MS PowerPoint (PPTX format)

Modelling longitudinal structural
change from serial MRI
Ged Ridgway – [email protected]
John Ashburner
Colleagues at the FIL (WTCN) and the
Dementia Research Centre
Wellcome Trust Centre for Neuroimaging
UCL Institute of Neurology
Motivation for longitudinal data
Need for appropriate statistical analysis
Benefits of longitudinal image processing
Risk of bias from asymmetric processing
Longitudinal imaging in SPM12
Unbalanced data and further extensions
Motivation for longitudinal data
• Development, growth, plasticity, aging, degeneration,
and treatment-response are inherently longitudinal
• Serial data have major advantages over multiple
cross-sectional samples at different stages
• Increasing power
– Subtlety of change over time vs. inter-individual variation
• Reducing confounds
– Demonstrating causality with interventions
– Separating within-subject changes from cohort effects
Example: Training & structural plasticity
• Intervention (training) + longitudinal data allows
causal interpretation of change, cf. just difference
• Draganski et al. (2004) Neuroplasticity: Changes
in grey matter induced by training
– “volunteers who learned to juggle … transient and
selective structural change in brain areas associated
with processing and storage of complex visual motion”
• Draganski et al. (2006) Temporal and spatial
dynamics of brain structure changes during
extensive learning
Example: Training & structural plasticity
• Scholz et al. (2009) Training induces changes in
white matter architecture
Example: Training & structural plasticity
• Comments & Controversies, NeuroImage, 2013, 73:225–267
• Thomas & Baker: Teaching an adult brain new tricks: A critical review
of evidence for training-dependent structural plasticity in humans
• Erickson: Evidence for structural plasticity in humans: Comment on
Thomas and Baker (2012)
• [ Jones et al: White matter integrity, fiber count, and other fallacies:
The do's and don'ts of diffusion MRI ]
• Draganski & Kherif: In vivo assessment of use-dependent brain
plasticity—Beyond the “one trick pony” imaging strategy
• Fields: Changes in brain structure during learning: Fact or artifact?
Reply to Thomas and Baker
Thomas & Baker: On evidence, biases and confounding factors:
Response to commentaries
Example: Alzheimer’s disease evolution
• Multiple sources of cohort effects
– Birth-year (nutrition, etc.)
– Disease onset-time cohorts
– “Healthy survivor effect”
• Timescales too long for pure longitudinal studies
– “Unstructured multicohort longitudinal designs”
• See Thompson et al. (2011)
– [source of figure on next slide…]
Example: Alzheimer’s disease evolution
Further statistical issues
• Even simple designed experiments have pitfalls
• Usually seek group-by-time interaction
– Not significant change in one group but not another
– Not group difference at one time-point but not another
• Can’t ignore dependence within-subject over time
• In an ANOVA with group and time factors:
– Time effects can relate to (smaller) within-subject var.
– Group differences must relate to between-subject var.
– Group-by-time interaction …
Benefits of longitudinal image processing
• Smaller within-subject variation motivates
longitudinally-tailored image processing methods
• Boundary shift integral (BSI)
– Intensity difference after rigid registration over region from
brain masks more precise than mask volume diff.
• Non-rigid registration “Jacobian-integration”
– JI over segmented region more precise than multiple
independent segmentations (example following…)
• Temporally-constrained/regularised “4D” methods
– E.g. Xue’s CLASSIC, Wolz’s 4D graph-cut
Longitudinal imaging animations
Interpolating rigidly
aligned images
Warping average by
interpolated transform
Interpolating volume
change (divergence)
relative to the average
Benefits of longitudinal image processing
• Anderson et al. (2012) Gray matter atrophy rate
as a marker of disease progression in AD
Risk of bias from asymmetric processing
• Within-subject image processing often treats one
time-point differently from the others
– Later scans registered (rigidly or non-rigidly) to baseline
– Baseline scan segmented (manually or automatically)
• Asymmetry can introduce methodological biases
– E.g. only baseline has no registration interpolation error
– Baseline seg. more accurate than propagated segs.
– Change in later intervals more regularised/constrained
Risk of bias from asymmetric processing
• Theory known for a long
time (but often ignored)
– Ashburner et al. 1999;
Christensen, 1999;
Cachier & Rey, 2000;
Smith et al. 2001
• Demonstrated in practice
recently as a serious issue
– Thomas et al. 2009;
Yushkevich et al. 2010;
Thompson & Holland 2011
Risk of bias from asymmetric processing
• Comments & Controversies, NeuroImage, 2011, 57:1-21
• Thompson & Holland: Bias in tensor based morphometry
Stat-ROI measures may result in unrealistic power estimates
• Hua et al: Accurate measurement of brain changes in
longitudinal MRI scans using tensor-based morphometry
• Fox et al: Algorithms, atrophy and Alzheimer's disease:
Cautionary tales for clinical trials
• Reuter & Fischl: Avoiding asymmetry-induced bias in
longitudinal image processing
Longitudinal image processing in SPM12
• Ashburner & Ridgway (2013)
• “Unified” rigid and non-rigid registration with
model of differential intensity inhomogeneity (bias)
• “Generative” – each time-point is a reoriented,
spatially warped, intensity biased version of avg.
• “Symmetric” with respect to permutation of images
• “Consistent” with direct registration between pair
• “Diffeomorphic” – complex warping without folding
Generative model
Registr. (velocity)
Example result – Alzheimer’s disease subject
• Above: Images aligned only rigidly (OASIS data)
• Below: Non-rigid volume change (divergence)
Example result – Group averages
82 subjects from OASIS longitudinal data (part 1)
DARTEL for between-subject spatial normalisation
Divergences transformed without modulation
Next step could be SPM statistical analysis…
Terminology: TBM, DBM & (longitudinal) VBM
• (Deformation) Tensor-based morphometry (TBM)
– Davatzikos et al. (1996); Chung et al. (2001)
– SPM-like (mass-univariate) analysis of Jacobian or div
• See also mass-multivar. “generalized” TBM (Lepore et al. 2008)
• Deformation-based morphometry (DBM)
– Ashburner et al. (1998)
– Multivariate analysis of displacement vector patterns
• Longitudinal VBM (Kipps et al. 2005)
– Tissue-specific volume-change (using segmentation)
Longitudinal statistical modelling in SPM
• “Balanced” data (e.g. designed experiment)
– Same number (and timing) of time-points over subjects
– Repeated-measures / within-subject ANOVA
– Dependence within specified factor(s)
• “Unbalanced” data (e.g. observational study)
– E.g. more frequent observation closer to onset (DIAN)
– Two-stage (fMRI-like) analysis of summary statistics
– E.g. straight line or polynomial regression coefficients
• Sub-optimal if times vary dramatically (singletons dropped)
Other statistical modelling approaches
• Bernal-Rusiel et al. (2012) Statistical analysis of
longitudinal neuroimage data with Linear Mixed
Effects models. [FreeSurfer]
• Chen et al. (2013) Linear mixed-effects modeling
approach to FMRI group analysis. [AFNI]
• Li et al. (2013) Multiscale adaptive generalized
estimating equations for longitudinal neuroimaging
data. [unbalanced … twin and familial studies]
• Bayesian spatio-temporal modelling in SPM…
Demo of longitudinal imaging in SPM12
• Beta version released in December 2012 (phew!)
– Frequent updates until final release
• Record (and ideally report) the SPM12 revision number (r5360)
– Longitudinal registration relatively stable
• No longitudinal examples in SPM manual yet
– Possibly after SPM course in May…
– Support on SPM list, or email me (don’t email John!)
No Longitudinal
button, but found
in Batch menu,
like Dartel, etc.
Choice of paired
or general serial.
No difference in
model, but easier
specification and
results for pairs.
Specify Time 1
scans for all
subjects, then all
Time 2 scans (in
same order!)
Vector (list) of
time intervals (yr)
Default values
can be left; NaN
to automatically
estimate (Rician)
noise levels
One module
per subject
(scripting required
if many subjects!)
Select all scans
for this subject
Vector (list) of
times relative to
arbitrary datum
(e.g. baseline=0)
Jacobian output
useful to quantify
interpretable ROI
volumes (in litres)
• Output/results
– Average image
– Jacobians or
– Deformations
• Next steps
– Segment avg
– Run Dartel/Shoot
– Warp e.g. dv to
standard space
– SPM stats on dv
– Or combine with
seg of avg (VBM)
Modelling longitudinal structural
change from serial MRI
Ged Ridgway – [email protected]
This work was supported by the Medical Research Council
[grant number MR/J014257/1]
The Wellcome Trust Centre for Neuroimaging is supported by
core funding from the Wellcome Trust [091593/Z/10/Z]

similar documents