactivations (not shown)
ICA also revealed areas not identified
by SPM including:
•primary motor (green)
•frontal regions anterior to the
frontal eye fields (blue)
•superior parietal regions (blue)
2010 MRN fMRI Course
9
ICA: Single Subject
The ICA maps from
one subject for the
visual and basal
ganglia components
are depicted along
with their time
courses (basal
ganglia in green and
visual in pink)
Note that the visual
time course
precedes the motor
time course
2010 MRN fMRI Course
10
Event-Averaged Time Courses
•Time courses from selected
voxels in the raw data (a) and
time courses produced by the
ICA method (b).
•In all cases the time courses
are event-averaged (according
to when the figure was
presented) within each
participant and then averaged
across all ten participants.
•Voxels from the raw data
were selected by choosing a
local maximum in the
activation map and averaging
the two surrounding voxels in
each direction.
•Dashed lines indicate the
standard error of the mean.
2010 MRN fMRI Course
11
Approach 2
• Group ICA (stacking images)
•
•
[V. D. Calhoun, T. Adali, G. D. Pearlson, and J. J. Pekar, "A Method for Making Group Inferences
From Functional MRI Data Using Independent Component Analysis," Hum. Brain Map., vol. 14, pp.
140-151, 2001.]
[V. J. Schmithorst and S. K. Holland, "Comparison of Three Methods for Generating Group
Statistical Inferences From Independent Component Analysis of Functional Magnetic Resonance
Imaging Data," J. Magn Reson. Imaging, vol. 19, pp. 365-368, 2004.]
• Components and time courses can be directly compared
Sub 1
ICA
Sub N
Sub 1
2010 MRN fMRI Course
Sub N
12
Group ICA
Data
Back-reconstruction
ICA
1
Subject 1
X
Subject N
2010 MRN fMRI Course
Ai
A1

A


Subject i

Si
S_agg
AN
13
Simulation
Nine simulated source maps and time
courses were generated, followed by
an ICA estimation. The red lines
indicate the t<4.5 boundaries
2010 MRN fMRI Course
14
Are the data separable? (Simulation)
•A natural concern is whether
the back-reconstructed maps
from individual subjects will be
influenced by the other subjects
in the group analysis
•This simulation was performed
in which one of the nine
“subjects” had a structured,
source #2 map (whereas all of
the nine “subjects” had a
similar, source #1 map).
•As one can see, in this
example, the back-reconstructed
ICA maps are very close to the
individual maps and there
appears to be little to no
influence between subjects
2010 MRN fMRI Course
15
The Stationarity Assumption
Stationary source S
common to all five
“subjects”
Sources S1-S5
differing across the
five “subjects”
S
S1
S2
S3
S4
S5
ICA results
2010 MRN fMRI Course
source #1
source #2
•The ICA estimation requires the
data to be stationary across subjects
•Some signals in the data (e.g.
physiologic noise) will most likely
*not* be stationary
•However it is reasonable to
assume the signal of interest (fMRI
activation) will be stationary
•A simulation was performed to
examine how non-stationary
sources would affect the results
•One stationary signal (fMRI
activation) and one non-stationary
signal were simulated for a fivesubject analysis
•The ICA results reveal that the
fMRI activation is preserved
16
Evaluation of Group ICA Methods
2010 MRN fMRI Course
E. Erhardt, S. Rachakonda, E. Bedrick, T. Adali, and V. D. Calhoun, "Comparison of multi-subject ICA
methods for analysis of fMRI data," in Proc. HBM, Barcelona, Spain, 2010.
17
Comparison of multi-subject ICA methods for analysis of fMRI data
2010 MRN fMRI Course
E. Erhardt, S. Rachakonda, E. Bedrick, T. Adali, and V. D. Calhoun, "Comparison of multi-subject ICA
methods for analysis of fMRI data," in Proc. HBM, Barcelona, Spain, 2010.
18
DIFF
STR
GICA3
Comparison of multi-subject ICA methods for analysis of fMRI data
2010 MRN fMRI Course
E. Erhardt, S. Rachakonda, E. Bedrick, T. Adali, and V. D. Calhoun, "Comparison of multi-subject ICA
methods for analysis of fMRI data," in Proc. HBM, Barcelona, Spain, 2010.
19
Default Mode Group Maps
GICA3
2010 MRN fMRI Course
STR
E. Erhardt, S. Rachakonda, E. Bedrick, T. Adali, and V. D. Calhoun, "Comparison of multi-subject ICA
methods for analysis of fMRI data," in Proc. HBM, Barcelona, Spain, 2010.
20
Methods
• Scan Parameters
•
•
•
•
•
9 slice Single-shot EPI
FOV = 24cm, 64x64
TR=1s, TE=40ms
Thickness = 5/.5 mm
360 volumes acquired
• Preprocessing
•
•
•
•
Timing correction
Motion correction
Normalization
Smoothing
+
Right
+
Left
t (secs)
+
0
90
180
270
360
• ICA
• An ICA estimation was performed on each of the nine subjects
• Data were first reduced from 360 to 25 using PCA, the data were
concatenated and reduced a second time from 225 to 20 using PCA
• An ICA estimation was performed after which single subject maps and
time courses were calculated
• Group averaged maps were thresholded at t<4.5, colorized, and overlaid
onto an EPI scan for visualization
2010 MRN fMRI Course
21
Are the data separable? (fMRI experiment)
•The same slice from nine subjects when the right (red) and left (blue) visual fields
were stimulated, (a) analyzed via linear modeling (LM), (b) back-reconstructed from a
group ICA analysis, or (c) calculated from an ICA analysis performed on each subject
separately. A transiently task-related component is depicted in green.
•The results between the two ICA methods appear quite similar and match well with
the LM results as well (note that there may be small differences due to different initial
conditions for the ICA estimation)
2010 MRN fMRI Course
22
Comparison with GLM Approach
R
L
V.D. Calhoun, T. Adali, G.D. Pearlson, and J.J. Pekar, "A Method for Making Group Inferences From Functional
MRI Data Using Independent Component Analysis," Hum. Brain Map., vol. 14, pp. 140-151, 2001.
2010 MRN fMRI Course
23
Sorting/Calibrating
•
A ‘second-level’ or group analysis involves taking certain parameters (estimated by
ICA) such as the amplitude fit for fMRI regression models, or voxel weights, and
testing these within a standard GLM hypothesis-testing framework
2010 MRN fMRI Course
Comp#
R2
1
0.81
10
0.81
4
0.017
Subject
Reg1
Reg2
1
2
1
2
1
1.89
2.28
0.28
0.65
-0.19
0.02
0.66
2.19
2.03
-0.40
2
-0.10
0.08
24
Prenormalization
1) No Normalization (NN),
where data is left in its
raw intensity units
(Calhoun, 2001)
2) Intensity Normalization
(IN), which involves voxelwise division of the time
series mean
3) Variance Normalization
(VN), voxel-wise z-scoring
of the time series
(Beckmann, 2004).
E. Allen, E. Erhardt, T. Eichele, A. R. Mayer, and V. D. Calhoun, "Comparison of pre-normalization methods on the accuracy of
group ICA results," in Proc. HBM, Barcelona, Spain, 2010.
2010 MRN fMRI Course
25
Result 1: AOD and rest data
produced highly similar networks
Comp# Comp# Description
Corr
Oddball Rest
A: Default mode
16
19
0.9577
B: Motor
11
9
0.9156
C: Sup parietal
13
12
0.9142
D: Medial visual
10
6
0.8628
E: Left lateral frontoparietal 0.8557
12
7
F: Lateral Visual
14
2
0.8170
G: Temporal2
17
13
0.8135
H: Cerebellum
8
11
0.8059
I: Temporal1
1
15
0.8048
J: Frontal
4
16
0.7838
K: Right lateral frontoparietal 0.8170
2
4
L: Anterior cingulate
5
0.035
V. D. Calhoun, K. A. Kiehl, and G. D. Pearlson, "Modulation of
Temporally Coherent Brain Networks Estimated using ICA at Rest and
During Cognitive Tasks," Hum Brain Mapp, vol. 29, pp. 828-838, 2008.
2010 MRN fMRI Course
26
Result 2: Though similar TCNs were
identified for AOD and rest, spatial and
temporal task modulation was induced
Description
Tar
Nov
A: Default mode
-8.44 (1.4e-9)
-5.79 (5.6e-6)
B: Motor
4.62 (2.3e-4)
1.11 (1.0)
C: Sup parietal
2.51 (8.9e-2)
-3.50 (6.5e-3)
1.09 (1.0)
0.12 (1.0)
E: Left lateral frontoparietal
2.41 (1.1e-1)
1.21 (1.0)
F: Lateral Visual
-4.34 (5.4e-4)
-3.92 (1.9e-3)
G: Temporal2
10.29 (6.2e-12)
7.76 (1.1e-8)
H: Cerebellum
4.09 (1.1e-3)
-2.59 (7.4e-2)
D: Medial visual
I: Temporal1
J: Frontal
13.67 (1.2e-15) 9.30 (1.1e-10)
-2.55 (8.1e-2)
-3.28 (1.2e-2)
K: Right lateral frontoparietal -12.00 (6.3e-15) -3.89 (2.1e-3)
V. D. Calhoun, K. A. Kiehl, and G. D. Pearlson, "Modulation of
Temporally Coherent Brain Networks Estimated using ICA at Rest and
During Cognitive Tasks," Hum Brain Mapp, vol. 29, pp. 828-838, 2008.
2010 MRN fMRI Course
27
Example of spatial sorting
2010 MRN fMRI Course
28
Example 1: ‘Default Mode’ Mask
• Using wfu pickatlas to define mask using regions
reported in Rachle 2001 paper
•
•
•
•
•
Posterior parietal cortex BA7
Occipitoparietal junction BA 39
Precuneus
Posterior cingulate
Frontal Pole BA 10
• Smooth in SPM with same
kernel used on fMRI data
• Sort in GIFT using
spatial sorting
2010 MRN fMRI Course
A.Garrity, G.D.Pearlson, K.McKiernan, D.Lloyd, K.A.Kiehl, and V.D.Calhoun, "Aberrant 'Default Mode' Functional
Connectivity in Schizophrenia," to appear Am. J. Psychiatry, 2006.
29
ICA to identify ‘Default Mode’ Network
Healthy
Schizo
Healthy vs
Schizo (N=26/26)
2010 MRN fMRI Course
+Symptoms
A.Garrity, G.D.Pearlson, K.McKiernan, D.Lloyd, K.A.Kiehl, and V.D.Calhoun, "Aberrant 'Default Mode' Functional
Connectivity in Schizophrenia," to appear Am. J. Psychiatry, 2006.
30
Spatial Sorting: Example 2
• Classification of Schizophrenia
• Mapping the brain via intrinsic connectivity
Patients
Controls
2010 MRN fMRI Course
31
Robustness of ‘modes’
.5 kHz
Standard
tone,
sweep,
whistle
1 kHz
Standard
2010 MRN fMRI Course
Target
Standard
Standard
Standard
Novel
Standard
32
The Challenge
• Accurate classification requires single-subject
accuracy -> very stringent requirement!
• We cannot use knowledge of the diagnosis in the
development of the classification algorithm
2010 MRN fMRI Course
33
Temporal Lobe Synchrony
• Supervised Classification
•
•
•
•
•
Step 1: Select Training Group
Step 2: Use ICA to extract temporal lobe maps
Step 3: Compute within-group mean images
Step 4: Subtract the mean images
Step 5: Set a positive and negative threshold
t
…
HC1
HCN
t+
ICA
…
Sz1
2010 MRN fMRI Course
SzN
Calhoun VD, Kiehl KA, Liddle PF, Pearlson GD: “Aberrant Localization of Synchronous Hemodynamic
Activity in Auditory Cortex Reliably Characterizes Schizophrenia”. Biol Psychiatry 2004; 55842-849 34
Temporal Lobe Synchrony in Schizophrenia
t
t+
2010 MRN fMRI Course
35
Temporal Lobe Synchrony in Schizophrenia
• Step 6: Form classification measure (average the values within each
boundary and subtract)
DF  t +, t , i    i , IM t +  i , IM t   .* MSK HC|SZ
• Step 7: Optimize group discrimination (using a sensible error metric)
min Err  t +, t      DF  t +, t , i   0 +   DF t +, t , i   0 
iSz
iHc
• Step 8: Apply classification to new data
2010 MRN fMRI Course
Calhoun VD, Kiehl KA, Liddle PF, Pearlson GD: “Aberrant Localization of Synchronous Hemodynamic
Activity in Auditory Cortex Reliably Characterizes Schizophrenia”. Biol Psychiatry 2004; 55842-849 36
Temporal Sorting: fBIRN SIRP Task
• Methods
• Subjects & Task
• 28 subjects (14 HC/14 SZ) across two sites
• Three runs of SIRP task preprocessed with SPM2
• ICA Analysis
• All data entered into group ICA analysis in GIFT
• ICA time course and image reconstructed for each subject, session, and
component
• Images: sessions averaged together creating single image for each subject and
component
• Time courses: SPM SIRP model regressed against ICA time course
• Statistical Analysis:
• Images: all subjects entered into voxelwise 1-sample t-test in SPM2 and
thresholded at t=4.5
• Time courses: Goodness of fit to SPM SIRP model computed, beta weights for
load 1, 3, 5 entered into Group x Load ANOVA
2010 MRN fMRI Course
fBIRN Phase II Data: www.nbirn.net;
NCRR (NIH), 5 MOI RR 000827 (2002-2006) and 1 U24 RR0219921 (2006 onwards)
37
Component 1: Bilateral Frontal/Parietal
2010 MRN fMRI Course
fBIRN Phase II Data: www.nbirn.net;
NCRR (NIH), 5 MOI RR 000827 (2002-2006) and 1 U24 RR0219921 (2006 onwards)
38
Component 2: Right Frontal, Left Parietal, Post. Cing.
2010 MRN fMRI Course
fBIRN Phase II Data: www.nbirn.net;
NCRR (NIH), 5 MOI RR 000827 (2002-2006) and 1 U24 RR0219921 (2006 onwards)
39
Component 3: Temporal Lobe
2010 MRN fMRI Course
fBIRN Phase II Data: www.nbirn.net;
NCRR (NIH), 5 MOI RR 000827 (2002-2006) and 1 U24 RR0219921 (2006 onwards)
40
Example 2: Simulated Driving Paradigm
*
0
2010 MRN fMRI Course
Drive
Watch
180
360
600
41
Previous Work
• Walter, 2001.
Driving
Watching
“Our results suggest that simulated driving engages mainly areas concerned with perceptualmotor integration and does not engage areas associated with higher cognitive functions.”
“our study suggests that the main ideas of cognitive psychology used in the design of cars, in
the planning of respective behavioral experiments on driving, as well as in traffic related political
decision making (i.e. laws on what drivers are supposed to do and not to do
during driving) may be inadequate, as it suggests a general limited capacity model of the psyche of
the driver which is not supported by our results. If driving deactivates rather
than activates a number of brain regions the quests for the adequate design of the man-machine
interface as well as for what the driver should and should not do during driving is still widely open.”
2010 MRN fMRI Course
42
Baseline Simulated Driving Results
N=12
Higher Order Visual/Motor:
Increases during driving; less
during watching.
Low Order Visual:
Increases during driving;
less during watching.
Motor control: Increases
only during driving.
Vigilance: Decreases only
during driving; amount
proportional to speed.
Error Monitoring &
Inhibition: Decreases only
during driving; rate
proportional to speed.
Visual Monitoring:
Increases during epoch
transitions.
*
Drive
Watch
V. D. Calhoun, J. J. Pekar, V. B. McGinty, T. Adali, T. D. Watson, and G. D. Pearlson, "Different Activation
in Multiple Neural Systems During Simulated Driving," Hum. Brain Map., vol. 16, pp. 158-167, 2002.
2010 MRN fMRI Course
Dynamics
43
SPM Results
2010 MRN fMRI Course
Calhoun, V. D., Pekar, J. J., and Pearlson, G. D. “Alcohol Intoxication Effects on Simulated Driving: Exploring AlcoholDose Effects on Brain Activation Using Functional MRI”. Neuropsychopharmacology 2004.
44
Interpretation of Results
2010 MRN fMRI Course
45
Functional Network Connectivity
(between groups)
A: Default
Key:
: ρpatient > ρcontrol
: ρcontrol > ρpatient
G: Temporal
B: Parietal
C: L. & M. Visual
Cortical Areas
F: Frontal
E: Frontal Parietal
Subcortical
2010 MRN fMRI Course
D: Frontal
Temporal Parietal
46
FNC Software
2010 MRN fMRI Course
47
Fusion ICA Toolbox (FIT)
500+ unique downloads
http://icatb.sourceforge.net
Funded by NIH 1 R01 EB 005846
2010 MRN fMRI Course
48
FMRI Snapshots (movie)
ERP (temporal) Components: T  t1
FMRI (spatial) Components: S  s1
FMRI Image Snapshot: MF t   T  ST t 
sN 
tN 
Calhoun, V.D., Pearlson, G.D., and Kiehl, K.A. (2006). Neuronal Chronometry of Target Detection: Fusion of
2010Hemodynamic
MRN fMRI Courseand Event-related Potential Data. NeuroImage 30, 544-553.
49
Target Stimuli
SNPs
0.55
rs1800545 ADRA2A
rs7412
APOE
rs1128503 ABCB1
rs6578993
TH
rs1045642 ABCB1
rs2278718 MDH1
rs4784642 GNAO1
rs521674 ADRA2A
Novel Stimuli
SNPs
2010 MRN fMRI Course
0.47
Genes
Genes
rs1800545 ADRA2A
rs7412
APOE
rs6578993
TH
rs2278718 MDH1
rs1128503 ABCB1
rs429358
APOE
rs3813065 PIK3C3
rs4121817 PIK3C3
rs521674 ADRA2A
50
Demo
• 3 subject ICA
• Sorting
• Component Explorer (split time courses, event-related
average)
• Orthogonal Viewer
• Composite Viewer
• Examine Regression Parameters
• Taking Images/Timecourses from GIFT to SPM
2010 MRN fMRI Course
51
2010 MRN fMRI Course
52
Comparison of multi-subject ICA methods for analysis of fMRI data
2010 MRN fMRI Course
E. Erhardt, S. Rachakonda, E. Bedrick, T. Adali, and V. D. Calhoun, "Comparison of multi-subject ICA
methods for analysis of fMRI data," in Proc. HBM, Barcelona, Spain, 2010.
53