m/eeg analysis contrasts, inferences and source localisation
Post on 12-Jan-2016
31 Views
Preview:
DESCRIPTION
TRANSCRIPT
M/EEG Analysis Contrasts, Inferences
and Source Localisation
Methods for Dummies 2011/12 - FIL
Suz Prejawa
‘Ōiwi Parker Jones
Let’s get right into it…
• experimental design• a note on images and lots on image construction in SPM•The SPM processing pipeline (1st and 2nd level analysis)• maybe a bit on other ways of looking at the data
+
spinster
+
time
real word trial
time
600 ms
600 ms
600 ms
Experimental Design
MEG data(imagine some technical specification about the MEG machine and the acquisition here… like number of sensors etc)
Randomised presentation of (i) 96 real words (spinster) (ii) 96 non-words (stinster)
+
spinster
+
time
real word trial
+
time
600 ms
600 ms
600 msstinster
+
time
non word trial
Aim:Identify at what point in time andover what sensor area the greatestdifference lies in the responses towords vs nonwords.
Steps (i) create an image of the data
= MEG 1st level analysis(ii) conduct statistical test
= MEG 2nd level analysis(iii) use SPM for source localisation
+
spinster
+
time
real word trial
time
600 ms
600 ms
600 ms
+
spinster
+
time
real word trial
A word on images
• Essentially what you want is a single image file– So, sensor data needs to be converted in SPM– You have a choice
(i) to average across your trials to obtain 1 image file per condition (ii) to obtain 1 image file for each trial in every condition (epoch data)
– (i) for 2nd level analysis across subjects– (ii) allows within subject statistical tests + comparison
across subjects and conditions (essentially looking at levels within conditions)
Stimulus/EventOnset
X real wordsO nonwords
Averaging trialsan illustration
Averaging
Averaging trials
And then you do this… with, let’s say, averaged data
From Kilner & Friston, 2010)
Note:
275 sensors
averaged ERFs
Time bins may be as many as 1 per ms
A word on interpolationfrom Litvak et al (2011)
“… Data in the time domain are converted into an image by generating a scalp map for each time frame and stacking scalp maps over peristimulus time. Scalp maps are generated using the 2D sensor layout specified in the dataset and linear interpolation between sensors. The user is asked to specify the output dimensions of the interpolated scalp map. Typically, we suggest 64 pixels in each spatial direction.” (p4)
In other words:
You create a 2D space by flattening the sensor locations and interpolating between them to create an image of M*M pixels
(where M=number of channels)
Construction of a 3D (space × space × time) data volume from sensor-space maps
Creation of a single image= what happens in MEG 1st level analysis
A
B
C
Construction of (space × space × time) summary statistic image- Again!
from Litvak et al (2011)
Example: if we were interested in the N170 component, one could average the data between 150 and 190 milliseconds.
0
1
Words
Nonwords
Realistically, you might want to choose a particular time slot for your analysis and not look at all time points
to reduce data that needs to be compared (and subsequently controlled with FWE)
Smoothing
• Important step to take before 2nd level analysis (In SPM, use smooth images function in the drop down other menu)
• Used to adjust images so that they better conform to the assumptions of random field theory
• Necessary for taking into consideration spatial and temporal variability between subjects
• General guiding principle: Let smoothing kernel match the data feature you need to enhance. Try to smooth the images with different kernels and see what looks best.
2nd Level Analysis
Given the contrast images from the 1st level, we can now
test for differences between conditions.
1Tc = 2X
2
+ 2
second levelsecond level
1 -1
2nd level contrast 2nd level model = used in fMRI
SPM output:
Voxel map, where each voxel contains one
statistical value
The associated p-value is adjusted for multiple
comparisons
2 sample t-testDesign Matrix
2nd
L
E
V
E
L
words
Non-words
from Litvak et al (2011)
And that is what you end up with…
This is not the brain (but sensor space and time)
showing where (in the scalp space) there is an effect (in polarity) between the 2 conditions and when
after it has been controlled for multiple comparisons
Summary- MEG in one slide
At the 1st level, we select periods or time points in peri-stimulous time that we would like to analyse.
Time is treated as an experimental factor and we create a 3D image that
contains information on polarity over space and time
to provide input to the 2nd level
0
1
•Similar to fMRI analysis. The aim of the 1st level is to compute contrast images that provide the input to the second level.
•Difference: here we are not modelling the data at 1st level, but simply forming images of data over time
Words
Nonwords
Example: when within the N170 component, is there a significant difference in polarity at any given sensor between my 2 conditions (and which sensor areas are these)?
Transform data into frequency spectrum
-100 -50 0 50 100 150 200 250 300 350 400
-60
-40
-20
0
20
40
60
ms
fem
to T
MRO33 (200)
-100 0 100 200 300 400
5
10
15
20
25
30
35
40
45
ms
Hz
MRO33 (200)
-1500
-1000
-500
0
500
1000
1500
Ideal for induced responses i.e. responses not phase locked to the stimulus onset
Different methods but SPM uses the Morlet Waveform Transform (mathematical functions which breaks a signal into different components)
Trade off between time resolution and frequency resolution
Comparing frequency bands
between 2 conditions
Time-Frequency analysis
Transform data into time-frequency domain
Not phase-locked to the stimulus onset – not revealed with classical averaging methods
[Tallon-Baudry et. al. 1999]
Useful for evoked responses and induced responses:
SPM uses the Morlet Wavelet Transform
Wavelets: mathematical functions that can break a signal into different frequency components.
The transform is a convolution
The Power and Phase Angle can be computed from the wavelet coefficients:
Source localization:I Aim / Application
II Theory
a) What is recorded (EEG / MEG)
b) Forward problem Forward solutions
c) Inverse problem Inverse solutions
d) Inverse solutions: discrete vs. distributed
I Aim
To find a focus of brain activity by analysing the electrical
activity recorded from surface electrodes (EEG) or SQUID
(Superconductive Quantum Interference Device; MEG)
Where does the data come from ?
1pT1s
I Application:
- focal epilepsy:
spikes
- evoked potentials:
auditory evoked potentials
somatosensory evoked potentials
cognitive event related potentials
- induced responses:
alpha/beta/gamma oscillations Fries et al., 2008. (also see Barnes et al., 2004 )
IIa What is recorded
Lopez daSilva, 2004
EPSP
-
Layer IV
radial
tangential
IIb Forward problem Forward solutionHow to model the surfaces i.e. the area between
recording electrode and cortical generator?
Plummer, 2008Realistic shape – (BEM isotropic, FEM anisotropic)
Skin, CSF, skull, brain
-6-4-20246x 10
-13
-6
-4
-2
0
2
4
6x 10
-13
-6
-4
-2
0
2
4
6x 10
-13
Estimated dataEstimated position
Measured data
Dipole Fitting
?
IIc Inverse problem Inverse solutionsDiscrete source analysis Distributed source analysis
Current dipole represents an extended brain area
Each current dipole represents one small brain segment
Number of sources < number of sensors Number of sources >> number of sensors
The leadfieldmatrix has more rows (number of sensors) than colums (number of sources)
The leadfieldmatrix has more colums (number of sources) than rows (number of sensors)
Result:Source model and source waveforms
Result: 3D Volume imagefor each timepoint
Useful priors for cinema audiences
• Things further from the camera appear smaller• People are about the same size• Planes are much bigger than people
Useful priors for MEG analysis
• At any given time only a small number of sources are active (Dipole fitting)
• All sources are active but overall their energy is minimized (Minimum norm)
• As above but there are also no correlations between distant sources (Beamformers)
Singh et al. 2002
ME
G c
ompo
site
fMR
IOscillatory changes can be co-located with BOLD response
MEG + EEG…?
These can also be usefully combined
Two aspects of source analysis are *original* in SPM:
- Based on Bayesian formalism: generic inversion it can incorporate and estimate the relevance of multiple constraints (data driven relevance estimation – Baysian model comparison)
- The subjects specific anatomy incorporated in the generative model of the data
SPM source analysis
IId
Conclusions/If you only remember one three things!
• MEG/EEG inverse problem can be solved…if you have some prior knowledge.
• SPM lets you test between priors in a Bayesian framework, and incorporate subject specific anatomy.
• Really exciting part is the millisecond temporal resolution we can now exploit.
Thanks!
- Gareth Barnes
- Vladimir Litvak
- Marta Garrido
- Calvin
- Hobbes
- And…
you, of course!
Sources – Part 2
- Look under figures
- Gareth Barnes’s “The M/EEG inverse problem and
solutions” lecture
- Stavroula Kousta / Martin Chadwick (2007, MfD)
- Maro Machizawa / Himn Sabir (2008, MfD)
- SPM 8 manual
- BESA tutorials (http://www.besa.de), M. Scherg
- Dipole Simulator (http://www.besa.de/updates/tools/)
References/ source of information for part 1
Vladimir Litvak, Jérémie Mattout, Stefan Kiebel, et al., “EEG and MEG Data Analysis in SPM8,” Computational Intelligence and Neuroscience, vol. 2011, Article ID 852961, 32 pages, 2011. doi:10.1155/2011/852961
Kilner J, Friston K. (2010) Topological inference for EEG and MEG. Annals of Applied Statistics, 4(3): 1272–1290.
Previous MEG MfD talks
Vladimir Litvak, Marta Garrido and Gareth Barnes
III The buttons in SPM :Graphical user interface for 3D source localisation
III EEG/MEG imaging pipeline
0) Load the file
1) Source space modeling
2) Data co-registration
3) Forward computation
4) Inverse reconstruction
5) Summarizing the results of the inverse reconstruction as an image
0) Load the file
1) Source space modeling
MRI
template
1) Source space modeling
Select mesh size:
- coarse
- normal
- fine
2) Data co-registration
Co-register
2) Data co-registration
Methods to co-register
– “select” from default locations
– “type” MNI coordinates directory
– “click” manually each fiducial point
from MRI images
3) Forward computation
Forward Model
Recommend
ation:
Single shell
for MEG
BEM for
EEG
3) Forward computation
4) Inverse reconstruction
Imaging
VB-ECD
Beamforming
4) Inverse reconstruction
Default – click “Standard”:• “MSP” method will be used. MSP : Multiple Sparse Priors (Friston
et al. 2008a)Alternatives:• GS (greedy search: default):
– iteratively add constraints (priors)• ARD (automatic relevance determination):
– iteratively remove irrelevant constraints• COH (coherence):
– LORETA-like smooth prior …
4) Inverse reconstruction
TIME Time course of the region with maximal activity
SPACEMaximal intensity projection (MIP)
5) Summarizing the results of inverse reconstruction as an image
Window
5) Summarizing the results of inverse reconstruction as an image
3D NIfTI images
allow GLM
based statistical
analysis
(Random field
theory)
top related