Preprocessing Flashcards
What are two possible kinds of data that we work with in MRI preprocessing/analysis?
- structural data (mostly acquired using T1 MPRAGE sequence): anatonical image of the brain
- functional data (mostly acquired using T2* EPI sequence): changes of brain activity over time
Define preprocessing and statistical analysis.
in MRI of course (:
Preprocessing:
Algorithmic correction of temporal and spatial artefacts, that arise from measurements
Statistical analysis:
Statistical model estimation and parameter inference on experimental variations of interest
Explain the terminollgy of fMRI (for experiments)
- Subjects: people who perform an experiment
- Session: a single appointmen for testing
- Run: A run of an experiment (multiple possible per session)
in SPM Runs are called sessions
Explain the terminology of fMRI (focus on data)
- In each run a number of volumes is acquired
- Each volume consists of different slices
- Each slice consists of differnt voxels
- The thickness of a slice is equivalent to the height of a voxel
- The in plane resolution describes the width and depths of voxels
- the Matrix size describes the amount of voxels in a slixe
- Matrix size and in plane resolution make up the field of view
What are the three brain axis in neuroimaging
- Coronal: view from front
- Sagittal: view from the side
- axial/transversal: view from above
Name the three main objectives of the SPM module for Matlab
- Displaying data
- preprocessing data
- analyzing data
Name 4 alternatives to SPM for preprocessing.
- AFNI
- FSL
- Fressurfer
- fMRIprep
Name the six typical steps of preprocessing.
IMPORTANT!
- Slice timing
- Realignment
- Coregistration
- Segmentation
- Normalization
- Smoothing
Why is slice timing necessary?
- The volume is acquired as a number of sequentially acquired slices. -> Each slice is acquired at a different point in time
- By using slice timing we are correcting the data, to estimate what the volume would look like if each slice were reorded simultaneously
- Slice time = Time of repition / Number of slices
- We can either acquire slices in ascending- , descending slice order or ascending interleaved slice order
How is slice timing performed mathemtically?
- Data from each each voxel is Fourier transformed
- Data for each time course per voxel is interpolated.
- Signal for each time point per voxel is phase shifted to the timepoint where the reference slice was recorded
- Data is backtransformed into signal space
What do we receive after performing slice timing in SPM?
- We receive volumes that are corrected to be acquired at a single timepoint per volume
- visually we can not distinguish slice timed and unprocessed images
- however the choice of selecting a certain reference slide is relevant to our statistical analysis (different results!)
Why do we perform Realignement?
- Subjects move during MRI acquisition, even if fixated and asked not to move, this can have certain effects:
- Activation shifts between voxels (i.e. activation that is in voxe a will appear in voxel b after movement) -> failure to detect local activations -> reduces sensitivity of analysis
- Experimental paradigm may be connected to movement -> spurious activations -> reduces specificity
What two steps are performed during realignement?
…and how is realignement performed?
- Estimation/Registration: determine rigid-body transformation from each acquired image to reference by minimizing the sqaured difference between original and realigned images (rigid-body transformation consists of 3 translation and 3 rotation parameters that are calibrated)
- Reslicing/Resampling: Applying estimated transformations to correct whole series -> register image first, then resampled image. Resampling = the transformation from one grid to another
What does realignement return?
A series of realigned images, the mean functional image and the values of realignment parameters over time.
Why should we perform coregistration?
- Before mapping images to standard space we must align the structural image to the functional images.
- This allows for a more precise anatomical localization of activations
- This relies on the same algorithms as realignment
What do we have to specifiy for coregistration in SPM?
- source image (usually structural T1 image)
- image that source image is realigned to (usuayll mean functional scan)
What is the output of coregistration in SPM?
- The critertion optimized to fit reference (e.g. normalized mutual information)
- The estimated realignement parameters (three rotations and three translations)
- the coregistered source image (structural) and the mean functional image
Why should we perform segmentation?
- Segmentation improves mapping to standard space
Is segmentaion performed independently in SPM?
No. In SPM segmentaion and Normalization are usually performed together in the model “unified segmentation”
How is segmentaion performed in SPM and what tissue types are segemented in SPM?
- SPM uses spatial priors (the so call tissue probability map) and the intensity of voxels in order to assess what tissue type a voxel belongs to.
- The tissue typed differentiated are:
- Skull (bone)
- CSF
- White matter
- Grey matter
- Meninges (dura mater, arachnoid mater, subarachnoid membrane, pia mater)
- (air)
What is the output of segmentaion in SPM?
- A number of Tissue probability maps (for different tissue types)
- Normalization parameters used for warping volume to standard space
Why should we perform Normalization?
- There is a great inter-subject variability when it comes to anatomy in MRI scans:
- If we perform experiments with multiple subjects, we want to increase sensitivity of our statistical analysis by having each voxel represent the same anatomical strucure for each participant -> standard anatomical template
- In order to compare studies, we want to map the brain into a common standard coordinate system -> standard template (e.g. MNI or Tailarach)
What two steps of Normalization are performed in SPM?
- Linear Normalization: Adjusting for global differences using affine transformations with 12 parameters (3 rotation x 3 translation 3 zoom x 3 sheer)
- Non-Linear Normalization: Adjusting for local differences using deformation fields based on smooth basis functions
How is linear normaliztaion achieved and what is its general goal?
There is a affien transformation using 12 paramers: 3 Translation -; 3 Rotaion-; 3 Zoom-; and 3 Sheer-Parameters
The goal is to have a rough agreement between source and reference images.
How is Non-Linear normalization achieved?
Non-Linear Normalization is based on Deformation fields
These deformation fields are created like this:
A number or marginal basis functions, describe different spatial frequencies in different direction. each marginal basis function has a bias - a weight.
By multiplying each marginal basis function we receive a number of joint basis functions. By adding these upe we receive the deformation field.
This deformation field describes how a certain - local region - must be deformed in order to optimally match the anatomical template.
Defining the deformation fields is acheived by estimating the beta values (biases) of the joint basis functions, that exist when the image is maximaly similar to the template.
The discrete cosine transform (DCT) represents an image as a sum of sinusoids of varying magnitudes and frequencie
Describe the relationship of segmentation and normalization and exlain a possible solution.
segmentation and normalization have a circular relationship:
* knowing what tissue type a voxel has helps with normalization
* knowing the location of the voxel helps with defining its tissue
-> the solution is to create a generative model that accounts for both:
how voxel intensity arises from a mixture of tissue type distributions and how tissue type have to be spatially deformed to match the template.
Describe two typical preprocessing pipelines and how they can be integrated for optimal results.
Pipeline 1: Take T1 structural image and normalize it to an anatomical template. Take normalization parameters and apply them to the whole series of functional scans
-> issue: anatomical data might not be perfectly alignes to functional data!
Pipeline 2: Take mean functional image from realignement and normalize it to an EPI template. Use estimated parameters and apply them to the whole series of functional scans.
-> issue: anatomical data is disregarded
Integrated pipeline: take mean functional EPI data and coregister anatomical data to it. Normalize anatomical data to template. Use normalization parameters from this step to correct the functional timeserie.
Why should we perform Smoothing?
- Even afer all prior steps there is no perfect anatomical alignement between participants -> smoothign helps with alignement
- Smoothing increases SNR
- Smoothing is required through Random Field Theory (RFT)
How is data smoothed?
We take a 3D Gaussian kernel and apply it to each voxel seperately (= convolution of of each voxel with a 3D Gaussian kernel)
Each value (of the voxels, that the kernel reaches) is multiplied by the kernel value. After this is done the products are summed up. This gives a smoothed image.
How is a smoothing kernel defined and how is it calculated?
The Kernel is defined by its width at half maximum (FWHM). This is calculated based on its direct relation to the standard deviation of a multivariate normal distribution.
If the distribution X is normally, randomly distributed, the FWHM is defined as: 2* SQR(2ln 2std).
What do we receive after smoothing?
- After smoothing each voxel becomes a the result of a weighted ROI
- This means that voxels that are further apart, have less influence on each other.
- The larger the FWHM is, the stronger the influence of further aways voxels is.
Why should we not smooth our data?
- Smoothing effectively enlarges the brain (as it smears out data to the empty space at the edges of the brain)
- Smoothing effectively reduces spatial resolution
- Smoothing can obscure fine grained information
Name three thumb rule for Smoothing
- The FWHM should be at least twice the normalized voxel size in order to satisfy RFT assumptions
- In univariate analysis smoothing should be performed before statistical analysis
- In multivariate analysis smoothing should be performed after statistical analysis
Name to further processing steps (one for spatial and one for temporal preocessing)
- Spatial: Masking
-> Use a thresholded TPM and multiply it with funcitional scans to receive a masked image (i.e. an image where analysis is performed only on certain voxels) - Temporal: Filtering
-> filter low frequencies in order to correct for drifts or trends in data (usually done during statistical model estimation)
