adaptive beamformer source reconstruction · adaptive beamformer source reconstruction: ... recent...
TRANSCRIPT
Adaptive beamformer source reconstruction:
Our recent developments toward spatio-temporal reconstruction of brain activities
Kensuke Sekihara1, Srikantan S. Nagayajan2
1Department of Engineering, Tokyo Metropolitan Institute of Technology2Biomagnetic Imaging Laboratory, University of California, San Francisco
ISBET 2003Santa Fe, November, 2003
Right median nerve stimulationmeasured by a 160-channel whole-head sensor array
Hashimoto et al., “Muscle afferent inputs from the hand activate human cerebellum sequentially
through parallel and climbing fiber systems”, Clin. Neurophysiol. Nov;114, pp.2107-17, 2003 .
Right posterior tibial nerve stimulation
measured by a 37-channel sensor array
Hashimoto et al., “Serial activation of distinct cytoarchitectonic areas of the human {SI} cortex after
posterior tibial nerve stimulation,” NeuroReport 12, pp1857-1862, 2001
7 mm
( )
( ) data vector: (t) =
( )
1
2
M
b t
b t
b t
⎡ ⎤⎢ ⎥⎢ ⎥
• ⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
b
Definitions( ) 1b t
( ) 2b t ( ) Mb t
data covariance matrix: ( ) ( )Tt t• =R b b
= [ , ,
source magnitude: ( , )
source orientation: ( ) ( ) ( ) ( )]
•
•
r
r r r rη Tx y z
s t
,t ,t ,t ,tη η η
( ) ( ) ( )
( ) ( ) ( )( ) =
( ) ( ) ( )
η
η
η
⎡ ⎤⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥
= ⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥ ⎣ ⎦
⎣ ⎦
1 1 1
2 2 2
( )
, ( ) ( ) ( )
( )
x y z
xx y z
y
zx y zM M M
l l l
l l l
l l l
r r rr
r r rL r l r L r r
rr r r
y
x
sx
l1x l2
x l3x
lMx
z
y
x
l1z l2
z l3z lM
z
z sz
y
x
l1y l2
y l3y lM
y
z sy
|sx|=|sy|=|sz|=1
Lead field vector for the source orientation ( )rη
Adaptive spatial filter
1
11
( )
(̂ , ) ( ) ( ) [ ( ), , ( )] ( ) ( )
( )
MTM m m
m
M
b t
s t t w w w r b t
b t=
⎡ ⎤⎢ ⎥
= = =⎢ ⎥ ∑⎢ ⎥⎢ ⎥⎣ ⎦
r w r b r r…
1
1
( )( )( ) ( )
TT
T
−
−⇒ =l r Rw rl r R l r
Minimum-variance beamformer
21
1(̂ , )( ) ( )Ts t −=r
l r R l r
subject to ( ) 1min T =T
ww Rw w l r
Spatial filter
Following problems arise when applying minimum-variance beamformer to MEG/EEG source reconstruction.
(1) Output SNR degradation.
(2) Vector source detection.
(3) Statistical significance evaluation.
Minimum-variance beamformer is very sensitive to errors in forward modeling or errors in sample covariance matrix.
Because such errors are inevitable in neuromagnetic measurements, minimum-variance beamformer generally provides noisy spatio-temporal reconstruction results.
Introducing eigenspace projection
⇓
⇓
Output SNR degradation.
and
1
1 1
0 0
0 00
, [ , , | , , ]0
0 0
0 0
ST TP P P M
NS N
M
+
⋅⎡ ⎤⎢ ⎥
⋅⎢ ⎥⎡ ⎤⎢ ⎥= = =⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦⎢ ⎥⋅
⎢ ⎥⎢ ⎥⋅⎣ ⎦
ΛΛ
… …R U U U U U e e e eE E
λ
λ
λ
Output SNR ∝+
2[ ( ) ( )][ ( ) ( ) ]
TS
TS
TN
l r l rl r l r
Γε Γ εΓ
error in estimating ( )l r
Even when is small, may not be small,because noise level eigenvalueλ +≈ ←22 /
TN
TN p j
ε ε εε ε ε
ΓΓ
Also, −= =1 1T TS S S S N N
−Ν ΝΓ Ε Λ Ε Γ Ε Λ Ε,
Some definitions:
Eigenspace projection
Extension to eigenspace projection beamformer
where or , ,TS S x y zµw E E wµ µ= =
∝2[ ( ) ( )]
[ ( ) ( )]
TS
TS
l r l rl r l r
ΓΓ
Output SNR
Output SNR ∝+
2[ ( ) ( )][ ( ) ( ) ]
TS
TS
TN
l r l rl r l r
Γε Γ εΓ
⇓
(non-eigenspace projected)
(eigenspace projected)
The error term arises from the noise subspace component of ( ).TNε Γ ε w r
Application to 37-channel auditory-somatosensory recordingeigenspace-projection results
Application to 37-channel auditory-somatosensory recordingNon-eigenspace projected results
The minimum-variance beamformer formulation should be extended to incorporate the vector nature of sources.
The electromagnetic sources are three dimensional vectors.
Two-types of extensions has been proposed: scalar and vector formulations.
⇓
Vector source detection
=1 1
1 1
( , ) ( )( , )( , ) ( , ) ( ) ( )
T T TT
T T T
− −
− −=l r η R η L r Rw r η
l r η R l r η η L r R L r η
Scalar MV beamformer formulation
uses a single weight vector, but it depends not only on but also on .r η
S. E. Robinson et al., Recent Advances in Biomagnetism, Tohoku University Press, 1999
1 1 1[ ( ), ( ), ( )] [ ( ) ( )] ( )T T Tx y z
− − −=w r w r w r L r R L r L r R
Vector MV beamformer formulation
uses three weight vectors which detect , , and source components. x y z
M. E. Spencer et al., 26th Annual Asilomer Conference on Signals, Systems, and Computers, 1992B. D. van Veen et al., IEEE Trans. Biomed. Eng., 1997
max max21
11
1(̂ , )( ) ( )
1( ( ) ( ) )min
T T
T T
min
s t
γ
−
−−
=
⎡ ⎤= =⎢ ⎥⎣ ⎦
η η
η
rη L r R L r η
η L r R L r η
minimum eigenvalue of 1: [ ( ) ( )]Tminγ −L r R L r
Scalar formulation:
max22
1 1
ˆ ˆ ˆ ˆmax( , ) [ ( ), ( ), ( )]
1max[ [ ( ) ( )] ]min
x y z
T T
s t s s s
γ− −
=
= =
ηη
η
r r r r η
η L r R L r η
Vector formulation:
Two types of formulations give the same output power.
Output power
Asymptotic output SNR
2
2 2200
ˆ ˆ ˆ[ ( ), ( ), ( )] 1max( )( )
x y zoptV
min
s s sZ
ασσ= =
η
r r r ηr
W r η
max12 2
2 2 1 220 00
(̂ , , ) 1 ( ) ( ) 1( ) min( ) ( )( , )
T Topt
nT T
miS
s tZ
ασ σσ
−−
−
⎡ ⎤= = =⎢ ⎥
⎣ ⎦η η
r η η L r R L r ηrη L r R L r ηw r η
Scalar beamformer
Vector beamformer
minimum eigenvalue of 11 2: ( ) ( ) ( ) ( )T T
minα−− −⎡ ⎤ ⎡ ⎤⎣ ⎦ ⎣ ⎦L r R L r L r R L r
Two types of formulations give the same asymptotic output SNR.
2
2 2200
ˆ ˆ ˆ[ ( ), ( ), ( )] 1max( )( )
x y zoptV
min
s s sZ
σ ασ= =
η
r r r ηr
W r η
Asymptotic output SNR for vector beamformer22 2
22 220
ˆ ˆ ˆ( ) ( ) ( )( )
( ( ) ( ) ( ) )x y zconv
V
x y z
s s sZ
σ
+ +=
+ +
r r rr
w r w r w r
(orientation optimized)
(conventional)
optVZ
convVZ
MV beamformer
−
−=
1
2
( )( )
( ) ( )
η L r Rw r
η L r R L r η
T ToptT
T Topt opt
11 2( ) ( ) ( ) ( )T Topt min optα
−− − =⎡ ⎤ ⎡ ⎤⎣ ⎦ ⎣ ⎦L r R L r L r R L r η ηwhere
1
1
( )( )
( ) ( )
−
−=η L r R
w rη L r R L r η
T ToptT
T Topt opt
1( ) ( ) γ− =⎡ ⎤⎣ ⎦L r R L r η ηTopt min optwhere
Weight normalized MV (Borgiotti-Kaplan) beamformer
Orientation optimized weight
Orientation optimized results always gives the best SNR
(̂ , ) ( ) ( ) ( ) ( ) ( ) ( )T T TSs t t t t= = +r w r b w r b w r n
(̂ , ) ( , ) ( ) ( , ) ( ) ( , ) ( )
( , ) ( ) ( , ) ( )
T T TS S
T T
s t t t t
t t
∆
∆
= = +
+ +
r w r b w r b w r b
w r n w r n
η η η
η η
( ) ( ) ( )St t t= +b b nData model (signal plus additive noise):
Beamformer reconstruction:
Parametric method
Thus, if , then 22 2
0 0ˆ( ) (0, ) ( , ) ( ( ) ( ), ( ) )T TSt N s t N tσ σn r w r b w r∼ ∼
Problem:
This is because the weight is obtained using a sample covariance.
Evaluation of the statistical significance:
Test this hypothesis at each voxel using a non-parametric statistics
The null hypothesis: no signal source
Signal sources: sources existing in the test (task) period but not in the control period.
Evaluation of the statistical significance:Our proposed method
assumed source waveform sensor array
} sources
y (cm)z
(cm
)
0-5 5
-5
-10
-15
computer-generated magnetic field spontaneous magnetic recordings
+
(̂ )js t
source waveform
source magnitude waveform
0(̂ )s t
use this distribution as an empirical distribution to test H0 at each (x,y,z).⇐
histogram of prestimulus period
(̂ )∈
j
j
s tt
If is significant. 0 0ˆ ˆ( ) , ( )> ths t s s t
prestimulus poststimulus
95% ths
(̂ )s t
⇓
where 22 2
( ) ( )( ) ( ) ( )
j jj j j
s t s tT t s t s tσ
σ
−= = −
Multiple comparison procedure
histogram of ( )js t histogram of ( )jT t
indicates the average over the prestimulus period.⋅
⇒
Standardization of the distribution
Maximum statistics
histogram of max( , )T x y
1 1( , )maxT x y
2 2( , )maxT x y
3 3( , )maxT x ythmaxT95%
Statistical thresholdingcross sectional view
( , ) ( , ) ( , )Σ σ= +thmaxx y T x y s x y
Threshold:
( , )x yΣ
2( , )s x y
( , )x yσ
Application to auditory-button-press measurements
Summary
My talk has addressed the problems caused from:
•output SNR degradation•vector source detection•statistical significance evaluation
and described our solutions for them:
•eigenspace projection•orientation optimized weight•a novel nonparametric statistical method.
(requires SPM 2)
will be soon available from UCSF.
MEG-MRI Toolbox
http://www.tmit.ac.jp/~sekihara/
Visit
The PDF version of this power-point presentation as well as PDFs of the recent publications are available.
Collaborators
Kanazawa Institute of TechnologyHuman Science Laboratory
Dr. Isao Hashimoto
University of MarylandLinguistics and Cognitive Neuroscience Laboratory
Dr. David Poeppel
Massachusetts Institute of Technology,Department of Linguistics and Philosophy
Dr. Alec Marantz