Transcript susanwhitfieldgabrieli

Quality Assurance
NITRC Enhancement Grantee Meeting
June 18, 2009
Susan Whitfield-Gabrieli & Satrajit Ghosh
RapidArt
MIT
Acknowledgements
THANKS!
Collaborators:
• Alfonso Nieto Castañón
• Shay Mozes
Data:
• Stanford, Yale, MGH, CMU, MIT
Funding:
• R03 EB008673: PIs: Satrajit Ghosh, Susan Whitfield-Gabrieli, MIT
fMRI QA
• Data inspection as well as artifact detection and
rejection routines are essential steps to ensure
valid imaging results.
• Apparent small differences in data processing
may yield large differences in results
QA in fMRI
Before Quality Assurance
QA in fMRI
Before QA
After QA
QA: Outline
• fMRI quality assurance protocol
• QA (bottom up)
• QA (top down)
Quality Assurance: Preprocessing
Bottom Up: review data
Raw
Images
Artifact
Detection
Preprocessing
Review Data
Check behavior
Create mean functional image
Review time series, movie
Interpolate prior to preprocessing
Quality Assurance: Post Preprocessing
Top Down: review stats
Bottom Up: review functional images
PreProc
Data Review
- time series
- movie
Artifact
Check
GLM
Artifact
Check
- Check registration
- Check motion parameters
- Generate design matrix template
- Check for stimulus corr motion
- Check global signal corr with task
- Review power spectra
- Detect outliers in time series, motion:
determine scans to omit /interp or deweight
RFX
Artifact
Check
Review Statisitcs
Mask/ResMS/RPV
Beta/Con/Tmap
Data Review
Global
mean
COMBINED
OUTLIERS
INTENSITY
OUTLIERS
Deviation
From mean
Over time
Thresholds
MOTION
OUTLIERS
Realign
Param
Outliers
Data Exploration
Including motion parameters as covariates
1.
2.
3.
Eliminates (to first order) all motion related residual variance.
If motion is correlated with the task, this will remove your task activation.
Check SCM: If there exists between group differences in SCM, AnCova
Power Spectra: HPF Cutoff Selection
.01 .02
Artifact Detection
Scan 79
Scan 95
Artifact Detection/Rejection
Artifact Sources:
Head motion *
Physiological : respiration and cardiac effects
Scanner noise
Solutions:
Review data
Apply artifact detection routines
Omit*, interpolate or deweight outliers
*Include a single regressor for each scan you want to remove, with a 1
for the scan you want to remove, and zeros elsewhere.
*Note # of scan omissions per condition and between groups
Correct analysis for possible confounding effects:
AnCova : use # outliers as a within subject covariate
BOTTOM UP
AUDITORY RHYMING > REST
Outlier Scans
ResMS
T map
Before ART
ResMS
After ART
T map
“TOP DOWN” 2nd level, RFX
Group Stats ( N = 50 )
Working Memory Task
Not an obvious problem:
Frontal and parietal
activation for a working
memory task.
Group Stats (N=50)
2B Working Memory Task
Find Offending Subjects: 2 of 50 subjects
Artifacts in outlier images
Scan 86
Scan 79
Scan 83
Scan 95
Comparison of Group Stats:
Working Memory (2B>X)
ORIGINAL
FINAL
Comparison of Group Statistics:
Default Network
Method Validation Experiment
• Data analyzed: 312 subjects, 3 sessions per subject
• Outlier detection based on global signal and movement
• Normality: tests on the scan-to-scan change in global BOLD signal after
regressing out the task and motion parameters. Normally-distributed
residuals is a basic assumption of the general linear model. Departures from
normality would affect the validity of our analyses (resulting p- values could
not be trusted) If all is well, we should expect this global BOLD signal
change to be normally distributed because: average of many sources (central
limit theorem )
• Power: the probability of finding a significant effect if one truly exists.
Here it represents the probability of finding a significant (at a level of p<.001
uncorrected) activation at any given voxel if in fact the voxel is being
modulated by the task (by an amount of 1% percent signal change).
Outlier Experiment
• Global signal is not normally distributed
In 48% of the sessions the scan-to-scan change in average
BOLD signal is not normally distributed.
This percentage drops to 4% when removing an average of 8
scans per session (those with z score threshold = 3)
Removing outliers improves the power
• Plot shows the average power to detect a task effect (effect size = 1% percent
signal change, alpha = .001)
• Before outlier removal the power is .29 ( 29% chance of finding a significant
effect at any of these voxels) After removing an average of 8 scans per
session (based on global signal threshold z=3) power improves above .70
THANKS!
Dissemination (NITRC)
- International visiting fMRI fellowships @ MGH
- 2 week MMSC @ MGH
- SPM8 Courses (local/remote)
-Visiting programs at MIT
Documentation
• Manuals, Demos, Tutorials
• Scripts