design constraints & optimisation in fmri · 2019-03-26 · fmri noise • physiological noise...
TRANSCRIPT
![Page 1: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/1.jpg)
Design constraints & optimisation in fMRI
Alexa Morcom
Edinburgh SPM course 2019
Thanks to Rik Henson, Cyril Pernet for slides
![Page 2: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/2.jpg)
Aim
• BOLD fMRI does not give an absolute measure of activity
Always compare activity across conditions using contrasts
Design task, conditions and timing to maximise an effect of interest
Large difference
Tiny difference
![Page 3: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/3.jpg)
Overview
04 Design constraint & optimisation
• fMRI signal and noise
• Temporal design
• Regressor correlation
• A word on power
![Page 4: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/4.jpg)
fMRI signal and noise
• fMRI signal is weak: functionally significant changes <5% signal intensity, so get all the signal you can
• Main sources of noise
– Thermal (intrinsic) noise
– System noise e.g. scanner drift, RF coil issues
– Image distortions e.g. susceptibility, Nyquist ghosting
– Physiological noise, mainly cardiac & respiratory rhythms/ movement
![Page 5: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/5.jpg)
fMRI noise
• Thermal noise = due to electron motion, in scanner circuits and participant
– This noise is unstructured (‘white’) and increases with higher resolution (smaller voxel size)
![Page 6: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/6.jpg)
fMRI noise
• Scanner drift = due to small changes in scanner magnetic field gradients over time
– Slow changes in voxel intensities with time which are not of interest = 1 type of low frequency noise
– Model separate ‘session constants’ per run
![Page 7: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/7.jpg)
fMRI noise
• Image distortions = nature of EPI plus specific sequence used and brain/skull properties*
– Nyquist ghost ‘wraparound’
– Ask physicist – can adjust scanner shim etc
* There are lots of types of distortion related to equipment &/or participant, this is a huge simplification
![Page 8: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/8.jpg)
fMRI noise
• Image distortions = nature of EPI plus specific sequence used and brain/skull properties
– Susceptibility artefact – magnetic susceptibility of scanned tissue, like air/brain interfaces
– Particularly orbitofrontal and anterior temporal
– Ask physicist about different slice angles or different sequences e.g. multi-echo
Structural (LEFT) and functional (RIGHT)
![Page 9: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/9.jpg)
fMRI noise
• Image distortions = nature of EPI plus specific sequence used and brain/skull properties
– Stretching of images
in slice ( z-) direction
– Inhomogeneity of field – physicist will minimise with shimming but can correct with fieldmap
– Also interacts with motion, correct w/ fieldmap
– And: artefacts due to metal objects
SPMmip
[-42
, -48
, -27
]
<
< <
SPM{T25
}
remembered vs. fixation
SPMresults: .\dm rfx\enc\fxvsbase\rvbasePHeight threshold T = 5.948441 {p<0.05 (FWE)}Extent threshold k = 0 voxels
Design matrix0.5 1 1.5 2 2.5
5
10
15
20
25
contrast(s)
3
0
5
10
15
SPMmip
[-42
, -48
, -27
]
<
< <
SPM{T25
}
remembered vs. fixation
SPMresults: .\dm rfx\enc\fxvsbase\rvbasePHeight threshold T = 5.948441 {p<0.05 (FWE)}Extent threshold k = 0 voxels
Design matrix0.5 1 1.5 2 2.5
5
10
15
20
25
contrast(s)
3
0
5
10
15
![Page 10: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/10.jpg)
fMRI noise
• Physiological noise = pulsations due to heart rhythm and breathing
• High frequency artefacts e.g. heartbeat every ~750 ms, sampled by slower scan e.g. TR 2 sec, ‘aliasing’ gives another type of low frequency noise
Respiratory CardiacLund et al. (2006)
![Page 11: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/11.jpg)
fMRI noise
• Physiological noise = head motion
• Very important – of order of half a voxel can cause serious issues due to partial volume effects, may need to exclude participant
• Correct in preprocessing and modelling steps
Lund et al. (2006)See Friston et al. (1995)
![Page 12: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/12.jpg)
Interim summary: fMRI signal & noise
• fMRI is very noisy
• Multiple sources due to equipment, participant, sequences and their interactions
• Most noise is low frequency – this also affects the timing of your task, and modelling
• MUCH better to reduce than to adjust, esp. motion
![Page 13: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/13.jpg)
Task timing
![Page 14: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/14.jpg)
Task design, signal and noise
• Must filter out large low frequency effects > ~0.01 Hz (1 cycle per 100 sec)
• To avoid removing effects of interest at the same time, need fairly high experimental design frequency
BUT
• ‘Sluggish’ BOLD response with slow response to neural changes effectively filters out high frequencies
• So also avoid very rapid expt. changes
![Page 15: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/15.jpg)
Task timing
Analysis of whole block
Large effects (=efficient)
Block model assumes summation of BOLD responses to successive stimuli
Analysis of single items
Smaller effects
Ability to distinguish BOLD responses to intermixed stimuli
[here: slow event-related]
![Page 16: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/16.jpg)
Alternating block design
• Block and rest (or control) alternating
• If shorter SOA greater summation of responses, increasing signal strength
• BUT at too short SOA may get psychological issues and/or refractoriness of BOLD response 1 / 3.3 / 5 s SOA
![Page 17: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/17.jpg)
Alternating block design
• Block and rest (or control) alternating
• If longer cycle > 10-15 seconds, get return to baseline between blocks so signal strength is higher
• Due to sluggish haemodynamic response
5 / 10 / 15 / 30 s blocks
![Page 18: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/18.jpg)
Alternating block design
5 sec blocks
45 sec blocks
15 sec blocks
![Page 19: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/19.jpg)
Interim summary: block design
• Short SOA and long blocks good
• BUT avoid block frequency overlapping with low frequency noise
• Trade-off
Optimum block duration ~16 s gives cycle length ~= duration of haemodynamic response function
![Page 20: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/20.jpg)
Why go event-related?
Advantages of event-related design
• Intermixing of conditions avoids unwanted psychological effects e.g. habituation, expectancy, loss of concentration
Old item
Old item Old
item Old item […]
New item
New item New
item New item […]
Old item New
items[…]
Old item
New item
![Page 21: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/21.jpg)
Why go event-related?
Advantages of event-related design
• Intermixing of conditions avoids unwanted psychological effects e.g. habituation, expectancy, loss of concentration
• Post-hoc classification of trials, e.g. Subsequent memory effect
Rem
Forget
[…]
Rem
Forget
![Page 22: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/22.jpg)
Why go event-related?
Advantages of event-related design
• Intermixing of conditions avoids unwanted psychological effects e.g. habituation, expectancy, loss of concentration
• Post-hoc classification of trials, e.g. Subsequent memory effect
• Some events can only be indicated by subject at particular time e.g. Spontaneous perceptual changes
![Page 23: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/23.jpg)
![Page 24: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/24.jpg)
Paradigm timing
Advantages of event-related design
• Intermixing of conditions avoids unwanted psychological effects e.g. habituation, expectancy, loss of concentration
• Post-hoc classification of trials, e.g. Subsequent memory effect
• Some events can only be indicated by subject at particular time e.g. Spontaneous perceptual changes
• Some events cannot be blocked, e.g. oddball
![Page 25: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/25.jpg)
Interim summary: event vs. block
• Detection power less for event-related designs – signal depends on how many events averaged
• Estimation power better for event-related, as can determine the shape of the haemodynamic response (though this means assumptions about its shape)
• Event-related much more flexible for cognitive studies
• But if you can go blocked, do!
![Page 26: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/26.jpg)
Individual stimuli 32s apart
Individual BOLD responses: back to baseline
NOTE: a slow event-related design avoids any overlap of event responses
Very slow!
Used sometimes where separation of responses is really critical
But less detection power as fewer trials
![Page 27: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/27.jpg)
Stimuli random 50% chance every 4s
= equivalent to having 50% trials of interest, 50% ‘null’ events
Mixture of ‘on’ vs. ‘off’ measurement & separation of event-types
This rapid event-related design takes HRF into account in the model
and allows overlap of event responses within & between conditions
MUCH improves temporal resolution
This = a simple randomised design
![Page 28: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/28.jpg)
Rapid event-related designs
• Random – on each trial, condition A or B determined randomly
• 1st approximation: good to randomise order, or SOA, of trials close together in time
• Semirandom – another systematic variation of probability of A or B over time
• Can be the most efficient
http://imaging.mrc-cbu.cam.ac.uk/imaging/DesignEfficiency
![Page 29: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/29.jpg)
Rapid event-related designs
What is jitter?
• Randomly or systematically varying the intervals between trials
• Can be done by varying SOA/ ISI or adding ‘null events’
Why jitter?
• To achieve both ‘clumping’ and separation of trial types/ conditions and improve efficiency
• To ensure that entire haemodynamic response is sampled over trials
http://imaging.mrc-cbu.cam.ac.uk/imaging/DesignEfficiency If long TR
Top = no jitterBottom = jitter
![Page 30: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/30.jpg)
Statistical efficiency
• To maximise our statistic, minimise the variance of the contrast involved
• Efficiency defined this way depends on the fitted model
and the contrast
and the design matrix
• Efficiency is relative, compare different models with same contrast & no. of scans
T = c'*b / sqrt(var( c'*b ))
var( c'*b ) = s2 c'*inv(X'X)*c
c = contrastb = the parameter estimates contrasteds2 = noiseX = design matrix
Optimise trial timing using e.g. OptSeq, or manually calculate efficiency
![Page 31: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/31.jpg)
Correlated regressors, for example:green: subject age or WM delay activityyellow: subject score or WM encoding
Testing for the green:
SPM contrast
Correlated regressors
![Page 32: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/32.jpg)
Highly correlated.Entirely correlated non estimable
Testing for the green:
Correlated regressors
Can test for ‘effect of either or both’ but cannot separate after the fact
![Page 33: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/33.jpg)
Correlated regressors
• If contrast involves correlated (collinear) regressors, it will be less efficient.
• E.g, mixed blocked/ event-related design for state and item-related activity (Chawla et al., 1999)
Efficiency again*
• e = 1 / trace (C (XTX)-1 CT)
• Efficiency is inversely proportional to design covariance
* ignoring noise and using matrix form
Pseudorandom distribution of SOAs min = 2 s
Fixed SOA = 4 s
![Page 34: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/34.jpg)
Interim summary
• Block design most efficient for detection of a response, highest statistical efficiency
• Low frequency experimental changes must be avoided, e.g. optimal block on/off ‘fits’ BOLD response ~32 s
• Event-related best for estimation of the course of the BOLD response, and for flexibility
• Random* intermixing of conditions and/or** jitter the SOA
• Statistical efficiency improved by low correlation of regressors
* pseudorandom might be better, see Liu et al., 2001** for some recipes see http://imaging.mrc-cbu.cam.ac.uk/imaging/DesignEfficiency
![Page 35: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/35.jpg)
Other important stuff
• How long should I scan each participants for?
As many trials as possible
As long as you and the participant can afford
• How many scanner runs (sessions) do I need?
For univariate studies, 1 is best*
More increases variance, reduces efficiency
Keep contrasts within-session
• How many participants do I need?
* The opposite applies for multivariate studies – Advanced Course!
![Page 36: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/36.jpg)
A note about power
• Statistical power in a multilevel model depends on no. of observations, and variance, at both levels of the model
• Insufficient no. of trials and insufficient N can mean your study is underpowered!
Too few trials will never get there
Too low N will never get there
NB just one example, set mean/ variance/ constraints, will NOT generalise!
![Page 37: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/37.jpg)
A note about power
Mumford & Nichols, 2008
What N? No. of needed cycles per participant for .80 powerN = 22 looks best
What N? Cost of achieving .80 power by varying cycle length and NN = 20 looks best
![Page 38: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/38.jpg)
A note about power
• Statistical power in a multilevel model depends on no. of observations, and variance, at both levels of the model
• Either insufficient no. of trials or insufficient N can mean your study is underpowered!
• As power depends on statistical efficiency, need to know design X to fully specify – can do using pilot data, see fmripower.org *
• Alternatively from old data can estimate effect size at group level and determine N using non-specialist software – but don’t ignore it!
neuropowertools.org also from Nichols & others; see andysbrainblog.blogspot.com
and https://www.biorxiv.org/content/10.1101/049429v1
![Page 39: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/39.jpg)
Summary
Main message: different designs for different questions
Don’t forget
• Scan for as long as possible (lots of trials, participants)
• your task-related changes of interest need to take place slowly but not too slowly
Want to know more?
• Temporal design efficiency and design optimisation
• http://imaging.mrc-cbu.cam.ac.uk/imaging/DesignEfficiency
![Page 40: Design constraints & optimisation in fMRI · 2019-03-26 · fMRI noise • Physiological noise = head motion • Very important – of order of half a voxel can cause serious issues](https://reader035.vdocument.in/reader035/viewer/2022070802/5f02c31d7e708231d405e29e/html5/thumbnails/40.jpg)
Thank you!