Computational Neuroanatomy John Ashburner [e-mail]

1

Concepts of SPM data analysis Marieke Schölvinck
2
EPI
structural
3
Basic idea
Make sure all images look the same
Make model of what you think brain activity in
your experiment should look like
And fit this model to the data see whether this
fit is statistically significant
within a single subject, and then over the
whole group
4
SPM user interface
spm fmri
Preprocessing
Analysis
Extra functions
5
Preprocessing
6
Preprocessing
(making sure that all images look the same)
1. Realignment align scans to each other 2.
Coregistration align scans to structural
scan 3. Slice timing make up for differences in
acquisition time 4. Normalisation to a
standard brain 5. Smoothing
7
1. Realignment
EPI (functional) images
8
1. Realignment

• Subjects will always move in the scanner
• therefore the same voxel in the first image
  will be in a different place in the last image!
• Correct by estimating movement and reorienting
  images accordingly

• Realignment involves two stages
• 1. Registration - estimate the 6 movement
  parameters that describe the transformation
  between each image and a reference image (usually
  the first scan)
• 2. Reslicing - re-sample each image according to
  the determined transformation parameters

9
2. Coregistration

• Its useful to display functional results (EPI)
  onto high resolution structural image (T1)
• Therefore warp functional images into the shape
  of the structural image.

10
3. Slice timing

• Each slice is typically acquired every 3 mm,
  requiring 32 slices to cover cortex
• Each slice takes about 60ms to acquire
• entailing a typical TR for whole volume of 2-3s
• 2-3s between sampling the BOLD response in the
  first slice and the last slice

11
4. Normalisation
MNI template brain
12
4. Normalisation

• Inter-subject averaging
• extrapolate findings to the population as a whole
• increase statistical power
• Reporting of activations as co-ordinates within a
  standard stereotactic space
• e.g. Talairach Tournoux, MNI

• You do it by a 12 parameter transformation
• 3 translations
• 3 rotations
• 3 zooms
• 3 shears

13
5. Smoothing

• Potentially increase signal to noise
• Use a kernel defined in terms of FWHM (full
  width at half maximum) - usually 6-8mm

14
Wrapping up preprocessing
1. Realignment align scans to each other 2.
Coregistration align scans to structural
scan 3. Slice timing make up for differences
in acquisition time 4. Normalisation to a
standard brain 5. Smoothing
MNI template brain
15
Analysis
16
Analysis
(fitting model to data and seeing whether this
fit is statistically significant)

• SOME TERMS
• SPM is a massively univariate approach - meaning
  that the timecourse for every voxel is analysed
  separately
• The experiment is specified in a model called a
  design matrix. This model is fit to each voxel to
  see how well it agrees with the data
• Hypotheses (contrasts) are tested to make
  statistical statements (p-values), using the
  General Linear Model

17
Model
voxel timeseries
model with 2 conditions

• How well does the model fit the data?

18
Design Matrix several models at once
1 gt 2
2 gt 1
other parameters (motion)
19
Contrasts
1 -1
-1 1

• T contrast are the values for condition 1 in
  this voxel significantly higher than the values
  during condition 2?
• F contrast are the values for both conditions
  significantly different from baseline?

20
Test every model for every voxel
1 -1
give me all the voxels for which this model
(condition 1 makes the voxel more active than
condition 2) fits the data significantly
21
A word on multiple comparisons
Because youre looking at thousands of voxels,
some will give a positive result just by chance.
You need to correct for this multiple
comparison problem using one of several options
in SPM FWE (family-wise error), FDR (false
discovery rate), or uncorrected (and say which
one you used!)
22
(No Transcript)
23
The End