fsl feat analysis

8/11/2019 fsl feat analysis

1/94

FMRI Group Analysis

GLM

Design matrix

Effect size subject-series

Voxel-wise group analysis

Group

effect size

statistics

Subject

groupings

1

1

1

1

1

1

0

0

0

0

0

0

0

0

0

0

0

0

1

1

1

1

1

1

Standard-space

brain atlas

subjects

Single-subject

effect size

statistics

Single-subject

effect sizestatistics

Single-subject

effect size

statistics

Single-subjecteffect size

statistics

subjects

Register

subjects into

a standard

space

Effect size

statistics

Statistic ImageSignificant

voxels/clustersContrast

Thresholding


2/94

uses GLM at both lower and higher levels

typically need to infer across multiple subjects,sometimes multiple groups and/or multiple sessions

questions of interest involve comparisons at the

highest level

Multi-Level FMRI analysis

Group 2

HannaJosephine Anna Sebastian Lydia Elisabeth

Group 1

Mark Steve Karl Will Tom Andrew

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 4

Difference?


3/94

Does the group activate on average?

A simple example

Group



4/94


0 effect size

A simple example

Group


Yk = Xk k + k

First-level GLMon Marks 4D FMRI

data set


5/94


6/94


0 effect size

A simple example

Group


Yk = Xk k + k

Marks

within-subject

variance


7/94


8/94


What group mean are we after? Is it:

1. The group mean for those exact 6 subjects?

Fixed-Effects (FE) Analysis

2. The group mean for the population from which

these 6 subjects were drawn?

Mixed-Effects (ME) analysis

A simple example

Group



9/94

Do these exact 6 subjects activate on average?

Fixed-Effects Analysis

Group


0 effect size

g =1

6

6

k=1

k

estimate group effect size asstraight-forward meanacross lower-level estimates


10/94



Group


0 effect size

YK= XKK+ K

K = Xgg

g =1

6

6

k=1

kXg =

1

1

1

1

1

1

Group

mean


11/94



Group


0 effect size

K = Xgg

Xg =

1

1

1

1

1

1

Group

mean


12/94


Consider only these 6 subjects estimate the mean across these subject only variance is within-subject variance


Group


YK= XKK+ KK = Xgg

Fixed Effects Analysis:


13/94


What group mean are we after? Is it:

1. The group mean for those exact 6 subjects?

Fixed-Effects (FE) Analysis

2. The group mean for the population from which

these 6 subjects were drawn?

Mixed-Effects (ME) analysis

A simple example

Group

Mark Steve Karl Keith Tom Andrew


14/94

0 effect size

Does the population activate on average?

Mixed-Effects Analysis

Group


0 effect size

YK= XK K+ K

gk

Consider the distribution over thepopulation from which our 6subjects were sampled:

2

g is the between-subject variance

g


15/94

0 effect size



Group


YK= XK K+ K

gk

Consider the distribution over thepopulation from which our 6subjects were sampled:

2

g is the between-subject variance

g


16/94



Group


YK= XKK+ K

0 effect sizeg

k

gK= Xgg + g

Xg =

1

1

1

1

1

1

Population

mean

between-subject

variation


17/94



Group


0 effect sizeg

k

gK= Xgg + g

Xg =

1

1

1

1

1

1

Population

mean

between-subject

variation


18/94


Consider the 6 subjects as samples from a wider population estimate the mean across the population between-subject variance accounts for random sampling


Group


YK= XKK+ K

Mixed-Effects Analysis:

K= Xgg + g


19/94

All-in-One Approach

Could use one (huge) GLM to infer group difference

difficult to ask sub-questions in isolation

computationally demanding

need to process again when new data is acquired

Group 2

HannaJosephine Anna Sebastian Lydia Elisabeth

Group 1


session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 1

session 2

session 3

session 4

session 1

session 2

session 3

session 4

Difference?


20/94

Summary Statistics Approach

At each level:

Inputs are summary stats from levels

below (or FMRI data at the lowestlevel)

Outputs are summary stats or

statistic maps for inference

Need to ensure formal equivalence

between different approaches!

In FEAT estimate levels one stage at a time

Group

Subject

Session

Groupdifference


21/94

FLAME

Fully Bayesian framework

use non-central t-distributions:

Input COPES, VARCOPES & DOFsfrom lower-level

estimate COPES, VARCOPES &

DOFs at current level

pass these up

Infer at top level

Equivalent to All-in-One approach

FMRIBs Local Analysis of Mixed Effects

Group

Subject

Session

Groupdifference

COPES

VARCOPESDOFs

Z-Stats

COPES

VARCOPESDOFs

COPESVARCOPES

DOFs


22/94

FLAME Inference

Default is:

FLAME1: fast approximation for all voxels (using

marginal variance MAP estimates)

Optional slower, slightly more accurate approach:

FLAME1+2:

FLAME1 for all voxels, FLAME2 for voxels close to

threshold

FLAME2: MCMC sampling technique


23/94

Choosing Inference Approach

1. Fixed Effects

Use for intermediate/top levels

2. Mixed Effects - OLSUse at top level: quick and less accurate

3. Mixed Effects - FLAME 1

Use at top level: less quick but more accurate

4. Mixed Effects - FLAME 1+2

Use at top level: slow but even more accurate


24/94

FLAME vs. OLS

allow different within-levelvariances (e.g. patients vs.

controls)

allow non-balanced designs

(e.g. containing behaviouralscores)

allow un-equal group sizes

solve the negative varianceproblem

0 effect size

pat ctl

GroupSubjectSession

< 0

>0?

0

subject

effect size


30/94

Single Group Average

We have 8 subjects - all in one group - and want themean group average:


estimate mean

estimate std-error

(FE or ME)

test significance of

mean > 0

>0?

0

subject

effect size


31/94



0

subject

effect size


32/94



0

subject

effect size


33/94



0

subject

effect size


34/94




35/94

Unpaired Two-Group Difference

We have two groups (e.g. 9 patients, 7 controls)

with different between-subject variance

Is there a significant group difference?

estimate means

estimate std-errors(FE or ME)


difference in means

>0?

0

subject

effect size


36/94


We have two groups (e.g. 9 patients, 7 controls)

with different between-subject variance


estimate means

estimate std-errors(FE or ME)


difference in means

>0?

0

subject

effect size


37/94



0

subject

effect size


38/94



0

subject

effect size


39/94



0

subject

effect size


40/94



0

subject

effect size


41/94




42/94




43/94

Paired T-Test

8 subjects scanned under 2 conditions (A,B)

Is there a significant difference between conditions?

0

subject

effect size


44/94

Paired T-Test



0

subject

effect size


45/94

>0?

Paired T-Test



0

subject

effect size

try non-paired t-test


46/94



data

0

subject

effect size

Paired T-Test


47/94



data

0

subject

effect size

Paired T-Test

subject meanaccounts for large prop.

of the overall variance


48/94



de-meaned data

0

subject

effect size

data

0

subject

effect size

Paired T-Test

subject meanaccounts for large prop.



49/94



de-meaned data

0

subject

effect size

data

0

subject

effect size

Paired T-Test

>0?subject mean

accounts for large prop.



50/94

subject

effect size00

Paired T-Test



51/94

subject

effect size00

Paired T-Test



52/94

subject

effect size00

Paired T-Test



53/94

subject

effect size00

Paired T-Test



54/94

subject

effect size00

Paired T-Test



55/94

subject

effect size00

Paired T-Test



56/94

Paired T-Test


EV1models the A-B

paired difference; EVs2-9 are confounds

which model out eachsubjects mean


57/94

Paired T-Test



58/94

Multi-Session & Multi-Subject

5 subjects each have three sessions.


Use three levels: in the second levelwe

model the within-subject repeated measure


59/94




Use three levels: in the third levelwe model

the between-subjects variance


60/94




Use three levels:

in the second level we model the within subject

repeated measure typically using fixed effects(!)

as #sessions are small

in the third level we model the between subjects

variance using fixed or mixed effects


61/94

Reducing variance


mean effect size small

relative to std errormean effect size large

relative to std error

0

s

ubject

effect size

>0?

0

s

ubject

effect size

>0?


62/94

mean effect size large


Reducing variance


mean effect size large


0

s

ubject

effect size

>0?

0

s

ubject

effect size

>0?


63/94

We have 7 subjects - all in one group. We also have

additional measurements (e.g. age; disability score;

behavioural measures like reaction times):


use covariates to

explain variation

estimate mean

estimate std-error(FE or ME)

0

subject

effect size

Single Group Average & Covariates


64/94




0

subject

effect size


fastRTslow


65/94





0

subject

effect size


fastRTslow


66/94





use covariates to

explain variation

estimate mean


0

subject

effect size


fastRTslow

S l G A C


67/94





use covariates to

explain variation

estimate mean


0

subject

effect size


fastRTslow

S l G A & C


68/94





use covariates to

explain variation

estimate mean


0

subject

effect size


fastRTslow

S l G A & C


69/94


additional measurements (e.g. age; disability score;behavioural measures like reaction times):


use covariates to

explain variation

estimate mean


0

subject

effect size


fastRTslow

Si l G A & C i


70/94


additional measurements (e.g. age; disability score;behavioural measures like reaction times):


use covariates toexplain variation

estimate mean


0

subject

effect size


fastRTslow

Si l G A & C i


71/94


use covariates to explain variation

need to de-mean additional covariates!


FEAT G A l i


72/94

Run FEAT on raw FMRI data to get first-level .featdirectories, each one with several (consistent) COPEs

low-res copeN/varcopeN .feat/stats

when higher-level FEAT is run, highres copeN/varcopeN .feat/reg_standard

FEAT Group Analysis

FEAT G A l i


73/94

FEAT Group Analysis

Run second-level FEAT to get one .gfeat directory

Inputs can be lower-

level .feat dirs or

lower-level COPEs

the second-level GLM analysis is run separately

for each first-level COPE

each lower-level COPE generates its own .featdirectory inside the .gfeat dir

Th t ll f lk


74/94

Thats all folks

A di


75/94

Appendix:

G F t t


76/94

Group F-tests

Gro F tests


77/94

Group F-tests

3 groups of subjects

Group F tests


78/94

Group F-tests

3 groups of subjects

Is any of the groups activating on average?

ANOVA: 1 factor 4 levels


79/94

ANOVA: 1-factor 4-levels



80/94


8 subjects, 1 factor at 4 levels



81/94



Is there any effect?



82/94




EV1 fits cond. D, EV2 fits cond A relative to D etc.



83/94




EV1 fits cond. D, EV2 fits cond A relative to D etc.

F-test shows any difference between levels



84/94




85/94


8 subjects, 2 factor at 2 levels. FE Anova: 3 F-tests give

standard results for factor A, B and interaction



86/94




If both factors are random effects then Fa=fstat1/fstat3,

Fb=fstat2/fstat3ME



87/94




If both factors are random effects then Fa=fstat1/fstat3,

Fb=fstat2/fstat3ME

ME: if fixed fact. is A, Fa=fstat1/fstat3



88/94




89/94


16 subjects, 3 factor at 2 levels.



90/94



Fixed-Effects ANOVA:



91/94



Fixed-Effects ANOVA:

For random/mixed effects need different Fs.

Understanding FEAT dirs


92/94


First-level analysis:



93/94


Second-level analysis:

Thats all folks


94/94

That s all folks

fsl feat analysis

Documents

Transcript of fsl feat analysis