fMRI Flashcards
What is another name for fMRI?
Blood Oxygenation Level Dependent Magnetic Resonance Imaging
== BOLD MRI
What is the sole purpose of fMRI?
We want to use MRI to produce activation maps in response to some kind of task
–> Taking the brain and finding out which regions are activated under a specific task condition
What is the mechanism of fMRI?
The fMRI BOLD signal is altered due to the increase of blood flow in response to brain activity
Stimulus = Neuronal activity = NVC = Haemodynamic response = Detection by MRI scanner = fMRI BOLD response
What did Angelo Mosso study?
Subject laid on a delicately balanced table which could tip downward either at the head or at the foot if the weight of either end were increased
The moment emotional or intellectual activity began in the subject, down went the balance at the head end, in consequence of the redistribution of blood in his system
How does blood flow change the fMRI signal?
The BOLD sequence isn’t measuring the blood flow, but is sensitive to the different magnetic properties of oxygenated and deoxygenated red blood cells
The increased magnetic field in the deoxygenated blood causes a magnetic field gradient resulting in adjacent water molecules precessing at different frequencies (larmour equation –> increase in magnetic field = protons speed up)
As water molecules are precessing at different frequencies in the gradient field, they dephase causing a loss of signal around deoxygenated blood compared to oxygenated blood
What is magnetic susceptibility?
Indicates whether a material is attracted into or repelled out of a magnetic field
Paramagnetic materials align with the applied field and are attracted to regions of greater magnetic field
Diamagnetic materials are anti-aligned and are pushed away, toward regions of lower magnetic fields
What is the magnetic susceptibility of oxygenated and deoxygenated blood?
Oxygenated blood = Diamagnetic = slightly reduces the magnetic field (it is oriented in the opposite direction to the main magnetic field)
Deoxygenated blood = Paramagnetic = adds to the magnetic field (Oriented in line with the main magnetic field)
Explain activation contrast
At rest, oxygenated blood travels through the arteries and exchanges in the capillary bed, deoxygenated blood leaves via the veins
The deoxygenated blood increases dephasing, reducing T2* signal
During activation, an excess of oxygenated blood flows into the activated region, swamping the deoxygenated blood
The excess of oxygenated blood reduces dephasing, increasing the T2* signal
What is the reason for the oversupply of oxygenated blood?
The reasons for the oversupply of oxygenated blood are still unclear, but may be related to the supply of other nutrients such as glucose
It could also be to provide a strong enough concentration gradient across the capillary wall so that the oxygenated blood can diffuse efficiently
An alternative hypothesis is that it is related to blood flow being controlled at the arteriole level - could be due to the control of the blood flow being higher up the vascular tree so to be safe of getting sufficient blood flow it increases it higher up the tree
What are we actually imaging in fMRI?
Not imaging individual red blood cells or even vessels
Imaging voxels are an average signal over mm3, typically 4 x 4 x 4 = 64 mm3 in fMRI
An fMRI voxel will contain many different vessels, but these still only make up around 3% of the voxel volume
What can be found in an individual voxel in fMRI?
Approx. =
6 million neurons
250km of axons
5 x 10 (to power 10) synapses
25km of dendrites
How do we image through k space?
We move through k-space by the negative phase encode and readout gradients
- Initially we have slice select at 90 degree pulse, selects a slice in the z direction
- The first line of k-space is acquired with a positive readout gradient during the first TR after applying the largest negative phase encode gradient
- During the second TR, the phase encode gradient is slightly less negative to read the second line of k-space
- We continue the process, incrementing the phase encoding gradient strength to read a different line of k-space during each TR
- Finally we apply the strongest positive phase encoding gradient to read the last line of k-space
What is a problem with acquiring an image in this way?
This process is far too slow for functional brain imaging
What method is used to image the BOLD signal which is faster?
Echo planar imaging
Explain echo planar imaging
- Negative y gradient, negative x gradient (moves southwest)
- Positive x gradient (moves along k space to the right)
- Apply a gradient blip in y axis (moves up k space tiny amount)
- Negative x gradient (moves back along k space to the left)
- Another gradient blip to move up k space
- Positive x gradient (move to right of k space)
- Continue until we have filled k space
What are the advantages of EPI?
Can be done in one TR
Means we can image the BOLD effect with a spin echo sequence
Single slice in under 100ms, so whole brain volumes in the order of a second or two
The whole of k space is acquired after a single 90 degree pulse - single shot imaging
What is EPI acquisition?
Echo planar imaging acquisition
What is a disadvantage of BOLD EPI acquisition?
Very low quality image
EPI is notorious for artefacts
Why does EPI create such a low quality image with artefacts?
Arises due to the imperfections in the rephasing-dephasing cycle of the rapidly switching bipolar frequency-encode gradient
We do it all after one excitation, so we don’t get a lot of signal at the two extremes, and those spaces at the edge of k space give a sharper image with more detail
What is the difference between sequential and interleaved slice acquisition?
Sequential slice acquisition acquires each adjacent slice consecutively, either bottom-to-top or top-to-bottom. Interleaved slice acquisition acquires every other slice, and then fills in the gaps on the second pass
Why do we normally interleave the slices?
Interleaving allows the use of a lower sampling bandwidth with a significant increase in signal-to-noise. The method also has the advantages of relative ease of implementation, no need for postprocessing to remove image distortion, and no need for shimming on a case-by-case basis
What causes signal drop out in BOLD MRI?
Huge susceptibility gradient between air and tissue which can cause signal drop out
How can we solve the problem of signal drop out in BOLD MRI?
Can reduce TE (echo time) and flip angle to manage signal drop out
What is a problem with reducing TE and flip angle?
It also reduced T2* weighting of the sequence and therefore reduces the BOLD signal that we measure
Why do we acquire dummy scans during BOLD imaging?
The first sequences have different brightness and contrast so we acquire dummy scans allowing the signal to reach steady state
What is the next step after acquiring the BOLD signal images?
We now have a BOLD EPI temporal imaging series and now we want to investigate the signal intensity time course voxel by voxel
How do we model the BOLD signal?
Model based
1. Linear regression
2. Nonlinear regression
Model free/data driven
1. PCA (principal component analysis)
2. ICA ( independent component analysis)
What is the purpose of using a model in BOLD analysis?
We want to fit a model to our data so that we can analyse which voxels show significant activation
Use a model to fit voxel signal time course
What is a problem with the model?
Model is a poor fit to the data - the data is clearly shifted in time with the response delayed compared to the stimulus
What is HRF?
Haemodynamic response function = the change in MR signal triggered by instantaneous neuronal activity
How do we account for the shift in the data?
Try and model the physiology - this is done using the haemodynamic response function which characterises the response to an impulse stimulus
It was derived experimentally and shows that the vascular response is delayed by a few seconds, peaking at around 5s before decaying
We can convolve the haemodynamic response function (HRF) with our experimental model to produce a more realistic estimate of the expected signal change
We can then fit the HRF convolved model to our data - it provides a much better fit to the voxel signal time course, but there appears to be some residual signal drift
How do we account for the residual drift in signal over time?
Add a linear ramp to the model
Next we need a value that tells us how well our model fits the data - how is this achieved?
Our data is a series of discrete signal measurements taken every few seconds
First thing is to estimate the amplitude - this is generated by the GLM fitting process
We then need to estimate the residual noise - calculated from the distance between the model function and acquired data at each point
What are EVs?
Explanatory variables
== a set of idealized predictions of what the hemodynamic response function (HRF) should look like if a voxel of interest became activated due to a task or stimulus
What is the General Linear Model?
GLM says that Y (the measured fMRI signal from a single voxel as a function of time) can be expressed as the sum of one or more experimental design variables (X), each multiplied by a weighting factor (β), plus random error (ε)
voxel time series data = design matrix (made up of EVs) x regression parameters + gaussian noise
What does the t statistic represent?
The ratio of the departure of the estimated value of a parameter from its hypothesised value to its standard error
== The ratio of the fitted amplitude to the residual noise
What creates a high or low statistic?
big amplitude / little residual noise = high
small amplitude / high residual noise = low
Is fMRI quantitative?
NO - cannot get a single subject to perform a task and measure activation in any meaningful units
fMRI relies on the signal contrast generated between states
What is an epoch?
A period of sustained neural activity
What is an event?
An isolated occurrence of a stimulus being presented or a response being made
What is ISI?
Inter-Stimuli Interval = time between successive stimuli
What is ITI?
Inter-Trial Interval = time between successive onsets of trials
What is SOA?
Stimulus onset asynchrony
What are the different experimental designs used in fMRI?
Block designs
Slow event related designs
Fast event related designs
Mixed design
Explain block designs
Earliest fMRI design mirroring that used in PET studies
Consists of many closely spaced successive trials over a short interval of time
Block design experiments utilise blocks of identical trial types to establish a task-specific condition
In a two condition block design what block lengths are statistically optimal?
Around 20 seconds
How is a model fitted to a block design?
The block design is convolved with the HRF to generate a GLM regressor
This is the prediction of what the measured (noise free) signal in an activated voxel should look like following our block design stimuli
What are the advantages of block designs?
It is the most efficient design for detecting signal amplitude differences between conditions
Fairly robust when there is uncertainty in the timing/shape of the HRF, as block duration is usually larger than HRF response
Can acquire more trials in less time than other designs because you don’t have to worry about spacing individual trials apart to get an estimate of each individual event
What are the disadvantages of block designs?
Stimuli are highly predictable - subjects know what is coming and may alter strategies accordingly
Inflexible for more complex tasks - impact of oddball stimuli, or stimuli/events that occur uncontrollably, cannot distinguish between trial types within a block
Does not account for transient responses at the start or end of a block (might behave differently from start of the block to end of the block after seeing stimuli several times)
Does blocking trials change the psychological process you are interested in?
Determining an appropriate baseline condition can be challenging
What are slow event related designs?
Consist of short stimuli separated by a fairly long inter-stimulus interval (ISI) to enable the HRF to fall back to baseline before the next trial
What is the recommended ISI in slow event related designs?
8s + 2 x stimulus duration
Effective loss of power is approx 35% so 9 minutes of slow event related design is equivalent to around 6 mins of a block design
How is a model fitted to slow event related designs?
As with block design, the slow ER design is convolved with the HRF to generate a GLM regressor
The ISI is set so that the HRF has decayed back to baseline before the following stimulus is presented
What are fast event related designs?
Consist of short stimuli separated by variable ISI
The inclusion of jittered fixation frames allows for more closely placed trials
Events can be truly randomised as you would do in a behavioural study
BOLD signal change is much lower even than in slow ER designs
Care with design is required - every combination of trial sequences must be used i.e., every trial type must be preceded and followed by every other trial type an equal number of times
What is jittering and why is it essential in fast event related designs?
Variable duration between trials
Needed in order to distinguish between the BOLD responses of multiple conditions
Adding jitter to experiments with conditions that are relatively close to each other (e.g., less than 10-15 seconds apart) allows the independent estimation of the hemodynamic response to each condition.
What is good practice in an fMRI experimental design?
- Evoke the cognitive or other process of interest
- Collect as much data as possible - because data is small and noisy
- Collect data on as many subjects as possible
- Choose stimulus and timing to create maximal change in BOLD signal for the cognitive process of interest
- Time stimuli presentation of different conditions to minimise overlap in signal
- Use software to optimise design efficiency for ER designs
- Get measure of subject behaviour in the scanner (ideally related to task) e.g., falling asleep or not concentrating
When can we reject the null hypothesis in a simple block design?
For a simple block design, we only have one regressor so all we can test is whether there is significant activation
If our t-statistic is greater than a specified value we can reject the null hypothesis and conclude that there is significant activation in our voxel
What are COPES?
Contrast of Parameter Estimates
Describe the statistical comparisons we want to make
These are input into software using contrast weight vectors
Describe the different vectors which would be applied if our two stimuli correspond to displaying faces (B1) and displaying objects (B2) for example
[1 0] B1 = Voxels where activation occurs due to faces
[0 1] B2 = Voxels where activation occurs due to objects
[1 1] B1 + B2 = Mean activation due to faces and objects
[1 -1] B1 - B2 = Voxels where face activated more than objects
[-1 1] = B2 - B1 = Voxels where objects activate more than faces
Describe COPEs from an imaging perspective
The GLM is fitted in every voxel
Therefore, we have values of beta and e (sigma squared) in every voxel, so we can produce images of them
How many regressors do we get in slow event related designs?
Three regressors and three parameter estimates beta 1, 2, 3
We can make many different statistical comparisons by assigning different contrast weight vectors
How do we decide which our voxels show statistically significant activation?
If we plot our t statistic, it is t distributed under the null hypothesis that beta = 0
p value = p(t > t’/ B = 0)
A small p value implies that the null hypothesis is unlikely to be true
A threshold can be set for the p-value (normally 0.05 which corresponds to a false positive rate of 5%
What is a significant p value in fMRI terms?
If the p-value is less than 0.05, the voxel is active
What does the t distribution depend on?
Degrees of freedom (DOF)
Similar t values mean very different things depending on the DOF
How can t-values be standardised so they can be interpreted in the same way?
A probability preserving transform can be applied to transform t-values into z-values
The z-distribution is just a standard normal distribution - this means that our t values can always be interpreted in the same way
What is the multiple comparisons problem?
The more hypotheses are tested on a particular data set, the more likely it is to incorrectly reject the null hypothesis
A p-value < 0.05 sets the FPR at 5% which means we expect around 5% of activations to be false positive - this becomes a problem when doing multiple comparisons
== Dead salmon effect
How do we overcome the multiple comparisons problem?
Multiple comparison correction methods
Name some of the multiple comparison correction methods
Bonferroni - strong control over false positives, least sensitive
Gaussian Random Fields - strong control over false positives, somewhat conservative
False-Discovery Rate - admits false positives, more lenient
Cluster enhancement
What is a problem with Bonferroni corrections in fMRI?
Bonferroni assumes that all tests are independent
Assumes all voxels are independent which is not the case, all from the same person and likely that neighbouring voxels share the same value so are correlated to an extent
What is first level analysis?
Analysis of fMRI data in individual subjects, making statistical inferences about which regions activate in those subjects
What is second/higher level analysis?
Following first level analysis, analysis of groups of data
What is the main research question of second level analysis?
Does the group activate on average?
What is essential for making voxel-by-voxel comparisons?
Need to register all the images into a common anatomical coordinate system = standard space
Explain the steps in group analysis
- Register all images into the standard space
- Each subject will have a beta value and an associated error value calculated from the first level fMRI analysis
- Can formulate the problem as another GLM
What are the two ways we can formulate the second level GLM?
Fixed effects model
Mixed effects model
Explain fixed effects models
These ignore the group variance, only considering the variance of the original MRI data propagated from the first level in the beta estimates
Explain mixed effects models
These consider the group (between subject) variance as well as the variance in the MRI data
Mixed effects analysis generalises to the population as a whole, while the fixed effects analysis is only applicable to the specific subjects studied
What pre-processing is needed in fMRI?
- Brain extraction
- Motion correction
- Slice timing
- Spatial smoothing
- Temporal filtering
- Unwarping
- Registration
What is the purpose of brain extraction?
Remove non-brain tissue for registration and masking purposes
What is the purpose of motion correction?
Ensures consistent anatomical coordinates between images
What is the purpose of slice timing?
Ensures consistent acquisition timing (uses temporal derivative in practice)
What is the purpose of spatial smoothing?
Spatial smoothing of functional MRI (fMRI) data is a standard pre-processing step used to increase signal-to-noise ratio (SNR) through noise averaging, to assist with cross-subject registration, and to help ensure that statistical assumptions regarding the smoothness of the images are met
What is the purpose of temporal filtering?
Highpass filtering to remove slow signal drifts
Voxel time courses of fMRI data often show low-frequency drifts, which is thought to be caused by physiological noise as well as by physical (scanner-related) noise. If not taken into account, these signal drifts reduce substantially the power of statistical data analysis
What is the purpose of unwarping?
Corrects for B0 inhomogeneity induced image distortion
What is the purpose of registration?
Registers images into a consistent coordinate frame e.g., standard space for group analysis
