Lecture 11: Source Localization in MEG Flashcards
So far in the module, we have discussed (sensor space) - (2)
activity in brain produces magnetic field which is picked up from sensors outside the head
We can plot activity across the sensors like this:
In sensor space, we detect the brain activity at sensors outside of the
head
In sensor space, we can perform assessments of
time and frequency of responses
In sensor space we can also look at approximately where the effects are happening - example
So yellow region show strongest effect happens towards the back of head but don’t have greater sense of ‘where’ it specifically is
Sometimes we want to know instead of sensor space , we want to know
where the activity is changing in the brain
What is source localization?
Tries to work out where the activity is changing in the brain by using a plausible model of the head and a set of assumptions about how signals propogate
Diagram of sensor space to (estimate) source space - (3)
No longer looking at…
In source space get better at…
This is an estimate source space meaning…
No longer looking at data per sensor but looking data due to positions within brain than the head
In source space, talk a bit better where responses are coming from - get this brain map like in MRI.
Its estimate of source space as difficult to transform sensor to source space so estimating based on what we know how signals move around and what we know about the anatomy of the participants head
Source localization gives our results in
source space
What does this diagram show? - (3)
Image things called dipoles in the brain
Has a positive charge on one side and a negative charge on other
This dipole produces our brain activity
We can model the sources in the brain as..
when the dipole turns on.. - (2)
dipoles in brain
When the dipole turns on it produces some brain activity that will be picked up at the sensors
Since we can model sources in the brain as dipoles we can consider the brain to include many
dipoles
Consider the brain to include many dipoles that are potential sources of
brain activity
Consider the brain to include many dipoles which are all potential sources of brain activity
what question are we then asking?
Which of these dipoles are turning on?
Consider the brain to include many dipoles which are potential sources of activity and
each dipole have a
location and an orientation
What does this image show? - (2)
The black dot shows the location of the dipole
Red arrow shows the orientation of dipole - which way is negative and which way is positive
Consider brain to have many dipoles and all potential sources of brain activity
To know where the possible sources of the brain activity (i.e., in diapoles) should be and likely orientation we need to know
the anatomy/structure of participants’ brain to see the possible sources of brain activity and likely orientation in cortex of grey matter
Participants have individual differences in how their
brain is structured
To know where the possible sources of the brain activity (i.e., in diapoles) should be and likely orientation we need to know the anatomy/structure of participants’ brain
how do we know the anatomy of the participants’ brain? - (2)
using their fMRI structural scan
or if we can’t then use average brain (e.g., MNI152 brain template)
Diagram of structural MRI and MNI
MNI is not as good as technique as structural MRI as..
but why is it used? - (2)
won’t know exact cortical folds and but some sense where their grey matter is
Used if some participants don’t show up for MRI session after EEG/MEG
From structural scan and MNI, we try to
produce 3D model of the brain using brain mesh
T1 Structural scans will take a series of
if we see an edge of scans then - (2)
2D slices that shows us the anatomy of the participants’ brain
compute a 3D model of brain
Intensity differences in 2D T1 anatomcial scan and T2 functional scans lets us work out the
3D shape of the brain , head and skull
From, Intensity differences in the 2D T1 and T2 scans let us work out the 3D shape of the brain, head and skull.
we can then build the
3D meshes of the brain
Intensity differences in the 2D T1 and T2 scans let us work out the 3D shape of the brain, head and skull
we can then build 3D meshes , one for brain, skull and head
This is typically created by
Freesurfer software
What are meshes?
meshes are made up of little triangles and each point of these triangles connect is called a vertx
Diagram of brain mesh and vertices
In each vertex of brain mesh, we can place a
dipole (potential source of brain activity) and decide (or estimate its orientation)
Putting dipoles in each vertex is essentially putting
use results of meshes for.. - (2)
dipoles in source space
use results of brain activity on meshes for data visualisation
Diagram of dipole in source space
+ where the dipoles placed in each vertex and why… - (2)
15,000 dipoles placed in vertexes on the cortex as we think MEG signal is coming from - surface of grey matter
dots are dipoles and covering cortical surface of a person’s brain
In each vertex of brain mesh, we can place a dipole (potential source of brain activity) and decide (or estimate its orientation)
OR we can allow dipole to have all 3 possible orientations
how to decide orientation or all orientations? - (2)
orientations
get software to learn all orientations and estimate it OR make simplified assumption that should be oriented perpendicular to the cortex (e.g., pyramidal cells)
Putting dipoles in each vertex of brain mesh allows the
analyses to let us look at the activity in each vertex (i.e., treat them like fMRI voxels)
Can understand how the current of diapole would propogate from a source based on (2)
shape of the brain and head
physics of the current flow
What does this diagram show in terms how MEG signals disperse from source (diapole)? - (2)
The diapole’s electric signal moves around the brain because in CSF and salty water has projections of electrical signal throughout the brain
The diapole has strong inital current but dispersal of signal electrically throughout brain
The MEG picks up the
stronger primary current and dispered volume currents (summed together)
What is volume currents which is hard to source localise
dispered currents in the photo
MEG picks up the stronger primary current and the dispersed volume currents (summed together) which is a bit harder for source localisation but
MEG is better for source localisation than
EEG
To estimate how MEG signal dispere from its source we - (3)
model the brain as one large sphere underneath each sensor (ignoring its detailed anatomy - like cortical folds as physics is hard). This makes many highly overlapping spheres centered on the brain.
Can model the brain as a simple sphere for MEG as magnetic fields are not strongly affected by the materials they are passing through.
Thinking how MEG dispere from source and reach sensor
Can model the brain as a simple sphere for MEG as magnetic fields are not strongly affected by the materials they are passing through.
Major advantage for MEG over EEG as - (2)
Electrical signals in EEG as its cares if its going through salty water or bones or metal
But magnetic signal disperes on the basis that it will be weaker the further away it gets but does not care what material it has gone through - reason why source localisation is better for MEG over EEG
Diagram of modelling the brain as simple sphere for MEG
More complicated and sophiscated model of how MEG signals dispere from its source (source localisation) takes..
skull and various tissues into account (known as Boundary Element or Finite Element Models) into Freesufer output
Individuals tend to not use much of the sophiscated models known as .. for MEG as they won’t work well - doesnt vary by materials
Boundary Element or Finite Elements Model
Whether we have overlapping spheres model or more sophiscated models of MEG (Boundary Element or Finite Element) of source localisation (how does signal dispere from source) is known as
head model - how we are thinking of anatomy of head so see how MEG signals move through
We can use more sophsicated models of source localisation for EEG - Boundary Element or Finite Element Models
which is needed for EEG as - (2)
as electrical signals are spread out more by the tissue they pass through and how much varies by the type of tissue (e.g., CSF and skin lets a lot through, bones don;t)
(although EEG source localisation will never be as good anyway due to this dispersal and the greater effect of the volume currents).
Additional step of producing head model for source localisation is to
know exactly where the brain is compared to MEG sensors
In co-registration in MEG, we can determine brain’s location in MEG helmet by
co-register (align) the head mesh (i.e., 3D meshes of the head and brain from MRI structural scan) to the MEG data (i.e., In MEG session, outline of the participants’ head is digitlsed)
Diagram of MRI structural session and MEG session where outline of head taken with lasers
Determining the brain’s location in the MEG helmet
MEG data also includes, beside outline, is - (2)
The sensor locations are recorded relative to reference coils placed at fiducial points on the head
Therefore, can see where the sensors are in relation to the head and brain
We want to go from
pattern of sensor responses recorded with MEG (sensor space) to underlying brain activity (source space)
What is the forward problem?
How to go from possible sources to sensor responses - source space to sensor space
How to overcome forward problem?
Do forward modelling
Diagram of forward vs inverse problem
we want inverse problem but its harder
Forward modelling is relatively easier than inverse modelling as
we can predict the pattern of responses we should
record at the sensors
The forward model is relatively straightforward to calculate for
simulated sources
The different sensors are differentially sensitive to in forward modelling
different sources and different sources more or less obvious at sensors (e.g,., orange one has big impact of sensor compared to others - blue isn’t perpendicular orientation)
The leadfield in forward model is the prediction of - (2)
what each sensor would show for a given source of activity
Gives matrix of each sensor x sources
What does this leadfield show? - (3)
orange one has big impact of sensor compared to others - blue isn’t perpendicular orientation
Numbers next to them are weights saying if we have activity at that source how much activity at this possible source gets to the sensor
If we do for sensory we get a matrix
We have to go to forward modelling (source to sensor space) first before
inverse modelling (sensor space to source space)
In forward models we can
stimulate what the activity would look like, ask them questions and stimulate different methods
Example of asking questions in forward model
Boto et al., (2016) did a forward model by looking what effect a simulated source would have with a sensor close to head (OPM) vs sensor far away from head (SQUIDs)
They compared on scalp OPM sensors and helmet-based SQUIDs which showed..
OPM sensors being closer to the brain than SQUIDs could result in a 5-fold increase in sensitivity!
What is inverse problem? - (2)
Want to go directly from a pattern of sensor responses (sensor space) to underlying estimated brain activity (source space)
Direction we want to go to
For inverse problem we need to produce a inverse model which is
much harder than producing forward model for forward problem
Another diagram of forward vs inverse problem - (3)
not simply forward and backward
Forward problem is from true source to MEG measurement
Inverse is from MEG measurement to estimated source
Inverse modelling is known as ill-posed problem which means that
an infinite number of different patterns of brain activity that could result in the same pattern of responses at the sensors
Inverse modelling is like
looking at shadow of activity rather than activity in brain itself
Example of inverse problem being an ill-posed problem - money
Example of inverse problem being an ill-posed problem - equation
Example of inverse problem being ill-posed problem - equation is not just an analology as - (2)
Not just an analogy – literally like this except not just x and y but ~15000 values you can weight by different amounts to get the value at one sensor
The problem is not constrained enough for there to just be one possible answer
With inverse problem we also can not distinguish -
Kinda like vision as.. (2)
a distributed source near the surface from a deep surface that is more forcal (e.g., deep source could be weak but source near surface could be weak as wlel)
Kinda like vision if something is weak and nearby it is similar to being strong and far away - can’t tell the difference
In inverse problem we can not distinguish many weak sources from one
stronger source
The problems of inverse problem - (3)
- Ill posed problem
- Can’t distinguish a distributed source near the surface from a deep source that is more focal.
- Can’t distinguish multiple weaker sources from one stronger source.
With inverse problem it is a hard problem with no … solutitions
unique
To overcome problems of inverse problem we can - (2)
work out the most likely pattern of brain activity for a given set of data
e.g., x + y = 10, x = 5 y = 5 OR x = 10 , y = 0
2 approaches of solving problems of inverse problem by working out the likely pattern of brain activity for a given set of data - (2)
- Dipole fitting
- Distributed models including minium norm estimation and beamforming
In dipole fitting we won’t get
whole brain map like we get in MRI
Dipole aims to find
a single source that best describes the sensor activity
Example of dipole fitting which says - (2)
- Wr got 15,000 places result could be coming from
- Which one is most likely?
The diapole fitting considers a source at each
possible location in the brain in turn
The dipole fitting runs the forward model for each - (2)
to estimate the pattern of activity at sensors
e.g., if it was activity at this dipole (source) what would it look like at the sensors?
The dipole fitting runs the forward model for each estimate the pattern of activity and each sensor and comapres each of these estimates to
(real)experimental observations
The dipole fitting runs the forward model for each estimate the pattern of activity and each sensor and comapres each of these estimates to real experimental observations and finds…
the location with the smallest error between the predicted and observed activity
The dipole fitting finds the location with smallest error between predicted and observed activity which in other words means:
whichever of possible sources when put through forward model to look like at the sensors, whichever looks most like what you what would get at sensors you say thats your single source
What is the outcome of dipole fitting?
a single location of where dipole is in the brain with an orientation (not an activity map)
Example of dipole fitting outcome - (2)
- On right, somtatosensory response map across sensors at particular time point
- On left we look at brain and one point and arrow (orientation)
(limitation)
The dipole fitting is likely to oversimply
complex (distributed) processes - only one of 15.000 sources causing the response
The distributed methods, unlike dipole fitting, produce a
whole brain activity map at each time point point = current density map
The images of whole brain activity at each time point (current density map in distributed methods ) images can make a
video showing where the active over time - across different conditions
In distributed methods it ask
how much activity is coming from each of the possible sources?
In distributed methods uses many diapoles (sources) in - (2)
set positions and usually set orientations across the brain
put along the cortex in 15,000 places and say which ones are important –> current density map
In distributed methods, use many diapoles in set orientation and position in brain and estimate
one specific solutio of how to get from sensor to source space using inverse model
How does inverse model work in distributed methods? - (2)
Apply this transformation matrix to the activity in sensor space at each time point, to estimate the activity in all the dipoles over time - getting into source space - number of each sensor to number of each dipoles (source)
can apply one inverse model at sensor activity at any time and at any condition which gives different brain maps
Whats two main approaches of distributed methods? - (2)
- Beamforming
- Minimum norm estimation (MNE)
Minium norm estimation uses all the
dipoles to estimate how to get from sensoor to source space in a big calculation
What is the minium norm in minium norm estimation (MNE)?
Of all the different possible solutions based on the forward model, finds the solution that minimises the total activity across the brain = minimum norm
The minium norm estimation minimises the - (2)
the sum of the activity in each dipole squared.
This gives a smooth solution
The miniumum norm in MNE can minimise the sum - (3)
of the activity in each dipole squared.
This gives a smooth solution
e.g., for x + y = 10, gives x = 5, y = 5
Aside from minium norm in MNE iminimies the sum of activity in each dipole squared (smooth solution)
can also optimise in other ways - (2)
e.g., to be more focal
(sparse - not smooth, as in x = 10, y = 0)
There are different approaches in MNE for miniumon norm such as …. which vary… (2)
There are several similar approaches (e.g., LORETA, dSPM, VESTAL)
varying on how the results are scaled to deal with the less sensitive deep areas and how they deal with noise
The main critque of minium norm estimation is it gives smooth results which may reduce
the spatial resolution
Beamforming are filters for
each spatial location ( ‘spatial filters’)
Beamforming asks
what activity from the sensors is caused by a specific source in brain
Beamforming estimate the activity at each location on the cortex
independently
In beam forming, in each location of brain it
estimate its contribution to the sensor data while minimising the effects of other brain regions
In beam estimation it For each location, we estimate its contribution to the sensor data, while minimising the effect of other brain regions (and noise)
Minimising this depends on
on knowing the relationship between the sensors, so depends on the data, and often between sensors in pure noise (empty room with no participant and taking MEG) data
In beam forming it works out - (3)
Works out how much a location’s activity depends on activity at each sensor.
Uses this to weight activity at each sensor.
Adds the weighted sensor activity together to get the activity at this location
Many different algorithims for Beamforming such as - (3)
- Least Constrained Minium Variance
- SAM (Synthetic aperture magnetometry) - non-linear beamforming approach
- DICs (Dynamic Imagning of Coherent Sources)
When to use DICs in beamforming? - (4)
You have multiple brain sources with potentially overlapping activity.
You want high spatial resolution to distinguish closely spaced sources.
Your data has good signal-to-noise ratio (SNR).
You’re interested in source dynamics (time-frequency analysis).
When to use SAM in beamforming? - (5)
You have a single dominant brain source or well-separated sources.
You have limited data or poor SNR.
You prioritize computational efficiency.
SAM is employed when studying brain activity and specifically focusing on spatial and temporal dynamics.
SAM is good at localizing brain sources, especially when these sources are dynamic or vary in space and time.
When to use LCMV in beamforming? - (5)
You need a basic spatial filter without strong assumptions about source location or orientation.
You have a single dominant source or well-separated sources.
You prioritize simplicity and interpretability.
LCMV beamforming is advantageous when you have prior information or want to focus on a specific brain source amidst background noise or interference.
It allows for the selective enhancement or suppression of specific brain sources based on prior knowledge or spatial filtering
Beam forming main critque is that
misses correlated signals in different locations (i.e., two regions working togheter) - opposite to MNE
In MEG/EEG results wen can transform the data /results into source space and visualise it either in
whole brain cortex or look in depth at one small area with region of interest (ROI)
Looking at one small area over time in MEG results it is called what in brainstorm
scout
What is a scout?
A collection of vertices on the cortical surface
Scout/ROI can be defined - (3)
- Based on coordinates
- Based on activity from localiser experiment
- Using anatomical atlases
The data is summarised across the scout by..
calulating the average OR absolute average OR maximum response etc..
The time-course from each participants’ ROI and scout can be used for
group-level anlayses
Source Localisation summary
To get from sensor space to an estimate of source space (source localisation) - (6)
We create a head model (spheres or more realistic)
We define our source space (dipoles across the cortex)
We build a forward model in the opposite direction (relatively easy)
We build an inverse model using one of these approaches (hard)
which gives us one reasonable solution
If we use distributed method, we can get whole brain activity maps and extract ROIs - scouts
