Preprocessing Flashcards
What’s the difference between functional and structural MRI?
Structural MRI provides an anatomical image of the human brain
Functional MRI measures changes in brain activity over time
What is the “field of view” and “matrix size” and what can we derive from them?
The field of view (FOV) is the width and length of a slice, usually 192mm*192mm.
The matrix size describes how many voxels are in the slice, e.g. 64x64.
These can be used to find the in-plane resolution, by dividing the FOV size by the matrix size, which gives the voxel size. 192/64 = 3mm voxel size. The height of the slice also describes the height of the voxels.
Temporal Preprocessing: Slice Timing
Why is it done?
How is it done?
When acquiring slices from the scanner, they are scanned sequentially. This means that two slices in the same volume were not acquired at the exact same time. We want to temporally align all the slices in relation to a reference time
This is done using “Fourier transformation”. For each voxel, the signal over time is described with an equation, and the equation is used to predict the activation at the timepoint of the reference slice. This is also called a Phase shift.
Keep in mind! when performing slice timing, different reference slices can give different statistical result
Spatial Preprocessing: Realignment
Why?
Subjects move their head in the MRI scanner, which shifts the measured signals between voxels. This can cause a reduced sensitivity by failure to detect local activations. Similarly, if the head motion is correlated to the experimental paradigm, this can cause reduced specificity and spurious activations.
Spatial Preprocessing: Realignment
How?
Two steps:
Estimation(Registration)
Reslicing(Resampling)
Estimation determines the rigid-body transformation at each scan, in relation to a mean scan or sometimes the first scan. Rigid-body transformation is how much the object(head) rotates around 3 axes, pitch, roll, yaw. For each scan, the difference to the mean mean scan per voxel is calculated.
This is then used to perform Reslicing/resampling, which shifts the scan so it fits with the mean image.
Spatial Preprocessing: Coregistration
Why?
How?
Before performing Normalization, the structural MRI needs to be in alignment with the functional MRI. This allows for a more accurate localisation of activations.
It is done using the same method as Realignment, where the structural MRI scan is realigned to the mean image extracted from the realignment.
Spatial Preprocessing: Segmentation
Why?
Segmentation is used to categorise a MR image into different types of tissue. - Gray matter - White matter - Cerebro-spinal fluid - Meninges - Skull - Air The segmentation also improves the normalization.
Spatial Preprocessing: Segmentation
How?
In SPM, there are tissue probability maps, which act as priors in the analysis of MR images. Based on these probability maps and the images, the segmentation provides a posterior probability that a certain brain area is a specific type of tissue.
Spatial Preprocessing: Normalization
Why?
There are substantial inter-individual differences in brain anatomy, which we want to account for. When using multiple subjects in the same analysis, normalization increases the sensitivity. Similarly, it is nice to make results from different studies comparable, by brining them into the same coordinate system (MNI Space)
Spatial Preprocessing: Normalization
How?
Two steps:
Linear registration
Non-linear registration
Linear registration consists of adjusting the global differences between the standard brain and the mean scan (e.g. brain size or skew). It is very similar to a rigid-body transformation, as the brain is adjusted based on 4 parameters: translations, rotations, zooms and sheers, which all have 3 axes. Thereby, 4x3=12 parameters describe how the mean scan has to be adjusted to fit the standard brain.
The non-linear registration deals with local differences - e.g. differences in specific gyri. This is done using deformation fields, which are 3d matrices that describe how voxels from the mean scan have to adjusted to fit the reference.
What is the relationship between normalization and segmentation?
It is circular!
This is because knowing which tissue type a specific voxel belongs to can help with the normalization. On the other hand, knowing where a voxel is in standard space helps the segmentation.
Therefore, often a generative model is made, which account for both at the same time.
Good luck following the arrows
Describes the usual preprocessing steps:
A mean image is calculated during realignment. The mean image is then coregistered to the structural scan, which is normalized to a standard brain. The normalization parameters are then used on the functional scans to normalized everything.
Or maybe neuroscience isn’t for you and you do something completely different? Open a bakery or move to the countryside or do mushrooms at Warschauer Straße the world is your oyster
Spatial Preprocessing: Smoothing
Why?
Even after normalization, there is still a residual anatomical mismatch between participants. To account for this mismatch, the activation is smoothed so that activation in voxels close to each other become more similar. This also reduces measurement artefacts, and increases the signal-to-noise ratio
Spatial Preprocessing: Smoothing
How?
By convolution with a 3D gaussian kernel. This basically means that a “kernel” moves through all voxels and transforms the value of the voxel so that it is a combination of its own value and all the neighbouring values.
The value “Full width at half maximum” FWHM describes how large the kernel has to be. The larger the kernel, the more neighbouring voxels are taken into account when smoothing.
Further preprocessing: Masking
If you’re only interested in specific voxels, masking can be utilised. E.g. to extract only gray matter voxels.
