intro to imaging analysis and data L3 Flashcards
what are the main neuroimaging data formats used in analysis?
- DICOM
- Analyse 7.5
- NiFTI
what are the typical preprocessing steps for neuroimaging data?
- correcting for mechanical artefacts
- poor signal-to-noise ratio
- slice timing corrections
-realignment
-normalisation
-spatial smoothing
what are the 6 degree of freedom (DoF) in rigid body registration?
3 translations (X,Y,Z) and 3 rotations
describe the characteristics of analyse 7.5 and NifTI formats
- analyse 7.5 only has info about image itself
- NifTI is an improved version of analyse 7.5 that adds MR parameters to the header like Tr and Te and voxel dimension
what is slice-timing correction and why is it necessary?
- it accounts foe the fact that different brain slices are sampled at different time points in the BOLD repsonse
- this is because we sample the brain in lots of 2d slices rather than getting a 3D image simultaneously
- without correction the time series data may be mixture of signals from different parts of the brain
what do MRI scanners output the data as? and what is it then converted to?
- output in the format of DICOM
which ahs both image data and metadata( patient details) - but since its not suitable for neuroimaging it is converted ti NifTI or Analyse 7.5
NifTI vs Analyse 7.5? whats the difference
nifti:
- has parameters that analyse 7.5 doesnt have like:
- Tr, Te, voxel size
- can handle 4D
- for more specific parameters
Analyse 7.5:
- has basic information about the image like dimension/orientation
- limited to only to 3D data and thus cant handle time0series data like fMRI
define the preprocessing tricks simply:
- realignment = corrects for patient movements
- normalisation = aligns brain images to a standard template
- co-registration = aligns functional (fMRI) and structural(anatomical) data from the same subject
- smoothing = applies filters to reduce noise and enhance signal
what is the difference between fMRI and MRI ?
fMRI ( function):
- measures the brain activity by detecting changes in blood 02 levels, allowing for the study of functional processes in the brain
MRI (structure):
- detail of anatomical images of brain and body structure
explain the concept of interpolation in slice-timing correction
- interpolation replaces a value based on the two adjacent time points and the reference slice
- ## sync interpolation is more commonly used than linear interpolation to avoid signal distortion
describe the process of realignment in fMRI data preprocessing? and what gets involved in terms of RBR( rigid body resgistration)?
- realignment ensures that each voxel samples the signal from the same part of the brain over time by correcting for head movement
- moves each volume towards the first or average volume using rigid body registration
- rigid body registration involves translation and rotation along the X,Y,Z axes resulting in 5 DoF
what are the challenges associated with head motion in fMRI?
- movement greater than 1 voxel is a problem
- realignment minimises the mean square error between adjacent brain volumes to correct for motion
explain the two steps involved in normalisation
- Affine (liear) transformation= involves 12 parameters:
- translation
- rotation
- stretches
- shears
and makes the brain approx the same shape and size of the template ( Talairach Atlas) - Non linear warping/deformation = this optional step uses DCT ( discrete cosine transform) ro refine the alignement
describe the role of DCTs and regularisation in non-linear warping
- show local deformation of the brain images
- they are flexible so we need to make sure that the adjustments don’t distort the actual shape of the brain
- regularisation minimises the mean square error between the template and source image ,AKA keeps the changes realistic
what is co-registration and why is it used?
- EPI images can lose signal esp in the frontal and temporal regions of brain.
- to improve normalisation accuracy, co-registration is used
- co reg aligns functional (EPI ) images and structural images
- since functional and structural images have different contrast, methods like MSE ( mean square error) are not good in measuring the alignment so instead mutual information or shared entropy is used to maximise the shared info between 2 types of images
what are the benefits of spatial smoothing in fMRI?
- Involves convolving the data with a smoothing kernel= Gaussian kernel
- it increases the SNR, masks anatomical variability, normalises errors/noise
what is the matched filter theorem and its relevance to smoothing?
- states that the kernel size should be equal to the size of the effect of interest
explain the aim of the fMRI data analysis
- the goal is to separate each voxel into two parts:
1. predictors = part explained by the experimental conditions
2. residual noise
Describe the mass-univariate approach in fMRI analysis? with the phases
- it tests each voxel independently meaning the analysis is done on a voxel-by-voxel basis
- has 2 main phases: single subject responses and group level responses
- the general linear model(GLM) is used as it explains the changes in the voxels signal over time
3 conditions for LTI
- linear scaling= BOLD response should scale directly proportional to input stimulus
- shift invarince= timing the response should depend on when stimulus happens
- additivity = réponse to lots stimuli should be sum of the response of each stimuli
whats is the role of the hemodynamic function (hrs) in fMRI analysis?
- HRF is a function that shows the delayed and dispersed nature of the BOLD response to a stimulus
- in fMRI the times when stimuli are shown are combined with HRF using convolution to predict what the signal will look like
- this convolution assumes that the BOLD response is a linear time-invariant system.
what are some considerations regarding HRF model?
- the shape of the BOLD response can vary across everyone and brain regions due to factors like vascular differences and ageing
- HRF is more for block designs
- variations can be seen using derivates of the canonical HRF such as temporal derivates and spatial derivates.
expain the concept of two-level analysis in fMRI
- 2 level analysis involves seeing data at both the individual subject level ( first level) and the group level ( 2nd level)
- at first level, each subject’s responses are estimated
- at 2nd level, group level responses are estimated by combining data from lots of subjects
how is low frequency drift addressed in fMRI data analysis?
- the BOLD signal shows low freq which can be removed with a HPF ( high pass filter)
- HPF removes low frequency components of the signal while keeping higher frequency
- DCT ( discrete cosine transforms) are used to implement the HPF
what is the FIR (finite impulse response) model?
- is good for capturing complex or non-standard BOLD responses , thus flexible
- but with flexibility comes a risk of overfitting to noise , meaning that model might fit random areas in data rather than the true signal
formula for regression
y= mx+c+error
describe the nuisance regressors in fMRI analysis
- are variables that can affect the fMRI signal
- variables like : realignment parameters ( head motion), breathing, large motion spikes
- including nuisance regressors in the model help remove sources of noise that are not needed in task
explain the importance of considering autocorrelation in fMRI data
- shows where signals at nearby time points are related
- not doing this can mess with the results and be biased
- so ReML ( restricted maximum likelihood) is used to correct for autocorrelation in data
- ignoring these autocorrelations can led to false positives and overestimating the actual relationship of signal.
differentiate between fixed effects, random effects and mixed effects models in fMRI group analysis
- fixed= individual differences within each persons brain activity
- random= differences in people
- mixed= to make interference the data is combined
why is lots of comparison correction important in fMRI analysis?
- fMRI data has lots of tests one for each voxel
- without correction the false postives results ( Type 1 error) increase
contrast FWER( false family wise error rate) and FDR(false discovery rate) correction methods and Bonferroni are all Multiple comparison. corrections define them
- FWER= controls the probability of any false voxel or cluster in the image
- uses method like random field theory
- FDR= controls the proportion of false positive results among the findings
- Bonferroni p>0.05/n , where n is number of independent tests
- it changes the threshold to fix the error
explain the concept of resolution elements in multiple comparison correction
-shows the number of independent tests that are taken
- accounts for spatial correction between voxels due to smoothing
- the number of resolution elements can be calculated as R= v/ smoothness , where v is number of tests
define spatial correlation
refers to the relationship between values at different locations in space, where nearby locations tend to show similar values or patterns
describe the difference between voxel-level and cluster level interference in multiple comparison corrections
- at voxel level, we keep only the strong voxels that are above the threshold as defined by Gaussian random field theory (p<0.05) to say they are active
- at cluster level:
- groups nearby voxels into clusters
- default threshold of p<0.001
steps for cluster level
1 step:
- set a threshold of p<0.001 to form cluster
- note: if p is low ( less than 0.05) means result is real but if high mean result is random and is thereofre consistent witht he null hypotheses ( means no effect or difference))
2nd step:
- only clusters that are large enough are kept
- this menthod is more sensitive ( better) but lose some detail on excact brain location ( spatial specificty is lost)
what is the region of interest (ROI) analysis >
- focuses on specific brain region
- when using a prior hypothesis and predefines ROI multiple comparison corrections across the whole brain isn’t needed
- ROI analysis can be done by extracting the signal from the ROI and doing statistical analysis or by using a small volume correction which controls peak level correction to the ROI
