Basics of fMRI Preprocessing
Douglas N. Greve
greve@nmr.mgh.harvard.edu
fMRI Analysis Overview
Subject 1
Raw Data
Preprocessing
MC, STC, B0
Smoothing
Normalization
First Level
GLM Analysis
X
Subject 2
Raw Data
Preprocessing
MC, STC, B0
Smoothing
Normalization
C
First Level
GLM Analysis
X
C
Higher Level GLM
Subject 3
Raw Data
Preprocessing
MC, STC, B0
Smoothing
Normalization
First Level
GLM Analysis
X
X
C
C
Subject 4
Preprocessing
MC, STC, B0
Smoothing
Normalization
Raw Data
First Level
GLM Analysis
X
C
2
Preprocessing
• Assures that assumptions of the analysis are met
• Time course comes from a single location
• Uniformly spaced in time
• Spatial “smoothness”
• Analysis – separating signal from noise
3
Input 4D Volume
64x64x35
85x1
4
Preprocessing
•
1.
2.
3.
4.
5.
•
Start with a 4D data set
Motion Correction
Slice-Timing Correction
B0 Distortion Correction
Spatial Normalization
Spatial Smoothing
End with a 4D data set
•
•
•
Can be done in other orders
Not everything is always done
Note absence of temporal filtering
5
Motion
• Analysis assumes that time course
represents a value from a single location
• Subjects move
• Shifts can cause noise, uncertainty
• Edge of the brain and tissue boundaries
6
Motion
• Sometimes effect on signal is positive,
sometimes negative
7
One Effect of Uncorrected Motion
on fMRI Analysis
• Ring-around-the-brain
From Huettel, Song, McCarthy
8
Motion Correction
Template
Target
Reference
Input
Time Point
Difference
(Error)
• Adjust translation and rotation of input
time point to reduce absolute difference.
• Requires spatial interpolation
9
Motion Correction
Raw
Corrected
•Motion correction reduces motion
•Not perfect
10
Motion Correction
• Motion correction parameters
• Six for each time point
• Sometimes used as nuisance regressors
• How much motion is too much?
11
Slice Timing
Ascending
Interleaved
• Volume not acquired all at one time
• Acquired slice-by-slice
• Each slice has a different delay
12
Effect of Slice Delay on Time Course
• Volume = 30 slices
• TR = 2 sec
• Time for each slice = 2/30 = 66.7 ms
2s
0s
Slice
Slice Timing Correction
TR1
TR2
TR3
X
• Temporal interpolation of adjacent time points
• Usually sinc interpolation
• Each slice gets a different interpolation
• Some slices might not have any interpolation
• Can also be done in the GLM
• You must know the slice order!
14
Motion, Slice Timing, Spin History
• Every other slice is bright
• Interleaved slice acquisition
• Happens with sequential, harder to see
15
Slice
Motion, Slice Timing, Spin History
TR1
TR2
TR3
• In TR3 subject moved head up
• Red is now the first slice
• Pink is now not in the FoV
• Red has not seen an RF pulse before (TR=infinite) and
so will be very bright
• Other slices have to wait longer to see an RF pulse, ie,
the TR increases slightly, causes brightening.
• Direction of intensity change depends on a lot of things
16
B0 Distortion
Stretch
Dropout
• Metric (stretching or compressing)
• Intensity Dropout
• A result of a long readout needed to get an entire slice
in a single shot.
• Caused by B0 Inhomogeneity
17
B0 Map
Magnitude
Phase
Voxel Shift Map
Echo 1
TE1
Echo 2
TE2
Voxel Shift Map
• Units are voxels (3.5mm)
• Shift is in-plane
• Blue = PA, Red AP
• Regions affected near air/tissue boundaries
• sinuses
19
B0 Distortion Correction
• Can only fix metric distortion
• Dropout is lost forever
20
B0 Distortion Correction
• Can only fix metric distortion
• Dropout is lost forever
• Interpolation
• Need:
• “Echo spacing” – readout time
• Phase encode direction
• More important for surface than for volume
• Important when combining from different scanners
21
Spatial Normalization
• Transform volume into another volume
• Re-slicing, re-gridding
• New volume is an “atlas” space
• Align brains of different subjects so that a given
voxel represents the “same” location.
• Similar to motion correction
• Preparation for comparing across subjects
• Volume-based
• Surface-based
• Combined Volume-surface-based (CVS)
22
Spatial Normalization
Native Space
MNI305 Space
Subject 1
Subject 1
MNI305
MNI152
Subject 2
Subject 2
Affine (12 DOF) Registration
23
Problems with Affine (12 DOF)
Registration
Subject 1
Subject 2 aligned with Subject 1
(Subject 1’s Surface)
24
Surface Registration
Subject 1
Subject 2 (Before)
Subject 2 (After)
25
Surface Registration
Subject 1
Subject 2 (Before)
Subject 2 (After)
• Shift, Rotate, Stretch
• High dimensional (~500k)
• Preserve metric properties
• Take variance into account
• Common space for group
analysis (like Talairach)
26
Spatial Smoothing
• Replace voxel value with a weighted average of nearby
voxels (spatial convolution)
• Weighting is usually Gaussian
• 3D (volume)
• 2D (surface)
• Do after all interpolation, before computing a standard
deviation
• Similarity to interpolation
• Improve SNR
• Improve Intersubject registration
• Can have a dramatic effect on your results
27
Spatial Smoothing
• Spatially convolve image with Gaussian kernel.
• Kernel sums to 1
• Full-Width/Half-max: FWHM = s/sqrt(log(256))
s = standard deviation of the Gaussian
0 FWHM
5 FWHM
10 FWHM
Full-Width/Half-max
Full Max
2mm FWHM
Half Max
5mm FWHM
10mm FWHM
Spatial Smoothing
29
Effect of Smoothing on Activation
• Working memory paradigm
• FWHM: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20
Volume- vs Surface-based Smoothing
14mm FWHM
• 5 mm apart in 3D
• 25 mm apart on surface
• Averaging with other
tissue types (WM, CSF)
• Averaging with other
functional areas
31
Choosing the FWHM
• No hard and fast rules
• Matched filter theorem – set equal to activation size
• How big is that?
• Changes with brain location
• Changes with contrast
• May change with subject, population, etc
• Two voxels – to meet the assumptions of Gaussian
Random Fields (GRF) for correction of multiple
comparisons
32
Temporal filtering
• Replace value at a given time point with a weighted
average of its neighbors
• Should/Is NOT be done as a preprocessing step –
contrary to what many people think.
• Belongs in analysis.
33
fMRI Analysis Overview
Subject 1
Raw Data
Preprocessing
MC, STC, B0
Smoothing
Normalization
First Level
GLM Analysis
X
Subject 2
Raw Data
Preprocessing
MC, STC, B0
Smoothing
Normalization
C
First Level
GLM Analysis
X
C
Higher Level GLM
Subject 3
Raw Data
Preprocessing
MC, STC, B0
Smoothing
Normalization
First Level
GLM Analysis
X
X
C
C
Subject 4
Preprocessing
MC, STC, B0
Smoothing
Normalization
Raw Data
First Level
GLM Analysis
X
C
34
Preprocessing
•
1.
2.
3.
4.
5.
•
Start with a 4D data set
Motion Correction - Interpolation
Slice-Timing Correction
B0 Distortion Correction - Interpolation
Spatial Normalization - Interpolation
Spatial Smoothing – Interpolation-like
End with a 4D data set
•
•
Can be done in other orders
Note absence of temporal filtering
35
36