What is the streaming order of pre-processing steps?
It depends on the software package, but in SPM it is:
- Distortion correction
- Motion correction
- Slice timing correction
- Spatial normalisation
- Spatial smoothing
- Statistical analysis
What are stereotactic spaces?
The x,y,z co-ordinates of the images in fMRI are derived from the native space of the MRI scanner. However, not all individuals or the same individual at different time points will align within the native space. Stereotactic spaces are templates to which individuals brains are standardised.
Talariach co-ordinates. This was the original stereotactic space, but it was developed from one persons brain.
MNI co-ordinates. These are more commonly used as it was developed from a large number of patients and thus is closer to the 'average' brain, reducing the transformations made.
Radiological conventions of neuroimaging representation
The image is flipped relative to the body - the right side of the brain is on the left side of the image. The image thus aligns with the patient when viewed from the foot of the bed.
Neurological conventions of neuroimaging representation
The left side of the brain is on the left side of the image.
The cost function is how misaligned two images are. A transformation model specifies the parameters needed to align the images. A simple model only has a few changes, and thus aligns gross but not fine aspects of the images. A more complex model can align fine features.
A good cost function should be small for images closely aligned, and become larger when they are more misaligned.
When a transformation model is applied to an image, the transformed co-ordinates do not fall ontop of old co-ordinates. It is therefore necessary to calculate intensity values at those intermediate points. This process is called interpolation.
These are linear transformations i.e., points that fell on a line before the transformation will continue to fall on a line afterwards.
Transformation: Shifting along an axis
Rotation around an axis
Scaling: 'stretching' along an axis.
Shearing: changing angles between points on a line
Each can be performed for each dimension. A full affine transformation where all of these are carried out across all three dimensions would be described by 12 parameters.
Rigid body transformations
For movement however, we can assume the image is moving in space without changing size or shape, therefore motion transformations can be described by only 6 parameters (3 transformations, 3 rotations). This is a rigid body transformation.
Piecewise linear transformation
The image is broken up and different affine transformation parameters are applied to piece of the image
These transform co-ordinates such that they no longer need to remain linear with one another.
T1 weighted images
White matter is brighter than grey matter
What is fMRI measuring?
When neurons become more active, the blood flow to that region of the brain increases to supply more oxygen for the increased metabolic demands of the cells i.e., the hemodynamic response. However, the increase in blood flow supply surpasses the increase in oxygen demand, and there is a surplus of oxygen. fMRI measures the increase in oxygen – the blood level oxygen dependent signal (BOLD). The BOLD response for a stimulus is mapped using the HRF, which assumes linear relationship between stimulus and neural response, and between neural response and BOLD response
The simplest fMRI method is to use a paired t-test to compare the brain between task and resting states. This requires long blocks of stimulation to allow the signals to reach steady states. This wastes the temporal resolution of fMRI. The t-test also does not consider the temporal structure of fMRI data, which violates the assumptions of the statistical model. Block designs however provide optimal sensitivity to differences between conditions.
Event related design
Event related design can be used to investigate the individual impact of brief stimuli. This was initially done by leaving large time periods between events to allow hemodynamic response to return to baseline – but this is not time efficient.
Events occurring more rapidly needed to be modelled. Eventually, the range of fMRI event designs for which the BOLD response behaved linearly (i.e. HRF can still be extracted and mapped for events close in time) was established. (Events must be separated by roughly 2 seconds). This linearity is essential, as it allows for simple analysis using general linear model.
Fixed effects analysis links the timeseries for all individuals and carries out analysis across all time points, ignoring the fact that the data are repeated measures from different individuals. This means that data from a single subject can drive results to significance within an analysis, therefore inferences cannot be made about the wider population.
Mixed effects analysis works instead by first working out the effect per subject per voxel, and then carries out a second level analysis to test for effects across subjects. This allows inferences to be made about the wider population. This approach however does not account for intraindividual variability.
MVPA measures patterns of activity across a region, rather than voxel-level analysis. It attempts to decipher the degree to which conditions can be separated on the basis of patterns of activity, and the type of information present within the patterns.