Multisite fMRI calibration and combination

1
Multi-site fMRI calibration and combination

• Jessica Turner
• Department of Psychiatry Human Behavior,
  University of California Irvine
• FIRST BIRN

2

• The challenges of combining data from different
  sites
• Rosen et alnot necessarily an issue at the study
  level
• But individual subjects from different sites?
• Needed for rare diseases, rare subtypes, etc.
• Statically
• Morph BIRN methods
• Dynamically
• Agar phantom
• Phase I study
• Replication and reliability
• Findings and methods
• 1) Variance STAPLE, variance components, FA
  measures
• 2) Methods
• Lees smoothness measures
• Greg Bs proportional scaling using SM
• Attempt to use the Phase I scalar and apply to
  MIND dataset not so good
• Our attempts to apply proportional scaling to MMN
  (conclusions?)
• Garys proportional scaling using BH
• 3) Conclusions

3
Biomedical Informatics Research Network

• NCRR funded initiative to develop solutions to
  large scale biomedical data sharing and mining
• Three neuroimaging test beds and a Coordinating
  Center
• Brain Morphometry BIRN
• Mouse BIRN
• Functional Imaging Research Schizophrenia Test
  bed (FIRST) BIRN
• BIRN CC

4
MRI reproducibility Sources of Variance

• 1. Measurement Noise
• 2. Geometric Differences
• 3. Long-term Geometric Instability
• 4. Short-term Geometric Instability
• 5. RF Field Differences
• 6. K-Space Trajectory
• 7. Field Strength
• 8. Registration/Alignment

• 9. Stimulus Presentation
  Differences
• 10. Physiological Noise
• 11. Subject Motion
• 12. Subject-Specific B0 Inhomogeneities
• 13. Subject Response Instabilities
• 14. Between-Subject Response Differences
• 15. Between-Subject Anatomical Differences

5
Variability

• Sources of variability
• Machine differences
• Geometric
• Temporal stability
• Activation variability
• Day to day variability within the same human
• Intersite differences imaging the same humans

6
Morphometry BIRNCalibration of Longitudinal and
Multi-site Structural MRI Studies An Evaluation
of Image Distortion and Image Intensity
Reproducibility J. Jovicich, D. Greve, E. Haley,
D. Kennedy, R.L. Gollub, B. Fischl, A. Dale,
Morphometry BIRN

• MRI Calibration Goals
• Characterize structural MRI test-retest
  reproducibility variance
• Understand sources of structural MRI variance
  correct them
• Initial focus
• 1.5T MRI, T1-weighted MRI based brain morphometry
  (AD, depression). Longitudinal multi-site
  studies.
• Common sequence (FLASH/SPGR) at multiple sites
• Source of variance investigated gradient
  non-linearities
• Phantom study
• Human test-retest study

7
Distortions from gradient non-linearities
Human test-retest
Phantom Scans from multiple sites
Single site
Multiple sites
No online in-plane distortion correction
Jovicich et al., 2004
8
Some vendors offer 2D gradient non-linearities
correction, but not 3D
Auto in-plane (sag) correction
No correction
3D-correction
Sagittal
?
OR
Coronal
This is what we want
Coronal
Sagittal
z-axis
9
Methods Gradients studied
Jovicich et al., 2004
10
Methods Phantom

• PHANTOM
• Diameter 250 mm
• Length 220 mm
• Plastic plates, spherical depressions
• 3D sphere-grid, spaced 20 mm
• Filled with H2O
• MRI Acquisition
• 3D Sagittal FLASH/SPGR
• Flip 300, TR20ms, TE6ms

11
Methods Human data Acquisition

• MRI ACQUISITION (at all sites)
• 3D Sagittal FLASH/SPGR
• TR20 ms
• TE 6 ms
• Flip angles 300, 200, 50, 30
• In plane resolution 1x1mm2
• Slice thickness 1.3 mm
• VOLUNTEERS
• 5 healthy volunteers
• Test-retest scans at all sites

http//www.nbirn.net/TestBeds/Morphometry/MRIAqui
sitionData.htm
12
Methods Data Analysis

• Gradient spherical harmonics from vendors
• Calculate displacement and intensity correction
  maps
• Distortion correction validation
• Phantom
• Test-retest reproducibility (within/across sites)
• Alignment of cortical surfaces
• Image intensity variability
• Subcortical volumes (in progress)
• Cortical thickness (in progress)

13
Phantom Results Distortion Correction
Deviation () from true phantom diameter along
z-axis
Phantom Images
Z-axis
Uncorrected Image
Corrected Image
Displacement Vector Field
Distance from magnets iso-center along z-axis
(mm)
Averaged across four sites
14
Human Results Improved co-registration
No distortion correction
3D distortion correction
Within Site (2 scans)
Across Sites (4 sites)
15
Reproducibility Effects Alignment of Surfaces
Siemens
GE
Same Subject Co-registered
CORTICAL ESTIMATES NO DISTORTION CORRECTION
CORTICAL ESTIMATES DISTORTION CORRECTION

• Distortion correction does improve cortical
  surface co-registration

Jovicich et al., 2004
16
Reproducibility Effects Image Intensity
Within-site
SINGLE SUBJECT EXAMPLE
Significant reduction of the mean intensity
variance
Across-site
Corrected
Uncorrected
Distortion corrected
Number of brain voxels
No correction
Relative reproducibility error
17
Improvement of across-site image intensity
reproducibility by using gradient distortion
correction
Distortion corrected
Mean image intensity variance and variance width
are reduced with distortion correction
Number of brain voxels
No correction
Relative reproducibility error

• Same subject scanned at
• MGH Siemens Sonata 1.5T (40 mT/m 200 mT/m/ms)
• UCSD GE Signa 1.5T (22 mT/m 120 mT/m/ms)
• Duke, BWH GE Signa 1.5T (40mT/m 150 mT/m/ms)

18
Reproducibility Effects Image Intensity

• Distortion correction does improve intensity
  reproducibility
• Other sources of variability are still present

Jovicich et al., 2004
19
FMRI variance and correction
EPI images from different scanners
20
Intersite Scanning Differences
Strength
K-space
Vendors
Raster
1.5T
GE
4T
Spiral
3T
Philips (Picker)
Dual Echo
Siemens
BWH MGH Duke/UNC MN NM IA Stanford
UCLA UCI UCSD
21
Intersite Scanning Differences
Strength
K-space
Vendors
1.5T
Raster
GE
4T
3T
Spiral
Philips (Picker)
Dual Echo
Siemens
BWH MGH Duke/UNC MN NM IA Stanford
UCLA UCI UCSD
22
Intersite Scanning Differences
Strength
K-space
Vendors
1.5T
Raster
GE
4T
3T
Spiral
Philips (Picker)
Dual Echo
Siemens
BWH MGH Duke/UNC MN NM IA Stanford
UCLA UCI UCSD
23
Human phantoms

• 5 healthy, right handed males
• Between 18 and 40 years old
• College students or graduates
• Lifelong English speakers
• Able to travel around the country from late July
  to mid September, 2003
• Common imaging parameters
• 35 slices
• 22 cm FOV
• 4 mm slices

24
Traveling Subjects Paradigm

• Scans at Each Site
• Cog
• Cog
• SM
• Rest
• SM
• BH
• SM
• BH
• Rest
• SM

AKA the Human Phantoms Five subjects Went to all
sites for scanning

• Cognitive Task Working Memory or Auditory
  mismatch
• SM - Sensory-Motor
• Rest - Eyes Open
• BH Breath Hold

25
Sensorimotor Paradigm
Block trials, 15s on, 15s off, 8 blocks. On
block Alternating contrast checkerboard.
Binaural tones generated (166 ms long with 167 ms
of silence.) Tones and visual contrast
change at 3Hz. Subject performs bilateral
finger tapping at 3Hz, in time with visual and
auditory cues. Off Block Fixation cross,
silence, no motion.
26
Breath holding paradigm
Block trials, 15s on, 15s off, 8 blocks. On
block Hold breath while circle of diminishing
size is projected. Off Block Fixation cross,
breathe normally Gray matter shows stronger
response than white matter.
27
Auditory paradigm

• Subjects saw a randomly luminance-reversing grey
  and white checkerboard, and pressed a button
  every time it reversed. This continued throughout
  the scans while in the background blocks of
• silent,
• constant, or
• mixed auditory stimuli were presented.
• The stimuli were either 1000 Hz or 1200 Hz tones,
  presented with an ISI of 300 ms each.

28
How repeatable are the results?

• Each subject was scanned in each scanner
• Each subject was scanned on two separate visits
• Each subject returned to their originating site
  at the end of the study for one more set of
  scans
• Are the results repeatable
• Within a session
• Across sessions
• Across scanners

29
Dependent measures
Percentage of voxels activated
Average signal strength
30
Sensorimotor task Site effects
Scan Cog Cog SM Rest SM BH SM BH Rest SM
Turner et al., 2004
31
Breathhold task Site effects
Turner et al., 2004
32
Intersite BOLD sensitivity Magnet Strength
Red 3T Blue 1.5T
Analysis performed by Lee Friedman, U New Mexico
33
Comparison across paradigms
34
Calibration possibilities

• Correct for inherent image sensitivity
  differences
• Black Box calibration scalar for each site
• Weight each site by its tendency to activate
• Which paradigm to use?
• Model-based analysis describing inter-group
  differences
• Bayes, GLM

35
Initial image smoothness and fMRI results
Sensitivity as a function of smoothness across
sites, showing high correlation between the two
and suggesting that if smoothness can be
equalized across sites, BOLD sensitivity will be
as well. (Analysis performed by L. Friedman, U
New Mexico)
36
Smoothing reduces intersite differences
UNSMOOTHED
SMOOTHED
Analysis performed by Lee Friedman, U New Mexico
37
Generalizability and dependability

• The generalizability coefficient divides the
  variance attributable to subject by the sum of
  the variance due to subject plus all other ANOVA
  effects involving subject, e.g. subject X site.
• Not sensitive to mean differences in site (or any
  of the other non-subject factors).
• The dependability coefficient divides the
  variance due to subject by the total variance.
• The dependability coefficient is increased by
  reproducibility between sites, i.e. by decreasing
  intersite variance.

Analysis performed by Greg Brown, UCSD
38
Proportional scaling SM

• Method Site-specific scalar based on the mean
  activation in an ROI over all sites

Analysis performed by Greg Brown, UCSD
39
Scalar method of reducing variance

• Using the regression weights as a measure of
  activation

Analysis performed by Greg Brown, UCSD
40
Auditory task results

• One subject, Mixed-tone condition only

MN 3T Duke 1.5T Duke 4T
UCI 1.5T BWH 3T
Top Day-1 Bottom Day-2
Analysis by J. Turner, UCI
41
Initial variability in auditory response
Right STG activation by visit (1 or 2), site
(coded 1-8), and grouped by magnet strength.
42
Effect of magnet strength
43
Behavior and STG activation
44
Using SM to calibrate Auditory results

• We tested four different calibration scalars,
  from the combination of two issues
• Which ROIs to include
• Which sites to include
• 1) ROIs Including six ROIs Left and right
  inferior and middle occipital gyrus, superior
  temporal gyrus, precentral sulcus
• Or using only L R STG
• 2) Sites to include Using all sites together to
  calculate the scaling factor
• Or combining 1.5T sites separately from 3 and 4T
  sites

45
Equations
46
Scalar calibration
The mean Z from all ROIs from the SM task
(MeanZsite).
47
Using SM to calibrate Auditory results

• Using Mean Z in Left and Right STG

48
Using SM to calibrate auditory results
Analysis performed by J. Turner, UCI and S.
Morris, UCSD
49
Relationship of BH and SM response
Analysis performed by Gary Glover and Lara
Foland, Stanford University
50
Reliability of BH over SM
51
Using BH to calibrate the SM results

• Using regression weights

Analysis done by Greg Brown, UCSD
52
Future Steps

• Continuing cross-site collaboration on a
  standardized dataset
• Reliability maps
• Bayesian approaches
• Smoothing methods
• Recommendations re calibration methods
• Smoothing a possibility
• BH seems more robust than SM
• BH shows promise for scaling cognitive tasks to
  improve generalizability and dependability
• Has to be tested on the auditory task
• Infrastructure to share and search these and
  similar data being developed

53
Infrastructure development

• Data sharing is essential and requires
  infrastructure
• Terabytes of data are not amenable to ftp or mail

INSTITUTION A
Imaging Data
Local Storage
DUP
Portal Mediator
Clinical Measures
Human Data Protection
Genotype Data
Federation distributed DB, autonomous, access
integrated resources Mediator translates
heterogeneous data sources to consistent
representation
54
Future directions and collaborations

• Future directions
• Application to other diseases
• Expansion to other cognitive tasks
• NIH Roadmap initiatives
• National Alliance for Medical Image Computing
  (NAMIC)
• www.na-mic.org
• Transdisciplinary Imaging Genetics Center (TIGC)
• Imaging Genetics Conference January 17th and 18th
  www.imaginggenetics.uci.edu