Voxel-based Lesion-Symptom Mapping Céline R. Gillebert
Voxel-based Lesion-Symptom Mapping
Céline R. Gillebert
Paul Broca (1861)
“Mr. Tan”
• no productive speech
• single repetitive syllable
‘tan’
Broca’s aphasia: problems with
fluency, articulation, word-finding,
repetition, production and
comprehension of complex
grammatical structures
Broca’s area: speech
production
Lesion-Symptom Mapping
= inferring the function of a brain area by
observing the behavioural consequences
of damage to that area
advantages
• stronger inference: Is brain area necessary for task?
fMRI, EEG, MEG: Does activity in brain area correlate with
task?
• infer function of node in network of areas
fMRI: difficult to understand the differential contribution of
areas that are simultaneously activated by the task
• clinical relevance: predict recovery or select best protocol
for rehabilitation of behavioural deficits
disadvantages
• Lesions do not respect the boundaries of functional
areas…
… and do not cover the whole brain, even not in the largest
possible sample of patients
• Lesions are permanent….
… although their relation to behavioural function depends on
the time to stroke (neuroplasticity)
• Lesions can cause dysfunction of structurally intact areas
at the distance
… lesion-symptom mapping is inherently a “localizationist
approach” http://www.strokecenter.org/
disadvantages
• Lesions do not respect the boundaries of functional
areas…
… and do not cover the whole brain, even not in the largest
possible sample of patients
• Lesions are permanent….
… although their relation to behavioural function depends on
the time to stroke (neuroplasticity)
• Lesions can cause dysfunction of structurally intact areas
at the distance
… lesion-symptom mapping is inherently a “localizationist
approach”
Example:
What brain injury leads to hemispatial
neglect?
Mort et al, 2003
Example: hemispatial neglect
Karnath et al. (2012).
Neuropsychologia
Example: hemispatial neglect
. . . . .
. .
.
. .
Demeyere et al. (under review). Psychological Assessment.
lesion overlap
• We can overlay the lesions of patients with a deficit on the
cancellation task.
• Example: Karnath et al. (2004). Cerebral Cortex
n=78
lesion subtraction
Patients with similar brain
damage but without the deficit
are critical to identify areas
related to the function on top
of areas that are commonly
damaged!
Karnath et al. (2004). Cerebral Cortex.
voxel-based lesion-symptom mapping
• Statistics to evaluate whether differences in lesion
frequency are reliable predictors of behavioural deficits.
• Example: Karnath et al. (2004). Cerebral Cortex
How to run a VLSM analysis?
How to run a VLSM analysis?
1. Acquisition of brain scan with visible
lesion
2. Delineation of the lesion
3. Normalization of lesion to a common
template
4. Statistics across a group of patients
CT versus MR scans
MRI scans
• no radiation (control data)
• higher spatial resolution
• different images with different
contrasts
CT scans
• clinical: acute haemorrhage visible
• when contraindication for MRI
• not ideal for research… but large database
Case RR, Oxford CNC Case RR, Oxford CNC
MR scans: different contrasts
T2-weighted scans
• Slower to acquire
• Excellent for finding lesions
• FLAIR attenuates CSF
T1-weighted scans
• Fast to acquire
• Good contrast between
WM and GM
• Excellent structural detail
Case RR, Oxford CNC Case RR, Oxford CNC
acute or chronic stroke?
• acute stroke: widespread dysfunction
• structurally intact brain areas are disrupted as they are
connected to the lesioned brain areas
• more clinically relevant
• chronic stroke: brain is plastic
• difficult to infer what a brain region used to do
• more stable, identifies functions that cannot be compensated
How to run a VLSM analysis?
1. Acquisition of brain scan with visible lesion
2. Delineation of the lesion
3. Normalization of lesion to a common
template
4. Statistics across a group of patients
lesion delineation
• Manual delineation of the lesion: “gold standard”
• requires experience and knowledge about brain anatomy
• time-consuming, only feasible for relatively small sample
sizes (but power of VLSM…)
• susceptible to operator bias
• Fully/semi-automated delineation
• replicable
• suitable for large sample sizes
• errors are inevitable
• “normal” signal varies from individual to individual
• lesions are heterogeneous in signal, also within an individual
Automated lesion delineation
• CT scans: Gillebert, C.R.,
Humphreys, G.W., & Mantini, D.
(2014). Automated delineation of
stroke lesions using brain CT
images. Neuroimage: Clinical,
4:540-548.
• MRI scans: Mah, Y.H., Jager, R.,
Kennard, C., Husain, M., & Nachev,
P. (2014). A new method for
automated high-dimensional lesion
segmentation evaluated in vascular
injury and applied to the human
occipital lobe. Cortex, 56:51-64.
Manual lesion delineation
Manual delineation of the lesion, slice by slice, using e.g.
MRIcron
Case RR, Oxford CNC Case RR, Oxford CNC
overview
1. Acquisition of brain scan with visible lesion
2. Delineation of the lesion
3. Normalization of lesion to a common
template
4. Statistics across a group of patients
normalization
• Alignment of brains to ‘template’ image in stereotaxic
space, necessary to compare lesions between individuals
• Linear and non-linear transformation to minimize
difference with template
normalization
• Alignment of brains to ‘template’ image in stereotaxic
space, necessary to compare lesions between individuals
• Linear and non-linear transformation to minimize
difference with template
• Use an appropriate (age- and modality-matched)
template:
Rorden et al. (2012). Neuroimage.
n=50,
73yrs
N=152,
25yrs
MNI152, SPM and FSL
n=30,
61yrs
n=366,
35yrs
Winkler et al. FLAIR Templates.
Available at http://glahngroup.org
normalization of CT scans: Gillebert et al. (2014) Neuroimage: Clinical
normalization
• ! Region of lesion appears different in image and template, and software will attempt to warp lesioned region
→ Solution: ignore the lesioned brain tissue in the process
→ Masked normalization: Brett et al., (2001) Neuroimage
→ Less of a problem with unified segmentation-normalization approach (Crinion et al. (2007) Neuroimage)
• Clinical toolbox for SPM
Clinical Toolbox in SPM Rorden et al. (2012). Neuroimage
http://www.mccauslandcenter.sc.edu/CRNL/clinical-toolbox
overview
1. Acquisition of brain scan with visible lesion
2. Delineation of the lesion
3. Normalization of lesion to a common
template
4. Statistics across a group of patients
visualization of lesion distribution
Molenberghs, Gillebert, et al., 2009
0 5 10 15 20 25 30 35 40 45 500
5
10
15
20
25
number of cancelled complete hearts
num
ber
of patients
Operationalization of behaviour
Demeyere*, Gillebert*, et al. (in preparation)
cut-off = 42
n=180
N=132
0 5 10 15 20 25 30 35 40 45 500
5
10
15
20
25
number of cancelled complete hearts
num
ber
of patients
Operationalization of behaviour
Demeyere*, Gillebert*, et al. (in preparation)
performance
Parametric or non-parametric statistics
• traditional: t-test for continuous data
• assumptions: data are normally distributed, two groups have similar variance, and
data represent interval measurements
• but
• assumptions difficult to test across the thousands of voxel-wise comparisons
• measures differences in the mean between two groups, not appropriate for
skewed distributions
• dependent variables often measured using an ordinal scale
• alternative: Brunner Munzel rank order test
• assumption free, also for variables on an ordinal scale
• Approaches normal distribution if n>= 10
correction for multiple comparisons
• Bonferroni-correction
• Strong protection against false alarms
• Overly conservatives when comparisons are not independent
• Permutation thresholding
• randomly relabeling and resampling the data, computing the maximum
observed statistic within the entire brain volume for each permutation
• lesions are formed from large contiguous regions, where each voxel is not
truly independent
• False discovery rate (FDR)
• controls the ratio of false alarms to hits
• sensitive where a signal is present in a substantial portion of the data
Some considerations…
• A t-test requires two groups and one continuous variable.
• The VLSM t-test is orthogonal to t-tests used for fMRI/VBM:
• fMRI/VBM t-tests:
• Deficit defines two groups.
• Voxel intensity provides continuous variable.
• VLSM
• Voxel intensity (lesion/no lesion) defines two groups.
• Behavioral performance provides continuous variable.
• Note VLSM group size varies from voxel-to-voxel.
• Statistical tests provide optimal power both groups have the same number of
observations (balanced).
• Therefore, VLSM power fluctuates across voxels
• We can not make inferences of voxels that are rarely damaged or always damaged (also true for
binomial tests).
Beyond VLSM…
• Track-wise “Hodological” Lesion-Deficit Analysis
• Thiébaut de Schotten et al. (2012) Cerebral Cortex
• maps of white matter tracts representing a probability of a given voxel
belonging to that tract
• calculating the size of the overlap (in cubic centimetres) between each
patient’s lesion map and each thresholded (50%) pathway map
→ Can these continuous measure of the pathway disconnection predict
behavioural deficits?
Beyond VLSM…
Chechlacz, Mantini, Gillebert, & Humphreys (under review). Cortex
Beyond VLSM…
• Track-wise “Hodological” Lesion-Deficit Analysis • Thiébaut de Schotten et al. (2012) Cerebral Cortex
• maps of white matter tracts representing a probability of a given voxel belonging to that tract
• calculating the size of the overlap (in cubic centimetres) between each patient’s lesion map and each thresholded (50%) pathway map
→ Can these continuous measure of the pathway disconnection predict behavioural deficits?
• Voxel-wise Bayesian Lesion-Deficit Analysis • Chen et al. (2008) Neuroimage
• Multivariate Lesion-Symptom Mapping (MLSM) • Zhang et al. (2014) Human Brain Mapping: Modelling the relation of the deficit to the entire
lesion map as opposed to each isolated voxel, using support vector regression
• Mah et al. (2014) Brain: capturing high-dimensional structure of lesion data using machine learning techniques