Session 4 - Preprocessing

Question 1

What is preprocessing (MRI)?

Answer 1

Preparation of data for anlysis
- apply statistical methods but not for analysis but for filtering data
- involves different steps:
1. Motion correction
2. Normalisation
3. Smoothing
4. Slice time correction
5. Unwarping
6. Reslicing

Question 2

Motion correction

Answer 2

• scanner moves due to strong magnetic field
• people move (head, breathing, swallowing, spine relaxes, sagging of head pads, position changes, etc.)
  –> re-align all images to create one final image: to increase sensitivity and specificity
• modern scanners can perform online motion correction

Question 3

reduced sensitivity

Answer 3

= major part of voxel-timeseries variance due to subject motion (~60%)
–> increases error variance: false negatives

Question 4

reduces specificity

Answer 4

• motion can be correlated with experimental paradigm
• motion ‘mimics’ experimental variance
  –> false positive activations

Question 5

Description of subject motion

Answer 5

• rigid body transformation:
  –> topology is preserved
• 6 transformation parameters
  –> 3 translations (in x,y,z)
  –> 3 rotations (on x,y,z)
• jumps occur after break between runs: new positioning
• spine relaxes: people become longer: up to 1 voxel size

aim = minimise the sum of squared differences between volume to register and reference volume –> iteratively
z_i = x-yi (z_i - difference, x - selected reference volume, y_i - other volumes to register)

• typically this step creates a file for each volume defining how it has to be transformed to best match reference volume

Question 6

What is reslicing?

Answer 6

= resample data to make it fit in the same space –> required interpolation
- different possibilities to interpolate

Question 7

Interpolation methods

Answer 7

1. Nearest neighbour:
   - take two real values and add one close to it
   - could cause a jump
2. Linear
   - take mean –> trilinear possible for 3D
3. B-spline: typically used, take different points and fit function to it
4. SINC: often used, can make images more sharp/blurry

Question 8

Residual error

Answer 8

• after motion correction and reslicing
• get number (overall) and for each voxel: motion in image

Question 9

Sources of residual error

Answer 9

Rigid body distortion
- slices of a volume are acquired successively
- motion occurs during the acquisition of a single volume not between volumes

• ghosts and other artefacts do not move synchronously with the rigid body
• motion can shift the slice position (eg slice 5 is not at previous location of slice 6)
• induces change of TR
• head motion distorts homogeneity of main magnetic field (B0) –> signal drop out

Solution
- include motion parameters as nuisance regressors in statistical model

Question 10

B0 unwarping/inhomogeneity of B0

Answer 10

• geometrical distortion along phase encoding direction
• signal drop (mainly front, face): darker, less signal
• measure strength: can differ
• create field map to calculate with a shift map how much needs to be adjusted –> needs to be done for each location

Question 11

slice time correction

Answer 11

• for long time a whole brain scan wasn’t possible –> causes a problem
• slices are acquired sequentially: take 2sec then the first slice has been recorded 2sec before the last one
  –> typical time delay 60-100ms
• take one slice time and correct all other slice times to the same time as the reference
• interpolation needed: realing points to a certain time
• why: important for event-related designs –> data is not collected at the same time but all is in reaction to the same event
• during statistical analysis: slices are treated as being acquisited simultaneously

Question 12

Spatial normalisation

Answer 12

Reason:
- different head sizes, different brains –> get template to normalise brain to standardised space
–> one way to perform statistics across subjects (also see ROI analyses)

• different standard brains (blurry): start with image –> sequeeze into standard brain –> get normalised brain
• there are different atlases and reasons to use them

Allows for report in standardised coordinates across studies –> databases, meta-analyses

Question 13

Parameters in spatial normalisation

Answer 13

12 parameters:
- 3 translations
- 3 rotations
- 3 zooms
- 3 sheers

Question 14

optimisation of nonlinear basis function

Answer 14

–> possible important for spatial normalisation

• cosine basis function: multiply origianl image
  –> shift right/left
• minimise squared difference between reference volume and volume to register weighted by set of basis-function
  –> ie find the optimal beta-weights for basis function

Question 15

Spatial smoothing

Answer 15

Aim:
- reduce noise
- improve signal to noise ratio of the data

Why:
- high frequency noise is independent for each voxel
- signal (BOLD) spreads across voxels

• functional homology across subjects increases –> very important in group studies

Matched filter theory:
- if the filter is of the same form and size as the signal, noise is filtered maximally

Random-Field theory:
- FWHM minimally three times the voxel size

How:
- blurr the image, typcial smoothing between 2mm
- often Gaussian kernel, take width between two points
- multiply intensity value with kernel: convolusion
–> take pixel that needs to be adjusted plus a little bit of the neighbouring pixel depending on distance

–> after smoothing, the data fit better to the assumptions of the Random field theory

Question 16

Reasons not to smooth

Answer 16

• brain gets enlarged (voxel beyond boundary)
• effective resolution is reduced
• fine-grained information can get lost –> multivariate pattern classification