automated mri analysis: the academic and commercial options
TRANSCRIPT
Automated MRI analysis: academic and
commercial options
RASAD 2012
Clifford R Jack Jr
Alexander Family Professor of Alzheimer's Disease Research,
Dept Radiology, Mayo Clinic, Rochester, MN
Outline
Measurements covered
Cross sectional sMRI
Longitudinal sMRI
fcMRI
Current options
Caveats
Opportunities for innovation
Automated sMRI measures
Atlas on template
Atlas registered to
individual AD subject
that has been
segmented
Components
Template with atlas labels
Segmentation
Registration
Extract data from ROI parcellation
Current options: available MRI analysis SW
Freesurfer - Fischl, MGH
Neuroquant/Quark - Dale, Cortechs Labs
SPM/VBM (with atlas) - Ashburner
FSL - Smith
Automated Non-linear Image Matching and Anatomical Labeling
(Animal) - MNI
Caraet and Surface-based Atlases Wash U.
LONI Pipelines, UCLA
Medical Image Processing, Analysis, and Visualization – NIH
Advanced Normalization Tools (ANTS) – Avant, Penn
http://idoimaging.com/programs (lists >250 tools)
Opportunities for innovation - multi
voxel, multi ROI – “AD Signature”
Selecting Regions - 3 approaches
Use established knowledge of pathology to select ROIs.
NFT Braak stage neurodegenerative atrophy
Combine best performing ROIs diagnostically
Unbiased training algorithm – train on a ―gold standard‖
data set, test in independent data sets. Gold standard could
be autopsy confirmed AD, clinically diagnosed, etc
Voxel-wise methods capture time dependent progression from MCI to AD (n = 33)
MCI
3 years before conversion to AD
MCI
1 year before conversion to AD
AD
At time of conversion to AD
L R RL RL
3D Maps from Multiple MRI Illustrate Changing Atrophy Patterns as Subjects Progress from MCI to AD
Whitwell et al Brain 2007
Dickerson et al – “AD signature” in impaired and
pre clinical subjects – neocortical association +
medial temporal limbic
Dickerson BC, Bakkour A, Salat DH et al. Cereb
Cortex. 2009;19:497-510
Dickerson BC, Stoub TR, Shah RC et al. Neurology.
2011;76:1395-1402
Becker JA, Hedden T, Carmasin J et al. Annals of
neurology. 2010;Apr 13 [Epub]
Dickerson et al. Cereb Cortex, 2009
(A) Medial temporal cortex, (B) Inferior temporal gyrus, (C) Temporal pole, (D)
Angular Gyrus, (E) Superior frontal gyrus, (F) Superior parietal lobule, (G)
Supramarginal gyrus, (H) Precunes, (I) Inferior frontal sulcus, (J) visual reference
STAND algorithm for Individual Subject
Diagnosis - Vemuri et al Neuroimage 2008
Main Component of the STAND-Algorithm Large library of (AD and CN) MRI scans from which regions differentiating AD from
CN are detected and used to score new incoming cases.
MRI Scan STAND Algorithm ≥ 0 ABNORMAL
<0 NORMAL
STAND algorithm for Individual Subject
Diagnosis - Vemuri et al Neuroimage 2008
Large library of (AD and CN) MRI scans from which regions differentiating AD from CN
are detected and used to score new incoming cases. Accuracy of the method ~ 90 %
MRI Scan STAND Algorithm ≥ 0 ABNORMAL
<0 NORMAL
Longitudinal sMRI Measures
Measure each time point independently
Register serial images and compute change in image
space
Boundary Shift Integral
Freeborough and Fox, 1997
Affine registration
Time 1 Time 2
TBMSyN – CN vs AD rates
Two sample T-test of
TBMSyN atrophy maps
between AD (N=51) and
CN (N=51) Mayo 3T
subjects. FDR corrected
p<0.05 threshold.
Baseline and a follow-up
T1 MPRAGE, separated
by 12 to 18 months
--------------------------------
M Senjem
Bias in longitudinal sMRI measures
Holland and Thompson Neuroimage 2010 – bias in
Hua et al (Neuroimage 2010) TBM resulting in
underestimated sample sizes
Algorithms, atrophy and Alzheimer's disease:
Cautionary tales for clinical trials, Nick C. Fox,
Gerard R. Ridgway, Jonathan M. Schott, Neuroimage
2011
Bias in TBM – from Thompson and Holland Neuroimage
2011
Caveats: “ Fox rules”; bronze standard
Commutative or ―symmetric‖: Are the absolute changes
from A→B the same as from B→A?
Transitivity: if three sequential scans are available does
summing measures for A→B and B→C reproduce the
result of directly measuring A→C?
Comparison with low-bias (but high-variability) manual
measurements even though they are too imprecise for
typical trials
Comparison with other more established techniques on the
same data-set
Caveats: “ Fox rules”; bronze standard
Assessment of ―reproducibility‖ with short interval scans.
Atrophy should tend to zero as the interval shortens. Whilst
there may be a degree of variability, should be no mean
group change with ―same-day‖ scans.
Comparison with the known disease biology and
pathological studies, where available. Example, cross-
sectional hipp vol reduced by ~15 to 20% at time of AD
dementia diagnosis. If atrophy starts several years prior to
symptom onset, this is compatible with a measured rate of
hippocampal atrophy of no more than 3–5% per year –
50%/yr is not plausible.
MR measures beyond sMRI
MRS, ASL, DTI: Difficult to see how to standardize
first step, acquisition, using vendor product sequences
complex spin preparation and read out – differ among
vendors
phantoms for standardized measures difficult – diffusion
and perfusion
fcMRI echo planar imaging: spin preparation and
read out standard (caveats: ramp sampling, spirals)
Functional MRI – task free vs task
based
Imaging of extrinsic perturbations of fMRI
time course by task activation task fMRI
Imaging of intrinsic network connectivity
(fcMRI) task-free fMRI
Task Free fMRI - history
First description of task free fMRI - Biswal, B., et al.,
Functional connectivity in the motor cortex of resting human
brain using echo-planar MRI. Magn Reson Med, 1995,
spontaneous low-frequency fluctuations (0.1-0.01 Hz)
observed in BOLD) signal were highly correlated
within sensory motor cortex
BOLD
Signal
Acquired
Images (Sampling over time) Time (Sec)
Resting State - fMRI Acquisition
Task Free fMRI (resting state) –
functional connectivity (fcMRI)
TF-fMRI = functional connectivity. Represents a measure of
correlated signal from two or more spatially distinct regions over
time
Low-frequency fluctuations are specific to gray matter and can be
used to identify the spatial extent of temporally correlated
networks of functional connectivity within the brain
These large-scale networks are present at all times in the living
human brain
resting state networks are more accurately referred to as intrinsic
connectivity networks (ICNs)
Task Negative Network aka DMN. 342 CN. Positive correl, 6
mm seed PCC. ICA (20 components) within the TNN (red-
anterior DMN, blue-posterior DMN, green ventral DMN)
TPN, 342 CN. PCC seed, negative correlations, aka anti-
correlations. Four ICAs within TPN (red-salience, blue-dorsal
attention, green-left executive control, violet-right executive)
Slice timing correction
Realignment
Normalization
(to SPM EPI template)
Spatial smoothing
(FWHM = 4mm)
Linear temporal detrending
Temporal bandpass filtering
(ideal filter with cutoffs of 0.01 to 0.08 Hz)
Remove covariables
(realignment parameters, global signal, WM
signal, CSF signal)
Standard
Pre
processing
TF-fMRI data analysis methods
extract the spatial and temporal extent of ICNs
seed-based correlation
seeds may consist of an individual voxel, group of contiguous
voxels, or larger functionally/anatomically derived regions of
interest (e.g. Brodmann areas)
Seed to brain
Seed to seed (node to node)
In phase or out of phase masking
Graph theory – node to node connectivity matrices
data driven multivariate analysis techniques, eg
independent component analysis (ICA)
Popular fMRI software
Standard Preprocessing fMRI Pipelines
SPM (http://www.fil.ion.ucl.ac.uk/spm/)
FSL (http://www.fmrib.ox.ac.uk/fsl/)
AFNI (http://afni.nimh.nih.gov/afni/)
Brain Voyager (http://www.brainvoyager.com/)
Resting state funct connectivity and network analysis packages
REST (http://www.restfmri.net/)
GIFT ICA (http://mialab.mrn.org/software/#gica)
Melodic ICA (http://www.fmrib.ox.ac.uk/fsl/melodic/index.html)
BrainScape (http://nrg.wikispaces.com/Brainscape+About)
CONN Toolbox (http://web.mit.edu/swg/software.htm)
Brain Connectivity Toolbox (https://sites.google.com/a/brain-
connectivity-toolbox.net/bct/)
fcMRI Characteristic connectivity pattern with AD spectrum
Decreased in phase connectivity in posterior DMN
Increased in phase connectivity in anterior network regions
Anterior – posterior disconnection
Connectivity changes may occur early in disease process
CN amyloid positive – ie preclinical phase of AD
Prior to amyloid - Posterior DMN connectivity may be
increased in APOE 4 early in life
History of MRI technology arc
Considerable potential, but only if standardized – ie opportunities
exist
Conclusions
Opportunities: innovation that
conforms to standards set by the field
Multi voxel diagnostics cross sectional – sMRI
Multi voxel diagnostics longitudinal – sMRI
fcMRI – pre processing and analysis
the recipe outlined for HV proposed earlier is
rationale and would serve as a good temple for future
efforts
Conformity vs innovation Not conflicting goals
Many examples of standards agreed upon by industry
competitors – have not impeded innovation
SMTP (send mail transfer protocol) – email
DICOM (Digital Imaging and Communications in Medicine) -
standard for handling, storing, printing, and transmitting
information in medical imaging
WIFI (802.11 is main protocol, b,g,n are flavors) - nearly all pad
computers can communicate over (ie speak) 802.11. Doesn’t
prevent the creation of vertical markets for improvements —
$500 iPad vs $300 Galaxy Tab
Conformity = standardization = widespread utility
How are standardized MRI
measures best distributed to the
medical community?
Proposal
By MR manufactures
Historically this is how technology is distributed
integrated chain running on scanner
acquisition
QC
Preprocessing artifact correction
extract relevant diagnostic info from individual scan
Relate individual scan to appropriate standard data base
Output value – that conforms to agreed upon standards
manufacturer assumes responsibility for insuring forward
compatibilty – consistency - with every upgrade
Difficult for stand alone SW developer
Proposal for MR manufacturers
having complicated 1 of a kind pipelines for both data pre
processing and analysis (and normative data bases) running
at every medical center is not a path to standardization
MR vendors are not going to distribute identical products,
but could distribute their own pipelines where the final
output meets standards set by the field
Standards established by an accredited or empowered expert
board with appropriate expertise, appropriate
representativeness, and without commercial or institutional
bias