fMRI part 1

1 of 60

fMRI

fMRI data structure

standard fMRI consist of both structural and functional data

balance spatial with temporal resolution

TR: temporal resolution

1: structural images: distinguish bw tissue types

T2*; functional images: can related changes in signal to an experimental manipulation

MRI images usually acquired in axial slices

sequentially

or

interleaved manner or acquisition

Field of View (FOV): extent of the brain that is inside the image

Slice thickness

Matrix size:

Experiment< subject< sessions< run persession< volume

fMRI part 1

2 of 60
Module 4

Psychological inference
reverse inference

Neuroimaging is the answer. What is the question?

Typical results: statistical map

colored areas indicate reliable non-zero eects

fMRI data are useful only if interpreted in the context of theories

Brain mapping provides forward inference

Forward inference: given an induced physiological state we observe brain activity

probability of observing an activity

Reverse inference: Can we infer a psychological state given the brain activity

fallacious reverse inference

Bayes rule: related forward and inverse inference:

[P(Brain/Psych)= P (brain) * P(psy/Brain)] / P (Psy)

positive predictive value:

P (enjoyment| caudate): the probability I am enjoying something given that the caudate is
activated

P (brain | caudate ): 0.9 < sensitivity

P( caudate | no enjoyment)= 0.4 < specificity
P (enjoyment) : baseline of enjoyment = 0.2

Calculate positive predictive value using Bayes rule:

P( caudate | enjoyment) = 0.2

when people make reverse inferences they assume that there is high positive predictive
value (PPV)

fMRI part 1

For a brain activation to have high PPV it must have high sensitivity and high specificity

When does reverse inference can be done

strategy 1: Leverage Neuroscience

strategy 2: Quantitative reverse inference
Assess activation of a region

Compute its sensitivity, specificity and positive and negative predictive values

Require testing many tasks, contexts

brain imaging is not good for

estimating eect sizes and predictive accuracy

testing assumptions

comparing evidence for dierent theories

3 of 60

fMRI part 1

4 of 60

Basic understanding of fMRI principles

Acquisition and reconstruction

net magnetization M is a vector of two components

longitudinal component: parallel to the magnetic field

transverse component: perpendicular to the magnetic field

in the absence of an external magnetic field there is no net magnetization. the nuclei of
magnetic atoms are randomly oriented

when placed in a strong magnetic field there is a net longitudinal magnetization in the direction
of the field

The nuclei precess about the filed with an angular frequency determined by the Larmor
frequency

A radio Frequency pulse is to used to align the phase and tip over the nuclei.

This causes the longitudinal magnetization to decrease and establishes a transversal

magnetization

The nuclei point to a direction orthogonal of the magnetic field

Longitudinal relaxation: The restoration of net magnetization along the longitudinal direction
as spins return to their parallel state

When the RF is removed the longitudinal magnetization grows back to its original size and the
transverse component decreases

A signal is created measured by a receiver coil

exponential grown described by time constant T1

Transverse relaxation: loss of net magnetization in the transverse plane due to loss of phase
coherence

Exponential decay described by time constant T2

fMRI part 1

5 of 60

dierent tissues (gray, white matter, CFS) have dierent T1 values

gray: 1000 msec

white: 600 msec

CFS: 3000 msec

TR: how often we excite the nuclei

TE: how soon after excitation we begin data collection

by altering the above we control which characteristic we emphasize on

The image represents the density or relaxation time of atoms in a tissue

Proton density: Long TR, short TE

T2 image: long TR, Long TE

T1: short TR short TE

T2* : combined eect of T2 and local inhomogeneities in the magnetic field. used to image
brain function

The scanner can eliminate or emphasize the eects of inhomogeneities.

by emphasizing them we obtain the BOLD signal

How can we use the signal to create an image

scan involve the construction of a 3D image from 2D slices

a brain slice is split into equally size voxels

(x, y): number of protons in each protons

magnetic field gradient: we sequentially control the spatial inhomogeneity of the magnetic field,
meaning that we slightly change the field

The measurements that we make are the Fourier transform of the image we would like to
reconstruct
the measurements are acquired in the frequency domain (k-space)

fMRI part 1

6 of 60

constructing an image of the whole brain is an inverse problem (). we need

enough measurements of kx and ky so that we reconstruct ( x,y)

discrete Fourier transforms

the better the spatial resolution the more k-space measurements I need to make (e.g. for 64 x
64 resolution I need 4096 measurements, meaning that I need to change the values of kx and
ky 4096 times.

methods of acquiring the data is k-space

EPI: echo plane imaging

spiral
the measured k-space data is complex valued

we work with the magnitude of the complex numbers

k-space< inverse fourier transform< image space

K-space
data is acquired in the k-space. By application of the inverse fourier transform we acquire an
image

there is not an one-to-one relationship between image and k-space. So each individual point in
an image space depends on all the points acquired in the k-space

low frequency portions of k-space: higher periodicity of the image waves

high frequency: low periodicity of the image waves

move left or right: change the direction of waves

if we are interested in the relative contribution of the high versus the low frequency waves in
the k-space, we select for one of the two

high frequency parts oscillate quickly and are responsible for detail in the picture. they
represent small structures whose size is on the same order as the voxel size (tissue
boundaries)

fMRI part 1

7 of 60

low frequency parts (center of k-space) give the contrast. They represent parts of the object
that change is a slow manner

Physiology, Signal and Noise

BOLD fMRI: ratio of oxygenated to deoxygenated hemoglobin in the blood

it measures the metabolic demands of active neurons

oxyhemoglobin is diamagnetic

deoxyhemoglobin is paramagnetic. it suppresses the MRI signal.

Active neurons increase blood flow< decrease in deoxy-hemo< increase T2* weighted image

hemodynamic response function (HRF):

As neuronal activation increases so do the metabolic demands for oxygen. So as oxygen is

extracted from the blood the hemoglobin becomes paramagnetic and therefore T2* decreases:

initial dip:i nitial decrease in T2* signal

peak BOLD 4-6 s after activation: over-compensation in blood flow dilutes the concentration
of deoxy-hemo and the BOLD signal increases

post-stimulus undersut: then the BOLD decreases below baseline levels

properties of HRF:

1. small magnitude of signal change

2. response is delayed and slow

3. exact shape of response vary across subjects and regions

how BOLD reflects neuronal activity:

fMRI part 1

8 of 60

it corresponds closely with the local field pontential (LFP)

point spread function of BOLD

but does not always reflect changes in neuronal activity: e.g .blood steal phenomena

Signal to noise ratio (SNR): the strength of a signal divided by an estimate of noise variability

basic measure of eect size

Contrast-to-noise (CNR): The dierence between two signals divided by an estimate of noise
variability

ways of calculation

spatial signals : calculate across one image

spatial SNR: mean intensity within signal area of interest () divided by standard
deviation outside signal area ().

spatial CNR : dierence in intensity between two tissue types divided by the variability
of measurements (1-2)/1,2

temporal signals: calculated at each voxel across time

time series measures: calculated at each voxel across time

temporal SNR (or functional SNR): mean signal across time divided by variability (i.e.
standard deviation) across time. /

Temporal CNR (or signal sensitivity)

dierence in intensity for on vs o states divided by variability

related to sensitivity to task

Scaling:

the absolute scaling of BOLD is arbitrary

fMRI part 1

depends on

1. the field strength

2. acquisition parameters

3. tissue type

implications

non-linearity in BOLD response

refractory eects

saturation: reductions in amplitude of a response as a function of inter-stimulus
intervals

Artifacts and Type of Noise

artifacts appear as

high frequency spikes

image distortions

periodic fluctuations

slow drift across time

how to prevent noise:

1. acquisition

2. Analysis:

Issues to check:

1. coverage

2. RF noise and malformed images

3. Transient gradient artifacts/spikes

4. Ghosting

5. dropout

6. task-correlated movement

Non-signal related components

drift: slow changes in voxel intensity over time. low frequency noise

9 of 60

fMRI part 1

10 of 60

main reason: scanner instabilities

experimental designs should use high frequency. quick changes for on to o

motion correction in the preprocessing stages of analysis

but

spin-history artifacts: due to complex interactions with the magnetic field

aliasing: if TR is too low. periodic signals that occur more rapidly than the sampling rate will
often be aliased back to lower frequencies.

to avoid aliasing we must sample at least twice as fast as the fastest frequency in the signal

Spatial and temporal resolution of BOLD

spatial resolution: below 1 mm

with BOLD fMRI we can collect info on :

functional maps

functional column

large-scale networks

BOLD point-spread function: related to microvasculature bed area aected by local neuronal
activity and venous/arterial flow contributions

what is the eect of resolution in group analysis: 1-15 mm

why?

artifacts

inter-subject normalization

ind. dierences in functional anatomy

diuse modulatory eects

spatial alignement

fMRI part 1

11 of 60

hyper acuity: the patter of activity across voxels may contain more info than any one voxel

can sometimes detect functional topography even if voxels are not small enough


to fit within one functional area

even if neurons are randomly intermixed MVPA may still identify patterns that are dierentially
associated with each

the classification of neuronal activation for events types is still possible because patterns of
activation for events are uncorrelated

hyperalignement: direct inter-subject alignment of brains in a functional space defined by

activity during movie watching

better cross subject matching

Temporal Resolution

limited in fMRI

peak of response: 5-6 s

onset: 2-3 s

Temporal hyperacuity:

if events are averaged can detect dierences in response latency of 100-200msec

sampling of the whole brain in 100-200 sec

Experimental and Design

Block design: Similar events are grouped

E.g. examining brain activation upon presentation of famous vs non famous faces. We present
a block of famous faces and after a while a block of non famous faces

Event-related design: Events are mixed

Rule of thumb: two condition block design with 16-20 sec blocks maximize power

16-20 sec optimal power

fMRI part 1

12 of 60

Goal: induce human subject to do or experience the psychological states you are studying

Considerations:

1. eect of stimulus predictability

Predictability influences psychological state. the predictability of a stimulus influences
how fast someone responds to it

2. eect of time on task

subjects should be engaged on the task as much time as possible

for this sometimes a rapid presentation of stimuli is needed

3. Participant strategy

dierent stimuli configurations aord dierent strategies. e.g stroop test: compatible
(the word and the colour are the same) vs

incompatible trials.

you can not block compatible and incompatible trials

4. Temporal precision of psychological manipulations

what we expect subjects to do should fit with what they can do.

E.g recall happy vs sad memories. people cant switch back and forth from sad faces

5. consequences of unintended psychological activity

e.g .spatial attentional shifting

subject shift attention to fixation cross spontaneously

Kinds of Designs
Design

How many independent variables?

What kinds of measure variables

Trial Structure

How will events be organized in time?

Blocked/event related/ mixed

rapid/single trial partial trial

fMRI part 1

13 of 60

Basic principle: designing a study that is powerful for one purpose vs designing a study for
many purposes

within person variables: manipulated across time .

1. within factors

2. within-levels of factors

within observed variables (e.g. peoples performance)

between-person variables:
between factors

between levels

between observed

fMRI part 1

14 of 60

Examples

Simple subtraction: compare task A vs task B.

look at negative and neutral pictures

condition A: just look at them

condition B: look and reappraisal

Pure insertion principal: processes in complex conditions are simply added on top of
those in simpler, baseline conditions. Cognitive subtraction

Problem: the processes might interact with context

individual dierences designs: by making correlations with individual dierences we can

increase the specificity of inference.

E.g. brain- behavior correlation: correlate brain activity during A stimulus with
the level of the psychological state the stimulus is supposed to induced. E.g. brain activity
during reappraisal of negative stimuli vs level of psychological reappraisal.

Multiple subtraction: triangulate between event A and brain process by subtracting multiple

kinds of events

If the dierent conditions that we are subtracting o have the di

characteristics we can control for some of the drawbacks of pure

insertion.

E.g. compare brain activation at faces and brain activation with objects

and intact faces with scrubbed faces. this is to make sure that the FFA is

activated when someone sees a face and not other things

process overlap/dissociation


double dissociation: task A activates more than task B in one area and B

activated more than A in another area.

separate modifiability: task A activates one area but B not and vice versa

factorial design: view two factors at once and test of dissocciations in the area activated.

e.g task switching experiment:

factor 1: internal switch of attention between objects (2 levels)

factor 2: external switch of attention (2 levels)

2 x 2 factorial space

fMRI part 1

15 of 60

parametric modulation: manipulation variables in a parametric fashion within persons

enhance specificity

e.g. parametric increase in blood flow as the task complexity

increases

fMRI part 1

16 of 60

Preprocessing of fMRI data

Goals: 1. minimize the influence of data acquisition and physiological artifacts

2. check statistical assumptions and transform data to meet the assumptions

3. standardize the locations of brain regions across subjects to achieve validity and

sensitivity in group analysis

Steps:

Visualisation/ Artifact Removal:

e.g. drift

principal component analysis

Slice Time Correction:

we might sample multiple slices of the brain during each individual repletion time


(TR) to construct a brain volume

fMRI part 1

17 of 60

Typically each slice is sampled at slightly dierent time points . e,g, the top of
the brain volume might be sample a second later than the bottom

slice time correction shifts each voxels time series so that they appear to have been sampled
simultaneously

slices 1, ,2 ,3 are
di time points
have di time

sampled at
so they
curves

fMRI part 1

18 of 60

How to correct for this?

temporal interpolation: use info from near time points to estimate the amplitude of the MR
signal at the onset of each TR

use a linear, spline or sinc function

phase shift: slide the time course by applying a phase shift to the furrier transform of the time
course

Head motion:

when analyzing the time series associated with a voxel we assume that it depicts the same
region of the brain at every point

by head motion might compromise this assumption

we correct it with rigid body transformation

the goal is to find the best possible alignment between an input image

and some target image

it involves 6 variable parameters. 3 sets of translation in the x,y direction and 3 sets of rotations
(6 DOF).

fMRI part 1

19 of 60

Rigid Body transformation (6DOF): 6 translations and 6 rotations

Similarity (7DOF): 3 translation, 3 rotation and a single global scaling

Ane (12 DOF): 3 translation, 3 rotation, 3 scaling, 3 shearing

Warping

Transformations where the equations relating the coordinated of the image are non-linear

in motion correction the target image is usually the first one

minimize some function: e.g. sum of square dierences, mutual information

Co-registration

of the functional image with the structural image

simplifies later transformation of the fMRI to a standard coordinate system

the functional and structural images do not have the same signal intensity in the same areas so
they cannot be subtracted

So we have to use at least one ane transformation and a cost function

fMRI part 1

20 of 60

Warping to atlas template

normalization allows one to stretch, squeeze and warp each brain so that it is the same as
some standard brain

the structural MRI image used in the coregistration is wrapped onto a template image

Talairach space
based on a single subject , on a single hemisphere

Montreal Neurological Institute: combination of many subjects but right handed only

fMRI part 1

Spatial Filtering (Smoothing)

increase signal to noise ration and remove artifacts

21 of 60

fMRI part 1

22 of 60

matched filter theorem: a filter that is matched to the signal will give optimum signal to noise

typically the amount of smoothing is chose airport and is idependent to the data

so

adaptive smoothing might be an option

Non-stationary spatial Gaussian Markov random field

smoothing varies across space and time

fMRI part 1

23 of 60

General linear model

two level hierarchical analysis

1. within subject

2. across subjects

do it in stages

i. design specification (model building)

ii. estimation . model combined with real data (1st level)

iii. group analysis (2nd level)

iv. inference about areas activated in the group

The GLM approach treats the data as a linear combination of model functions (predictors) plus
noise (error)

the model functions have known shapes (lines, curve) but unknown amplitudes
simple regression

anova

Multiple regression

mixed eects/hierarchical

time series (e.g .autoregressive)

fMRI part 1

24 of 60

robust penalized regression (LASSO, Ridge)

non normal errors (logistic regression)

in most cases there is a closed form solution while others require iterative solutions

Simple regression

one prediction one outcome

Step 1: specify the model: there is a linear relationship between the predictor and the outcome

Step 2: Estimate: estimate slope and intercept

Step 3: Statistical inference: test slope and get a p value: how likely is that i have observed this
slope under the null hypothesis

Steo 4: Scientific interpretation

GLM: one continuous DV

repeated measures design:

paired t test

generalised least square (iterative): correlated errors. e.g .time series

each measure is not independent

from the previous one

aim at regression model: calculate

we do so by minimizing sum of squared residuals

fMRI part 1

25 of 60

fMRI part 1

26 of 60

apply GLM to fMRI

Typically two-level hierarchical analysis

within subjects

between groups

do it in stages

i. design specification (model building)

ii. estimation . model combined with real data (1st level)

iii. group analysis (2nd level)

iv. inference about areas activated in the group

Mass Univariate Analysis

typical model

we construct a separate model for every voxel

Brain activity in each voxel is the outcome (Y)

Stimulus, task and/ or behavior are the predictors

assume voxels are independent

Consider a single voxel in a single subject

block of famous vs non famous faces

fMRI part 1

27 of 60

matrix of bloc design

intercept: constant: mean level of fMRI signal across time

task regression: capture the eect of famous vs nonfamous faces

1: activation parameter estimate: estimated response amplitude

estimate of how large the dierence in activation is between famous and nonferrous
faces

event related design:

fMRI part 1

28 of 60

famous and non famous faces have their own beta

hemodynamic delay: BOLD responses are delayed and dispersed relative to neural activity

Assume impulse response in model

Common model: Fixed linear

combination of two gamma functions

How to turn assumed neural responses

into a predictor in a GLM model?

Assume a linear time invariant system: the

neural activity acts as the input or impulse
and the HRF (hemodynamic response)
acts as the impulse response function

x (t)= v(h(t))
x(t): fMRI signal over time

v(t): stimulus function

h(t): hemodynamic response

stimulus function (blocks or

events )convolved with
hemodynamic response function

linear: same HRF no matter what

came before

fMRI part 1

29 of 60

Details of building GLM models

GLM for more than two conditions: we specify number of conditions (e.g. A, B, C, D) convolved
each of them with an assumed basis function and end up with a design matrix.

multiple predictors and design contrast

block design;
a single predictor that captures the dierence

event-related design:
one regression for each condition

in this case we can assess

the dierence between famous and non famous

faces

each one separately

their average

dierent linear contrast for each parameter

fMRI part 1
Contrast: linear combination of GLM parameters

T-contrast: single, planned contrast > t test

Specified by a vector of weights (c) so that cT= a scalar value

signed: positive and negative values

dierence contrast [0 (for intercept) 1 (for famous) -1 (for non-famous) ]

Sum (average) contrast (famous vs rest) [ 0

1
0]

Single event contrast[ 0

1
1]: test only the famous faces against intercept

Factorial repeated measures design

Apply contrast

linear combination which equals the

parameter estimate for column A plus
B minus C plus D

Rules for T contrast

1. C can be a matrix

columns are applied independently

first column reflects main eect of factor 1

second column: main eect of

factor 2

third column: cross over

interaction: test if the eect of
factor 1 depends on the level of

30 of 60

fMRI part 1

31 of 60

factor 2

2. the scaling aects the magnitudes but not the inferences I make (t, p values)
Contrast weights must be
all participants. so no
runs

3: contrast weights typically sum to 0

0 is the null hypothesis

Exception: I can test the average of one or more conditions against the implicit

baseline

in order to construct a linear design matrix we have to assume that:

the same for

missing

fMRI part 1

32 of 60

1. neural function is correct

2. HRF is correct

3. Linear time interval system

Linear Basis Set

How to relax these assumptions

often fixed canonical HRF is used to model responses to neuronal activity

but if this is not the case the power of analysis is reduced

The HRF shape depends both on the vasculature and the time course of neuronal activity

so assuming a fixed HRF is not

usually appropriate

first picture: o in timing

second picture: missed shape of

peaks later

bottom left: missed duration

bottom right:

response, the true response

Also HRF varies across brain

regions

Temporal basis functions

model HRF as a linear combination of temporal basis functions, f(t) , such that the overall
response is a sum of three activation parameters * the respective factors

Data fit with three choices of basis set:

fMRI part 1

one predictor per time point

33 of 60

FIR model

purple predictor: captures what is happening in the

first couple seconds following even onset

and so on

so 4 predictors

Canonical HRF: two predictors for the onsets

3-parameter model: regressors, each of those is 3

basis functions convolved with the event onsets

FIR: six regressors for event time

Choosing basis set:

1. accuracy(bias): can the model capture the true response without systematic
variance?

canonical model of HRF: strong assumptions about shape, more bias

FRI: weak assumptions about shape but less bias

2. precision(variance): are the model parameters estimated with littles error variance?

canonical model: few parameters so high precision

FRI: many parameters, low precision (noisy)

fMRI part 1

34 of 60

bias/variance tradeo

SO we need a model which will be

1. simple:

few parameters so high precisions

parameters are interpretable measures of neuroscientific interest

3. Accurate in ways that count:captures the true response amplitude int he
physiological range

Filtering a Nuisance Covariance

model factors associated with known sources of variability but that are not related to
experimental hypothesis need to be included in the GLM

Examples:

signal drift

physiological artifacts (e.g.
respiration)

head motion

Drift
slow changes in voxel-intensity over time
due to scanner instabilities

need to include drift parameters in our model

design matrix has 11 columns:

1 corresponds to task

2. baseline

we want to remove the black line and get to the

red line

Transient gradient artifacts

control for spikes

Physiological Noise

fMRI part 1

35 of 60

Respiration and heart beat give rise to periodic noise often alliased into task frequencies

if the TR is too low there will be problems with alliasing

the sampling rate must be twice as big as the frequency of the curve you seek to model

Head Motion
Basic motion correction in the pre-processing stages takes into account gross dierences bet

How to deal with head motion artifacts?

Nuissance regressors:movement and CFS related artifacts

Scrubbing: drop images with high movement estimates. treat as missing data

physiologically implausible whole brain activation or deactivation

fMRI part 1

how to estimate ?

36 of 60

GLM Estimation
t

the desired line is the line that makes the square of the residuals e1, e2, e3 as small as
possible

The least square criterion:

Q= (Y-X) (-)

taking the derivative with respect to and setting it to 0 gives us the normal equations:

XX(hat)=XY

fMRI part 1

37 of 60

by solving for we get the ordinary least squares (OLS) estimations:

Properties:

the estimated value of hat equals to beta

the variance of beta hat

that means that any other unbiased estimator of beta will have a larger variance that the OLS

so hat is the best linear unbiased estimator (BLUE)

fMRI part 1

38 of 60

R= residual inducing matrix

Noise Models
is not typically the identity matrix because fMRI data typically exhibit significant
autocorrelation caused by physiological noise and drift that has not bee appropriately modeled

typically modeled with AR(p) or ARMA (1,1)

Autoregressive model of order 1

the autocorrelation function (ACF) between adjacent time points depends on how closely
dierent time points are to one another

fMRI part 1

39 of 60

the autocorrelation is equal to 1 if the lag is 0. However if we have a lag of one time point the
autocorrelation is equal to and decays as we move further

In general the form of the comatrix is unknown therefore it

needs to be estimated.

Therefore, estimating V depends on and estimating

depends on V. We need a iterative procedure (assume a value
of V estimate and the re estimate V)

Iterative Procedure

fMRI part 1

40 of 60
Inference, Contrast and t-tests

After fitting the GLM we use the estimated parameters to determine whether there is significant
activation present in the voxel.

our estimate is normally

distributed

Contrasts
test whether linear combinations of parameters are significant

0 for 1

1 for 2

-1 for 3

null hypothesis: 2=3

in order to test the null hypothesis against the non null we use t statistics

fMRI part 1

41 of 60

if we want to make simultaneous test of several contrast at once, c is a contrast matrix

we want to test if 1 and 2 are simultaneously or both equal to 0

c is just an indicator of the drift components

if s are simultaneously equal to 0 the drift will not

contribute

if we want to examine whether the drift components

contribute to the model we test whether there is a
dierence between the full model (including the design matrix for the drifts) from the reduced
model (excluding the design matrix for the drift)

we test that using F Statistics

So for each voxel a hypothesis test is performed

and the statistics corresponding to that test is
used to create a statistical image over all voxels

fMRI part 1

42 of 60

image of T statistics across space:

Use GLM to analyze fMRI data

1. construct a model for each voxel of the brain

1. massive univariate approach: separate models are fixed to each voxel of the brain

2. Perform a statistical test to determine whether

there is task related activation in each voxel. test the null hypothesis. Get a statistical image

3. Choose an appropriate threshold for determining statistical significance.

Get a statistical parametric map

How to determine the threshold?

Problems: many false alarms

many hypothesis tests are performed

simultaneously so many test statistics are
inflated due to noise

choosing a threshold in to determine a balance between sensitivity (true positives) and

specificity (true negative rate)

=0.05 threshold so 5000 false positive voxels

Group-level

fMRI part 1

43 of 60
Analysis I

Muti-level Analysis
fMRI experiments are often repeated for:

several runs in the same session

several sessions on the same subject

several subjects

the first level deals with individuals subjects

the second level deals with groups of subjects

contrast images between subjects are compared

backwards from group results to individual subject result

fMRI part 1

44 of 60

the most basic kind of group result is contrast values between task and control. Each dot is a
score for one subject

the spread between the dots reflects the variance of the data or the noise

Group Analysis II

multi-level models allow dierent variance

components to be introduced

at each level (within or between subjects)

Mixed-Eect Models

Hierarchical Models

Random-Eects (RFX) Models in

neuroimaging

all have multiple sources of variation

fixed eect model only one source of

variation

Sources of variation of the signal strength

a). measurement error

b) response magnitude

Fixed Eect: always the same from experiment

to experiments.

sex

fMRI part 1

45 of 60

drug type

Random eects

subject

word (e.g. choose one positive words and one negative)

if eect is treated as fixed, error terms in model do not include variability across levels

we cannot generalize to unobserved levels

so if we treat subjects as fixed eects we cannot generalize

Contrast design matrix: all 0s and 1s

Group Level Analysis III

fMRI part 1

46 of 60

Design matrix of the second-level

for four subjects

the second level relates the subject

specific parameters contained in
to the population parameters g

fMRI part 1

47 of 60

it assumed that the first level parameters are randomly sampled from a population of possible
regression parameters

Statistical techniques define the loss function that should be minimized in order to find the
parameters of interest

used techniques: minimize loss function, reduced minimized loss function

algorithms define the manner in which the chosen loss function is minimized

fMRI part 1

48 of 60

Multiple Comparison Problem in fMRI

Null Hypothesis: no eect

1-2=0

Test statistic T :measures the compatibility between the null hypothesis and the data

P-value: probability that the test statistic would take a value as or more extreme than actually
observed if Ho is true

P(T>t| Ho)

Fixed threshold: significance level:

Threshold u controls false positive rate at level

= P(T>|o)

The probability that the test statistic lies above that value is equal to value a

Errors:

1. Type I error: the null hypothesis is true but we mistakenly reject it (false positive)

control by significance level . small controls for that

2. Type II error: the null hypothesis is false but we fail to reject it

Power of the test: probability that a hypothesis test will correctly reject a false null hypothesis

Choosing an appropriate threshold

more than one hypotheses are performed so the the risk of making a single Type 1 error is
greater than the value for a single test.

the more tests one performs the greater the likelihood that he will perform a false positive

100.000 voxels with a 0.05 threshold will give us 5.000 false alarms

choosing a threshold is a balance between sensitivity (true positive) and specificity (true
negative)

How to quantify the likelihood of obtaining false positives?

fMRI part 1

49 of 60

1. Family-Wise Error Rate (FWER)

probability of making any false positives

2. False Discovery Rate (FDR)

controls the proportion of

Family Wise Error Rate (FWER)

Ti= value of the test statistic at voxel i

the probability of making one or more Type I errors in a family of tests under the null hypothesis

Bonferroni correction

Random field theory

In the FWER hypothesis, Ho states that there is no activation in any of the m voxels

If we reject a single voxel null hypothesis we reject the whole FWER hypothesis

a false positive at any voxel gives a Family-Wise Error (FWE)

we want to control the probability that any of the test statistics in any of the voxels is above
some value u

if it is above u we will reject the null hypothesis in that voxel

Bonferroni correction
m= total number of voxels

fMRI part 1

50 of 60
if we have 10 hypothesis and we
want to control for the probability
that the sum of them is above 0.05
then we take the threshold for each
to be 0.005 (0.05/10)

100 x 100 voxels from iid N (0.1)

distribution

500 false positives

Bonferroni is conservative and leads to a decrease in the power of the test

in general not optimal for correlation data

Random Field Theory

allow one to incorporate the correlation into the calculation of the appropriate threshold

it is based on approximating the distribution of the maximum statistic over the whole image

fMRI part 1

51 of 60

The FWER is the probability of getting a FWE

So we choose the threshold u such as the

max only exceeds it % of the time

Random Field is a set of random variables

defined at every point in D-dimensional
space

A Gaussian random field

has a Gaussian distribution at every point and every collection of points.

and so it is defined by its mean and covariance

we consider a statistical image to be a lattice representation of a continuous random field

Random field methods are able to:

approximate the upper tail of the maximum distribution which is the part needed to find
the appropriate thresholds and

account for spatial dependence

Euler Characteristic: the property of an image after it has been thresholded

fMRI part 1

52 of 60

no of holes

counts the number of bolbs-

if we threshold at u=0.5 we get 27 bolbs (whites in the picture) and one hole (in one of the
bottom bolbs)

Euler characteristic =2

Euler characteristic=1

How do we know the expected Euler

characteristic?

fMRI part 1

53 of 60

Properties:

as u increases, FWER decreases

as V (no of voxels we are controlling for) increases FWER increases

as smoothness increases, FWER decreases

RFT assumptions

1. The entire image is either multivariate Gaussian or derived from multivariate Gaussian
images

2. The statistical image must be suciently smooth to approximate a continuous
random field

FWHM smoothness of 3-4 voxel sizes is preferable

False Discovery Rate Correction

FWER: controls the probability of any false positives

FDR controls the proportion of false positives among all rejected tests

Suppose we perform test on m voxels

V: false positive

U,V, T and S are unobservable radom

variables

R is an observable random variable

FWER=P(V>or = 1)
V: number of false positives

FDR= E*( V/R)

fMRI part 1

54 of 60

FDR is 0 if R=0

A procedure controlling for FDR ensures that on average the FDR is no bigger than a prespecified rate q which lies between 0 and 1

anything below the black line is active and anything above is not active

Pitfalls and Multiple Comparisons

Cluster-level inference

1, define clusters by arbitrary threshold Uclus

2. Retain clusters larger than -level threshold Ka

fMRI part 1

55 of 60

better sensitivity especially to weak distributing signals

Worse spatial specificity (cluster of voxels instead of individual voxels)

Threshold-free cluster enhancement (TFCE)

Integral M (blue are) must be

above some threshold

combines info about the intensity (how big the t statistics are and how many voxels there are)

What people are doing?

fMRI part 1

56 of 60

Uncorrected thresholds: many studies use them especially when the sample size is small
because correcting for thresholds would decrease the power

Extent Threshold: use an arbitrary threshold and retain clusters of k contiguous active voxels
(e.g. p<0.001, 10 contingent voxels)

but problematic because false-positives are contiguous regions of multiple voxels due to
smoothness

the thresholds used are the default ones in each package (!)

in each blob (orange, blue) there

is at least one are that is
significantly activated but we do
not know what this is.

False discoveries: most

(45-70%) activated maps are not
truly active (Wood et al. 2014)

fMRI part 1

uncorrected thresholds must be set at 0.001

57 of 60

fMRI part 1

58 of 60

MTI fMRI lecture

in the linear combination predictors are dierent conditions

e.g .x1 is predicted signal upon presentation of words

x2 is the predicted signal upon presentation of nonsense stimuli

1 is the predictor estimate which tells you how much the bold signal in one voxel increases in
response to sentences

2 is the predictor estimate that tells you how much the signal increases in response to
nonsense words

1 and 2 are weighted meaning that if a voxel responds more to words that nonsense
stimulation 1 should be greater than 2

Analysis: the approach that works

1. find beta weights that best approximate a voxels signal

the best approximation is the one with less error

2. compare the beta weights for dierent predictors (sentences vs nonsense words)

in a GLM model we know

the design matrix

the BOLD signal

and we look for the beta weights

we look for beta weights that give the best approximation that is the one the minimizes the
sum of the squared errors

in the GLM we include as predictors : six predictors for head motion and predictors for time
derivatives that enable us to shift our approximation one time point back or forward.

that is because not all participants have the same hemodynamic function

GLM

1 Extract the signal time-series from a given voxel

2. run a GLM (with the signal and the design matrix as input)

3. compute the sum of squared errors (SSE)

4. Define the contrast and test it (with a t value)

fMRI part 1

59 of 60

We repeat that for every voxel

significant values are those that have t values that are not random

RFX using Summary statistics

fixed eect: models the mean

random eects: models the variance. random university eect

1. dierent GLM for each subjects or multiple subject-separable GLM

2. contrast vector for each subject

3. feed contrast vector to GLM

subjects should have identical design matrices

balanced design matrices at the first level

fewer voxels deemed active compared to average eect analysis. Why? because it takes into
account between subject-variability

Study Random Eects with:

1. Summary Statistics: we take the sample mean for each subject

2. Maximum Likelihood Estimators: we take the observed mean

Smoothing:

simple smoothing by mean: replace the values of 10 voxels by the mean value

with a Gauss kernel: weighted average for each voxel, in relation to the neighbors

RPV image: shows rebel per voxel= smoothness of data in each voxel

Bonferroni: uses number of independent voxels

Random Field: uses number of independent vessels

Convolusion of HDR:

canonical HRF: same for all brain regions

temporally basis function: dierent HRF for dierent brain regions

basis functions are determined by the parameter estimates

fMRI part 1

60 of 60

in the GLM framework the convolution of the stimulus function with each basis function is a
column in the design matrix

Finite Impulse response: one HRF for each event type in each voxel

Methods of convolution of HRF:

1. canonical HRF: fixed HRF

2. canonical HRF+ temporal derivative

3. canonical HRF+ temporal derivative+ dispersion derivative

4. FIR

5. smooth FIR

six. Inverse Logit Model

parametric modulation: a stimulus is parametrically varied across repetitions and this might
be reflected in the neuronal response. they are used to model trial to trial variation

we account for it by including an additional regressor in the design matrix

temporal autocorrelation:

nearby time-points are positively correlated because of

physiological noise

drift

Modeling of noise:

AR (p)

ARMA (1,1):