Preprocessing

Francis Loayza
Functional Neuroimaging Laboratory
Center for Applied Medical Research
University of Navarra
 So
Reconstructed me
fMRI Data S tep Pre
Slice Timing s pro
AdjustmentRealignment ce
ssi
Motion ng
Correction
Coregistration

fMRI Normalization Display

Study Warping Smoothing
Rendering of
Spatial & Temporal Results on
Meta-Data Filtering Anatomy
Data Modeling
Experimental Statistical
Statistical
Design Analysis
Matrix Maps
Engine

Table of
Local Max
Tools for preprocessing
 SPM (fMRI, VBM)
 FSL (fMRI, VBM)
 Brain Voyager (fMRI, VBM)
 VoxBo (fMRI, VBM)
 AFNI (fMRI, VBM)
 ANTs (for Deep Learning)
 SITK libraries (for Deep Learning)
Basic terminology for different scales
a period of time during which the subject is
session in the scanner, motionless (~1-2 hours)

a single run of an MR measurement

run sequence (< 10 minutes) [AKA:” series”]

volume
slice
voxel
Functional MRI (fMRI)
 Uses echo planar imaging (EPI) for fast
acquisition of T2*-weighted images.
 Spatial resolution:
3 mm (standard 1.5 T scanner)
 < 200 μm (high-field systems)
 Sampling speed:
1 slice: 50-100 ms
 Requires spatial pre-processing and statistical
analysis.
Quality control
• It is essential that you perform QC checks at
multiple points in your data analysis stream
 You MUST know how to look at raw data!
 You can’t trust your analysis package to do everything
for you.
 Things to look for
 Data acquisition artifacts
 Subject-related artifacts
 Analysis-related artifacts
Quality control
 MRI data artifact:
 Things to look for:
 Spikes
 Ghosting
 Geometric distortions
 Signal dropout
 Tools:
 ForSPM: art_repair
(matlab toolbox)
Quality control
 Spikes Ghosting
Quality control
Geometric distortions Susceptibility dropout

M Cohen.

Ojemann et al.
Quality control
 Movement artifact
The traditional preprocessing
1. Slice Timing
2. Realignment
3. Coregistration
4. Normalisation
5. Smoothing
Slice Timing
Some researches don't
recommend to perform this steep

exposed last

exposed first
Slice Timing
exposed last

exposed first
Slice Timing
 Correction shifts each voxel's time series so
that all voxels in a given volume "appear" to
have been captured at exactly the same time
TR RAW

TA
0 3 6

time in seconds
Slice Timing
 Slices of an EPI image acquired separately
 Therefore, information from different slices come
from varying points in time after task events
 Different orders of slice acquisition:
 ascending vs. descending
 sequential vs. interleaved
Slice Timing
Only needed if:
 Temporal dynamics of evoked responses are
important and if
 TR is sufficiently small to permit interpolation
( <3 seconds )

 BioPhysical latency is on the order of seconds

 Usually unnecessary if latency differences are
modelled in SPM analysis “proper” using
temporal derivatives.
SPM Slice timing
Old Format “3D”
afilename.hdr
afilename.img New Format “3D”

afilename.nii
afilename.mat
FSL Slice timing
Included in FEAT module
Realignment (motion correction)
 hemodynamic response is small compared to signal from
movement
 increase sensitivity of T-Test (movement contributes to
variance)
 When to do it
 Must be before Normalization
 Can be either before or after slice time correction
(disadvantages to both options)
 For interleaved acquisitions it’s recommended to slice time
correct first
 For sequential acquisitions it’s recommended to realign first
 May introduce errors due to interpolation and interaction
with motion
 Applying rigid-body transformations: 3 translation and 3
rotation
Métodos: análisis de imagen  preprocesado
Activation

 Imágenes RMf:

 Realineación

 Detección de artefactos
Realignment (3D motion correction)
 Subjects move, like it or not
 EPI image also moves from B0 shift
 Along phase encode
 (L->R or A->P)
Shift X, Y, Z (in mm)
 Rotate X, Y, Z (pitch, roll, yaw)

 If voxel at high-contrast
edge, then time series
will show changes
just from motion
Realignment
Movements could be due to stimulation
Realignment
 Gadgets for prevent artifacts movements

A B

C D
Realignment
In extreme cases, up to 90% of the variance in
fMRI time-series can be accounted for by effects
of movement AFTER realignment:
 subject movement between slice acquisition
 interpolation artefacts
 nonlinear distortion due to magnetic field
inhomogeneities (EPI distortion)
Realignment and Unwarp
 EPI images are distorted relative to the structural scans
 Bigger magnet = more distortion
 Different tissues have different magnetic susceptibilities
 Magnetic field warps at tissue boundaries
Realignment and Unwarp
 Either anatomical scan or functional data is used to estimate
warping parameters, using one of the following models:
 12 parameter affine transformation
 Low frequency basis spatial functions
 Vector field specifying the mapping for each voxel SPM
FSL (FLIRT)
 in function of degrees of freedom
 Realign: 6 DF
Unwarp: 6 DF
Total 12 DF
Realignment
 SPM adds a prefix r to each
file
 rafilename.nii
Coregistration
 Align different modalities (eg PET & MRI)
 Align functional (EPI) with structural (T1)
and different runs between each other
 Optimize parameters describing rigid body
transformation to match functional with
structural
Coregistration
 Advantages
 Aids in normalization
 Allows display of activation on anatomical images
 Allows comparison across modalities
 Necessary if no coplanar anatomical images

 Disadvantages
 May severely distort functional data
 May reduce correspondence between functional and
anatomical images
Normalization
Normalization
 Advantages
 Allows generalization of results to larger population
 Improves comparison with other studies
 Provides coordinate space for reporting results
 Enables averaging across subjects

 Disadvantages
 Reduces spatial resolution
 May reduce activation strength by subject averaging
 Time consuming, potentially problematic
 Doing bad normalization is much worse than not normalizing
Normalization
 Template
 152 MNI; Voxel size: 2x2x2
 SPM: Prefix w warfilename.nii
Normalization Template Normalized Data
fMRI to structural
Matching the functional image
to the structural image
– Overlaying activation on
individual anatomy
– Better spatial image for
normalisation
Two significant differences
between co-registering to
structural scans and motion
correction
– When co-registering to
structural, the images do not
have the same signal intensity
in the same areas; they
cannot be subtracted
– They may not be the same
shape
Problem: Images are different
Differences in signal intensity between the
images

Normalise to appropriate template (EPI to EPI; T1 to T1), then segment

Normalisation: Methods
Methods of registering images:
– Label-based approaches: Label homologous features in source
and reference images (points, lines, surfaces) and then warp
(spatially transform) the images to align the landmarks (BUT:
often features identified manually [time consuming and
subjective!] and few identifiable landmarks)
– Intensity-based approaches: Identify a spatial transformation
that maximises some voxel-wise similarity measure (usually by
minimising the sum of squared differences between images;
BUT: assumes correspondence in image intensity [i.e., within-
modality consistency], and susceptible to poor starting
estimates)
– Hybrid approaches – combine intensity method with user-
defined features
SPM: Spatial Normalisation
SPM adopts a two-stage procedure to determine
a transformation that minimises the sum of
squared differences between images:
Step 1: Linear transformation (12-parameter affine)
Step 2: Non-linear transformation (warping)
High-dimensionality problem
The affine and warping transformations are
constrained within an empirical Bayesian
framework (i.e., using prior knowledge of the
variability of head shape and size): “maximum a
posteriori” (MAP) estimates of the registration
parameters
Step 1: Affine Transformation
Determines the
optimum 12-parameter
affine transformation to
match the size and Rotation Shear
position of the images
12 parameters = 3
translations and 3
rotations (rigid-body) +
3 shears and 3 zooms Translation Zoom
Step 2: Non-linear Registration
Assumes prior approximate
registration with 12-parameter
affine step
Modelled by linear
combinations of smooth
discrete cosine basis functions
(3D)
Choice of basis functions
depend on distribution of warps
likely to be required
For speed and simplicity, uses
a “small” number of
parameters (~1000)
Matthew Brett
Step 3: Dipheomorphic Registration
2-D visualisation (horizontal
and vertical deformations):

Ashburner; HBF Chap 3

Brain
visualisation:
Source Template

Deformation
field
Warped
image
Smoothing
 Application of
Gaussian kernel
 Usually expressed in
#mm FWHM
 “Full Width – Half
Maximum”
 Typically ~2 times
voxel size
 Smooth 5 to 8 mm.
 Smoothing

Before convolution Convolved with a circle Convolved with a Gaussian

Métodos: análisis de imagen  preprocesado

 Corregistro

 Normalización

 Filtrado espacial
Smoothing FSL:

SPM:
Smoothing
 Advantages
 Increases Signal to Noise Ratio (SNR)
 Matched Filter Theorem: Maximum increase in SNR by filter with
same shape/size as signal
 Reduces number of comparisons
 Allows application of Gaussian Field Theory
 May improve comparisons across subjects
 Signal may be spread widely across cortex, due to intersubject
variability

 Disadvantages
 Reduces spatial resolution
 Challenging to smooth accurately if size/shape of signal is not
known
Smoothing effects

SPM:
Adds prefix s

swarfilename.nii
 Smoothing
Unsmoothed Data

Smoothed Data (kernel width 5 voxels)

Preprocessing using DARTEL
(SPM)
 There are some limitations for normalizing
using the traditional method.
 For this, it have been development some
algorithms to perform normalizations.
The traditional processing
 Affine registration
Diffeomorphic registration

 You can use different packages:

 ANTs The algorithm consist in create and store the
deformation flow field for each subject.
 SPM – DARTEL The same field can be used for both forward
 ART or backward deformations

 IRTK
SPM-Dartel steps
Important..! process all subjects
 Realign
 New segment (split off the differents brain tissues: gray,
white, CSF, bone, etc)

 Dartel tools => Run Dartel: create template based on all

subjects and determine the non linear deformations
 Dartel tools => normalize to MNI space: apply the
non linear deformations to each selected image (EPI and T1).
This script smooth the EPI images.