Voxelbased lesionsymptom mapping: An introduction and a tutorial

1
Voxel-based lesionsymptom mapping An
introduction and a tutorial

• Stephen M. Wilson
• Ahmanson-Lovelace Brain Mapping Center
• University of California, Los Angeles
• Center for Research in Language, UCSD
• October 22, 2004

2
Collaborators

• Elizabeth Bates, UCSD
• Ayse Pinar Saygin, UCSD
• Frederic Dick, UCSD
• Marty Sereno, UCSD
• Robert Knight, UC Berkeley
• Nina Dronkers, UC Davis
• Thanks also to David Wilkins and Carl Ludy. This
  work was funded by the Department of Veterans
  Affairs Medical Research, the National Institute
  of Neurological Disorders and Stroke (PO1
  NS17778, NINDS 21135, PO1 NS40813), and the
  National Institute of Deafness and Communication
  Disorders (NIH/NIDCD 2 R01 DC00216).

3
Lesion-symptom mapping

• Until a few decades ago, themain sources of
  evidence onbrain/behavior relationshipswere
  cases which came toautopsy.
• The advent of noninvasive structural imaging (CT,
  MRI, etc.) has made possible the acquisition of
  much larger datasets.
• This makes it important to establish quantitative
  methods for making inferences about the
  relationship between lesions and the symptoms
  they produce.

4
Voxel-based lesion-symptom mapping (VLSM)

• VLSM (Bates et al., 2003)exploits continuous
  behavioraland continuous lesioninformation.
• The method does not requirepatients to be
  grouped either bylesion site or by
  arbitrarybehavioral cutoffs.
• Instead, statistical analyses of the relationship
  between tissue damage and behavior are carried
  out on a voxel-by-voxel basis.

5
Previous approaches

• Groups defined by behavior
• Groups defined by lesion
• Voxel-based approaches (Adolphs et al., 1996,
  2000)

6
Groups defined by behavior

• Patients are divided into groups according to
  whether or not they exhibit a particular
  behavioral deficit (e.g. apraxia of speech in
  Dronkers, 1996 aphasic syndromes in Kertesz,
  1979).
• The lesions of the impaired patients are
  overlaid to determine if there is a common locus
  of infarction.
• An overlay of the sparedpatients lesions can
  also bemade to confirm that theidentified area
  is spared.

7
Groups defined by behavior (2)

• The main limitation of this approach is that it
  forces a yes/no decision to be made about whether
  a patient is impaired
• In situations where deficits are not binary, an
  arbitrary cutoff must be stipulated and
  information about varying degrees of performance
  is lost.

8
Groups defined by lesion

• Patients are divided into groups based on broad
  lesion locations (e.g. dorsolateral prefrontal
  cortex in Chao Knight, 1998).
• Behavioral measures of interest are then compared
  to other groups or to controls.
• This approach therefore dichotomizes the data not
  based on behavior but based on lesion. Again,
  potentially valuable information about patients
  lesion locations cannot enter into the analysis.

9
VLSM methods

• VLSM is fully continuous in both the brain and
  behavior domains.
• In this study, we analyzed speech fluency and
  language comprehension data for 101
  left-hemisphere-damaged stroke patients.
• Behavioral measures consisted of subscales from
  the Western Aphasia Battery (Kertesz, 1979).
  Fluency scores reflect a combination of
  articulatory, word finding, and sentence
  production skills, while the auditory
  comprehension measure represents the average
  score on yes/no questions, single word
  recognition, and enactment of 1, 2 and 3-part
  commands.
• Patients lesions were reconstructed onto
  templates bya board-certified neurologist (RTK)
  with expertise inbehavioral neurology but
  blinded to the clinical statusof the patient.
• VLSM algorithms were programmed in MATLAB andare
  freely available online at http//crl.ucsd.edu/vls
  m.

10
Voxel-by-voxel statistics

• At each voxel, patients are divided into two
  groups according to whether their lesions do or
  do not include that voxel.
• These groups are thencompared (e.g. with a
  t-test)and the resultant statisticsare
  displayed as color maps.
• So this is a massive univariateapproach similar
  to neuro-imaging analysis and subjectto similar
  challenges

11
Results for fluency and comprehension
12
Anterior insula for fluency

• The anterior insula has also been implicated as
  important for fluency in several neuroimaging
  studies (Wise et al., 1999 Blank et al., 2002).

13
MTG for comprehension

• Classic Wernickes area is the posterior STG. But
  recent neuroimaging studies (e.g. Scott et al.,
  2000 Davis Johnsrude, 2003) agree with VLSM
  that the STS and MTG are more important for
  comprehension.

14
Comparing VLSM and fMRI

• This data is from the exact same experiment (2AFC
  environmental sound to picture matching) on
  patients (Saygin et al., 2003) and fMRI of
  controls (Dick et al., 2003).

15
Basic methodological issues

• Differential power in different regions
• Which statistic to plot? (t, d, mean of lesioned
  patients, p, )
• Multiple comparisons
• Choice of scale
• Slice selection

16
Distribution of lesions

• Most aphasic stroke patients have MCA lesions, a
  few have PCA, ACA or other.
• This means that power is greatest in perisylvian
  regions.

17
Plotting effect size

• d difference in means divided by pooled
  standard deviation.
• For large samples, results are similar to t.

18
Multiple comparisons

• A pervasive problem in imaging analysis.
• Bonferroni correction is too conservative.
• False discovery rate is a useful procedure for
  VLSM.
• Proposed by Benjamani Hochberg (1995).
• Controls the expected proportion of false
  positives at a stated alpha level.
• Spatial autocorrelation is still a problem.
• Best not to attach too much importance to
  significance.

19
Correlated lesions

• As is always the case in lesion studies, an area
  may emerge as relevant because it plays a direct
  causal role, or because of a diaschitic effect
  involving highly correlated lesions some distance
  away.
• Could the insula be emerging just because it is
  close to Brocas area? Could the MTG be
  identified just because it is adjacent to
  Wernickes area?
• VLSM allows questions such as these to be
  addressed.

20
Approaches

• If one had an enormous number of patients (say
  1000), then it might be possible to treat dozens
  of ROIs and interactions between them as factors
  in a big ANOVA.
• In the real world, we can ask about the
  potentially confounding effects of single voxels
  (or regions).
• At each voxel, an ANCOVA is performed where the
  lesion status of a voxel of interest is covaried
  out.

21
ANCOVAs

• We selected four voxels-of-interest based on
  anatomical criteria in the centers of
• Brocas area
• the anterior insula
• Wernickes area (posterior STG)
• MTG
• Four maps were then constructed by performing an
  ANCOVA at each voxel covarying out the lesion
  status (lesioned vs. intact) of the reference
  voxel-of-interest.

22
ANCOVA results
23
Comparing maps

• It can be useful toquantitativelycompare VLSM
  maps.
• One way of doing thisis to run a regressionwith
  voxels as subjects.
• For fluency vs. comprehension, a correlation
  value of r 0.59 was obtained, because
  perisylvian regions are more important than
  peripheral areas in both domains.

24
Sample size

• How many patients does it take to make reliable
  inferences?
• Few neuropsychological studies have more than
  about 20 patients, and the mode number of
  patients is of course 1!
• With a metric for comparing maps (correlation),
  we can use repeated sampling of our 101 patients
  to discover how similar two maps of the same
  measure tend to be, as a function of sample size.
• 10 pairs of maps were made each for sample sizes
  ranging from 10 to 50.

25
Sample size

• Increasing reliability is roughly linear.
• Hard to interpret r2, need to look at maps to
  get a sense of similarity.

26
Example of a pair of images

• N 25, r2 0.3

27
More applications

• Grammaticality judgment
• Auditory word recognition
• Environmental sounds
• Pantomime comprehension
• Biological motion perception
• WCST
• Complex grammatical deficits (CYCLE)

28
Grammaticality judgment
29
A role for anterior areas in comprehension
30
ANCOVA factoring out MTG
31
Environmental sounds
32
Summary

• VLSM is an improvement on previous lesion-symptom
  mapping techniques because it uses all available
  information, eliminating reliance on cutoff
  scores, clinical diagnoses, or specified regions
  of interest.
• VLSM makes use of graphic formats and analytical
  methods closely related to those used in
  functional neuroimaging analysis, facilitating
  quantitative comparisons between lesion and
  imaging results.

33
References

• VLSM website http//crl.ucsd.edu/saygin/vlsmpape
  rs.html
• To download software http//crl.ucsd.edu/vlsm
• Bates, E., Wilson, S. M., Saygin, A. P., Dick,
  F., Sereno, M. I., Knight, R. T. Dronkers, N.
  F. (2003). Voxel-based lesion-symptom mapping.
  Nature Neuroscience, 6, 448-450.
• Rorden, C. Karnath, H.-O. (2004). Using human
  brain lesions to infer function a relic from a
  past era in the fMRI age? Nature Reviews
  Neuroscience, 5, 813-819.

34
Tutorial

• Using MATLAB
• Lesion file formats
• Specifying the patient set (vlsm_openpatientset)
• The behavioral data file (vlsm_openbehavdata)
• Creating statistical maps (vlsm_overlay)
• Displaying statistical maps (vlsm_display)
• Exporting figures (vlsm_export)