01/02/2022 HC 01: Human Neuroimaging
Used in neuroscience to relate brain processes to behavioral traits:
- Performance on cognitive tasks
- Inter-individual differences
- Disease conditions
But there are theoretical and practical pitfalls and recognizing them is important.
Some human neuroimaging techniques:
- EEG
- Radioactive ligand imaging
- FMRI
- Structural MRI
- Optical imaging etc.
Can be divided into structural and functional clusters.
Structural:
- Magnetic resonance based
a. Voxel based morphology = density of the tissue, related to number of
synapses/neurons and compare voxels among people
b. Diffusion mRI = how much fibers are there
c. MR spectroscopy = resonance frequency of several molecules)
- X-rays
- Ultrasound echography
- Metabolic
Functional:
- Electrophysiology (EEG, MEG, local field potentials etc.)
- Blood oxygen level dependent (optical or fMRI)
- Metabolic (PET now different and can say something about how much glucose is
being metabolized and you can say something about the function)
The field is exploding with options within human neuroimaging (see picture) and this brings
difficulties with it. It is also a good thing to pick the technique based on the size you want to
look at based on temporal and spatial density.
,FMRI - theory and data handling
Issues - fMRI needs extensive processing and no general agreement and requires special
software tools, You have some basic steps:
1. Signal noise (both EEG and fMRI)
a. Electrode resistance/physiological noise, muscle artifacts, radio waves from
outside > EEG
b. Hardware/temperature fluctuations, head motion (debate on how to account
for it and is a big one), physiological movements > MRI (it changes the
magnetic field and therefore the signal)
2. Summary per participant
a. Regression model to control it, but if it is related to your signal you throw a lot
out since you take it as a regressor into your model. If you can model it, you
can control it. You predict the variance at subject level for every voxel,
common regressors are task regressors, head motion parameters,
physiological measures, etc.
b. Beneath you see what it looks like (first level statistics), you start at a certain
time with collecting data and at every time point you indicate what is going on
(task regressors, because you get instructions for example). For every voxel
in the brain within that volume you have a regression analysis > how well
does the bèta wave in that voxel predict the signal > baseline from condition.
,3. Group level analysis
a. You then have to do this on a group level according to an ANOVA, MANOVA,
ANCOVA model etc. and that modeling is tricky, because if you do it not right
you end up with different results
b. Carefully construct your design
- 2 conditions: stress and methylphenidate.
- People had stress or no stress and if you want to know the main effect
to stress you put them in a certain column (1) and you can also do this
for methylphenidate
- Are you going to model the mean/add in other factors/model individual
factors (if you do not do this, you blow up your significance of results)
c. Normalisation
- Every brain in the same dimension > average your activation over
subjects, you want them to have them similar. However, if you want to
compare AD to a healthy brain, you do not want to normalize them
fully > optimum in normalization.
d. Choose statistical test
e. Choose MCC methods (multiple comparisons)
, 4. Correction for multiple comparisons
- EEG has 24 to 128 electrodes and for fMRI you have 100k-300k voxels and
this ends up for 3200 false positives for EEG and 10.000 for fMRI with alpha
level of 0,005 and for the Bonferroni correction this is ridiculous and you are
asking something unrealistic from your data > enormous effect for
significance.
- Rationale = voxels are activated close to each other, much more likely to be
together activation, however that has also has problems and there is inflation
of false-positive rates > shows you that this field is still progressing
- Reduction is a basic principle for fMRI
Seems that DLPFC does one function and ACC the other, but we are ignoring the functional
and structural connectivity between them and integration of information.
Realistic expectations:
Negative tone:
- Imaging only a proxy for neural activity
- Very noisy
- Advanced statistics with big multiple comparisons problem
- Data reductions complicate interpretation > looking at regions of interest is not
informative for the entire brain > see above example!
Require:
- Understanding the imaging modality
- Understanding the processing and analysis steps
EEG - theory and data handling
Hans Berger described the alpha wave in humans in 1924 and EEG is much older than fMRI
(1990). We use these techniques now more to understand brain disorders and use these
actually in the clinic.
EEG is used for clinical application as for example EEG used to look at people with
movement disorders (tracking DBS network effects and non invasively via EEG to see how
we move). Enhanced beta activity in basal ganglia to the primary motor cortex and other
Used in neuroscience to relate brain processes to behavioral traits:
- Performance on cognitive tasks
- Inter-individual differences
- Disease conditions
But there are theoretical and practical pitfalls and recognizing them is important.
Some human neuroimaging techniques:
- EEG
- Radioactive ligand imaging
- FMRI
- Structural MRI
- Optical imaging etc.
Can be divided into structural and functional clusters.
Structural:
- Magnetic resonance based
a. Voxel based morphology = density of the tissue, related to number of
synapses/neurons and compare voxels among people
b. Diffusion mRI = how much fibers are there
c. MR spectroscopy = resonance frequency of several molecules)
- X-rays
- Ultrasound echography
- Metabolic
Functional:
- Electrophysiology (EEG, MEG, local field potentials etc.)
- Blood oxygen level dependent (optical or fMRI)
- Metabolic (PET now different and can say something about how much glucose is
being metabolized and you can say something about the function)
The field is exploding with options within human neuroimaging (see picture) and this brings
difficulties with it. It is also a good thing to pick the technique based on the size you want to
look at based on temporal and spatial density.
,FMRI - theory and data handling
Issues - fMRI needs extensive processing and no general agreement and requires special
software tools, You have some basic steps:
1. Signal noise (both EEG and fMRI)
a. Electrode resistance/physiological noise, muscle artifacts, radio waves from
outside > EEG
b. Hardware/temperature fluctuations, head motion (debate on how to account
for it and is a big one), physiological movements > MRI (it changes the
magnetic field and therefore the signal)
2. Summary per participant
a. Regression model to control it, but if it is related to your signal you throw a lot
out since you take it as a regressor into your model. If you can model it, you
can control it. You predict the variance at subject level for every voxel,
common regressors are task regressors, head motion parameters,
physiological measures, etc.
b. Beneath you see what it looks like (first level statistics), you start at a certain
time with collecting data and at every time point you indicate what is going on
(task regressors, because you get instructions for example). For every voxel
in the brain within that volume you have a regression analysis > how well
does the bèta wave in that voxel predict the signal > baseline from condition.
,3. Group level analysis
a. You then have to do this on a group level according to an ANOVA, MANOVA,
ANCOVA model etc. and that modeling is tricky, because if you do it not right
you end up with different results
b. Carefully construct your design
- 2 conditions: stress and methylphenidate.
- People had stress or no stress and if you want to know the main effect
to stress you put them in a certain column (1) and you can also do this
for methylphenidate
- Are you going to model the mean/add in other factors/model individual
factors (if you do not do this, you blow up your significance of results)
c. Normalisation
- Every brain in the same dimension > average your activation over
subjects, you want them to have them similar. However, if you want to
compare AD to a healthy brain, you do not want to normalize them
fully > optimum in normalization.
d. Choose statistical test
e. Choose MCC methods (multiple comparisons)
, 4. Correction for multiple comparisons
- EEG has 24 to 128 electrodes and for fMRI you have 100k-300k voxels and
this ends up for 3200 false positives for EEG and 10.000 for fMRI with alpha
level of 0,005 and for the Bonferroni correction this is ridiculous and you are
asking something unrealistic from your data > enormous effect for
significance.
- Rationale = voxels are activated close to each other, much more likely to be
together activation, however that has also has problems and there is inflation
of false-positive rates > shows you that this field is still progressing
- Reduction is a basic principle for fMRI
Seems that DLPFC does one function and ACC the other, but we are ignoring the functional
and structural connectivity between them and integration of information.
Realistic expectations:
Negative tone:
- Imaging only a proxy for neural activity
- Very noisy
- Advanced statistics with big multiple comparisons problem
- Data reductions complicate interpretation > looking at regions of interest is not
informative for the entire brain > see above example!
Require:
- Understanding the imaging modality
- Understanding the processing and analysis steps
EEG - theory and data handling
Hans Berger described the alpha wave in humans in 1924 and EEG is much older than fMRI
(1990). We use these techniques now more to understand brain disorders and use these
actually in the clinic.
EEG is used for clinical application as for example EEG used to look at people with
movement disorders (tracking DBS network effects and non invasively via EEG to see how
we move). Enhanced beta activity in basal ganglia to the primary motor cortex and other
