Lesion-Symptom Mapping Guide
Lesion-Symptom Mapping Guide
Purpose
This skill encodes expert methodological knowledge for lesion-symptom mapping in clinical and cognitive neuroscience. A competent programmer without neuropsychology and neuroimaging training will get this wrong because:
- Lesion distributions are not random. Stroke lesions cluster in the middle cerebral artery (MCA) territory, creating systematic collinearity between brain regions. Standard voxelwise methods confound truly causal regions with regions that are simply co-damaged (Sperber, 2020).
- Lesion volume is a massive confound. Larger lesions produce worse behavioral deficits simply because more tissue is damaged. Any analysis that does not control for total lesion volume will attribute behavioral deficits to whichever regions are most often part of large lesions (DeMarco & Turkeltaub, 2018).
- White matter disconnection matters as much as grey matter damage. A focal lesion can disrupt distant regions via white matter pathway disruption (diaschisis). VLSM misses this entirely; disconnection analysis is needed (Foulon et al., 2018).
- Standard multiple comparison correction is insufficient. Permutation-based FWE correction is required because voxelwise tests are massively non-independent (lesion voxels are spatially correlated), violating assumptions of FDR and parametric corrections (Kimberg et al., 2007).
- Small samples produce false localizations. With fewer than 50 patients, VLSM lacks power and produces unreliable maps that do not replicate (Sperber, 2020).
When to Use This Skill
- Planning a voxel-based lesion-symptom mapping (VLSM) study
- Choosing between VLSM, multivariate, and disconnection-based approaches
- Setting statistical thresholds and correction methods for lesion analyses
- Evaluating whether a published lesion study adequately controlled for confounds
- Implementing lesion segmentation and registration pipelines
- Conducting network-based lesion mapping with normative connectome data
Do NOT use this skill for:
- Radiological lesion diagnosis (requires clinical neuroradiology)
- General fMRI analysis (see
fmri-glm-analysis-guide) - White matter tractography methodology (upstream of this skill)
Research Planning Protocol
Before executing the domain-specific steps below, you MUST:
- State the research question -- What specific question is this analysis/paradigm addressing?
- Justify the method choice -- Why is this approach appropriate? What alternatives were considered?
- Declare expected outcomes -- What results would support vs. refute the hypothesis?
- Note assumptions and limitations -- What does this method assume? Where could it mislead?
- Present the plan to the user and WAIT for confirmation before proceeding.
For detailed methodology guidance, see the research-literacy skill.
⚠️ Verification Notice
This skill was generated by AI from academic literature. All parameters, thresholds, and citations require independent verification before use in research. If you find errors, please open an issue.
Method Selection Decision Tree
What is your research question?
|
+-- "Which brain voxels are associated with a behavioral deficit?"
| |
| +-- N >= 50 patients, continuous outcome
| | --> VLSM (mass-univariate)
| |
| +-- N >= 50 patients, binary outcome
| | --> VLSM with Brunner-Munzel or chi-square
| |
| +-- N >= 100 patients, distributed representations expected
| | --> SVR-LSM (multivariate)
| |
| +-- N < 50 patients
| --> Underpowered for VLSM; consider ROI-based approach
| or case-series descriptive analysis
|
+-- "Which white matter pathways mediate the deficit?"
| --> Disconnection analysis (BCBToolkit, Disconnectome)
| Can supplement VLSM or replace it when tracts are the question
|
+-- "Which brain networks, when lesioned, produce this symptom?"
--> Lesion network mapping (normative connectome)
Maps lesion location to network disruption
Lesion Segmentation
Methods
| Method | Description | Accuracy | Time per patient | Source |
|---|---|---|---|---|
| Manual tracing | Expert traces lesion on each slice | Gold standard | 30--60 min | Brett et al., 2001 |
| Semi-automated (lesion_gnb) | Gaussian naive Bayes on FLAIR | Good for chronic WM lesions | 5--15 min | Pustina et al., 2016 |
| LINDA | Random forest on T1 | Good for chronic stroke | 5--10 min | Pustina et al., 2016 |
| U-net / deep learning | Trained CNNs | Approaching manual accuracy | < 1 min | Kamnitsas et al., 2017 |
Domain judgment: Manual tracing remains the gold standard. Semi-automated methods should always be visually inspected and manually corrected. Never trust a fully automated segmentation without visual QC of every patient (Brett et al., 2001).
Registration to Standard Space
-
Cost function masking (CRITICAL): When registering a lesioned brain to MNI template, the lesion MUST be masked out of the cost function. Without masking, the registration algorithm warps healthy tissue to fill the lesion, distorting the spatial normalization (Brett et al., 2001).
-
Recommended pipeline:
- Segment lesion on native T1 (or FLAIR)
- Create binary lesion mask
- Register T1 to MNI using nonlinear registration (e.g., ANTs SyN) with lesion mask as cost function mask
- Apply the same warp to the lesion mask
- Verify registration quality visually
- Enantiomorphic normalization: For large lesions, replace the lesioned hemisphere with a flipped version of the intact hemisphere before registration. This improves normalization quality for large lesions (Nachev et al., 2008).
Voxel-Based Lesion-Symptom Mapping (VLSM)
Prerequisites and Sample Size
| Requirement | Minimum | Recommended | Source |
|---|---|---|---|
| Sample size | N >= 50 | N >= 100 | Sperber, 2020; Kimberg et al., 2007 |
| Lesion overlap per voxel | >= 10% of sample (or N >= 10) | >= 15% | Kimberg et al., 2007 |
| Behavioral measure | Continuous preferred | -- | Bates et al., 2003 |
Domain judgment: VLSM with N < 50 is severely underpowered. Sperber (2020) showed that with N = 30, VLSM produces maps that fail to replicate and have unacceptably high false discovery rates. If N < 50, consider ROI-based analyses with a priori regions or descriptive lesion overlap approaches.
Statistical Tests
| Test | When to Use | Advantages | Source |
|---|---|---|---|
| t-test | Continuous behavior, binary lesion status per voxel | Simple, widely used | Bates et al., 2003 |
| Brunner-Munzel | Non-normal behavioral data, unequal variances | Robust to non-normality and unequal group sizes | Rorden et al., 2007 |
| Regression | Controlling for covariates (age, lesion volume) | Flexible; includes confound control | Sperber, 2020 |
| Liebermeister | Binary behavioral outcome (impaired/spared) | For binary classification | Rorden et al., 2007 |
Multiple Comparison Correction
| Method | Description | Recommended? | Source |
|---|---|---|---|
| Permutation-based FWE | Permute behavioral scores 5000+ times; threshold at 5th percentile of max statistic | YES -- gold standard | Kimberg et al., 2007 |
| FDR (Benjamini-Hochberg) | Controls false discovery rate | Acceptable alternative but assumes independence | Kimberg et al., 2007 |
| Bonferroni | Divide alpha by number of voxels | Too conservative; almost never detects effects | Expert consensus |
| Uncorrected | No correction | NEVER for publication | Expert consensus |
Domain judgment: Permutation testing is strongly preferred because lesion maps violate the independence assumptions of FDR and parametric corrections. The spatial correlation structure of lesions means neighboring voxels are highly non-independent. Permutation testing implicitly accounts for this correlation structure (Kimberg et al., 2007).
Controlling for Lesion Volume
Lesion volume MUST be controlled. Methods (DeMarco & Turkeltaub, 2018):
- Direct regression: Include total lesion volume as a covariate in the voxelwise regression model
- Behavioral residualization: Regress behavior on lesion volume first; use residuals as the dependent variable in VLSM
- Both yield similar results, but direct regression is preferred for interpretability (DeMarco & Turkeltaub, 2018)
CRITICAL: Failing to control for lesion volume is the single most common error in VLSM studies. Large lesions damage more regions and produce worse deficits, creating a spurious correlation between any frequently-damaged voxel and behavioral impairment (DeMarco & Turkeltaub, 2018).
Multivariate Lesion-Symptom Mapping
SVR-LSM (Zhang et al., 2014)
Support vector regression-based lesion-symptom mapping considers the full lesion pattern simultaneously, addressing the collinearity problem of mass-univariate VLSM.
| Parameter | Recommended Value | Source |
|---|---|---|
| Kernel | Linear | Zhang et al., 2014 |
| C parameter | 30 (default for LSM) | Zhang et al., 2014 |
| Feature reduction | Remove voxels with < 10% lesion overlap | Zhang et al., 2014 |
| Statistical inference | Permutation testing (5000+ permutations) | Zhang et al., 2014 |
Advantages over VLSM (Zhang et al., 2014):
- Considers spatial covariance of lesion damage (voxels are analyzed jointly, not independently)
- Better handles the collinearity created by vascular territories
- Can detect distributed patterns
Limitations:
- Computationally expensive (hours per analysis with permutation testing)
- Requires larger samples (N >= 80--100 recommended) for stable SVR weights
- Interpretation of SVR weight maps is less straightforward than VLSM t-maps
Machine Learning-Based Mapping (MLBM)
Other multivariate approaches (random forests, LASSO regression) can be applied:
- LASSO is useful for variable selection among brain regions
- Random forests handle nonlinear relationships
- All require permutation-based significance testing
- Cross-validation (leave-one-out or k-fold) is mandatory
Disconnection Analysis
Rationale
A focal lesion disrupts not only the damaged tissue but also white matter pathways passing through the lesion, disconnecting distant brain regions (Foulon et al., 2018). VLSM maps only the lesion site; disconnection analysis maps the affected structural connections.
Methods
| Tool | Approach | Data Required | Source |
|---|---|---|---|
| BCBToolkit | Maps lesion to disconnected tracts using normative tractography atlas | Lesion mask in MNI space | Foulon et al., 2018 |
| Disconnectome maps | Pre-computed: for each brain voxel, which tracts are disconnected | Lesion mask in MNI space | Thiebaut de Schotten et al., 2015 |
| Individual tractography | DTI/DWI tractography in each patient | Patient DWI data | Expert consensus |
BCBToolkit Pipeline (Foulon et al., 2018)
- Register lesion mask to MNI space
- Use the normative tractography atlas (from 170 healthy controls) to identify tracts passing through each lesion
- Generate a disconnection probability map: for each voxel outside the lesion, the probability that it is disconnected
- Use disconnection maps (instead of or in addition to lesion masks) as predictors in VLSM-like analyses
Advantages Over VLSM
- Captures remote effects (diaschisis) that VLSM misses entirely
- Accounts for the fact that two lesions in the same voxel can disconnect different pathways depending on their extent
- Particularly important for white matter lesions, where the damaged tissue itself has no gray matter function
Network Lesion Mapping
Lesion Network Mapping (Boes et al., 2015; Fox, 2018)
Maps a lesion to the brain-wide functional network it disrupts, using normative resting-state fMRI data:
- Use the lesion as a seed region in a normative resting-state fMRI connectome (e.g., 1000-subject dataset)
- Compute the functional connectivity map of the lesion location in healthy brains
- The resulting map shows which brain regions are functionally connected to the lesion site
- Across patients, identify the network that is commonly disrupted
Key Considerations
- Normative connectome: Results depend on the quality and size of the normative dataset. Larger datasets (N >= 500 healthy controls) provide more reliable connectivity estimates (Fox, 2018).
- Seed definition: Use the entire lesion as a seed, not just the peak voxel. Consider weighting by lesion probability if lesion masks are probabilistic.
- Validation: Compare network maps to known functional anatomy. The disrupted network should be biologically plausible for the observed deficit.
- Limitations: Assumes that healthy-brain connectivity predicts the effect of a lesion. Does not account for reorganization or compensation (Fox, 2018).
Common Confounds
1. Lesion Volume Correlation
Larger lesions produce worse behavioral deficits and damage more voxels. Without controlling for lesion volume, VLSM maps reflect voxels that are part of large lesions rather than voxels critical for the behavior (DeMarco & Turkeltaub, 2018).
2. Non-Random Lesion Distribution (MCA Territory Bias)
Stroke lesions are not uniformly distributed across the brain. The MCA territory (lateral frontal, temporal, parietal, insular cortex) is disproportionately affected. This means:
- High statistical power in MCA territory, low power elsewhere
- Voxels outside MCA territory may be critical but undetectable
- Collinearity between MCA-territory voxels inflates false positives for non-critical regions (Sperber, 2020)
3. Time Post-Onset
| Phase | Time | Concern | Source |
|---|---|---|---|
| Acute (< 2 weeks) | Edema, diaschisis, penumbra | Lesion extent overestimated; behavior worst | Karnath et al., 2004 |
| Subacute (2 weeks -- 3 months) | Recovery, reorganization | Lesion stabilizing; behavior improving | Expert consensus |
| Chronic (> 3 months) | Stable lesion | Preferred for VLSM; most stable brain-behavior relationship | Sperber, 2020 |
Domain judgment: Chronic-phase data (> 3 months post-onset) is strongly preferred for VLSM because both lesion extent and behavioral deficits have stabilized. Acute-phase data confounds true lesion effects with transient diaschisis and edema (Karnath et al., 2004; Sperber, 2020).
4. Covariates
Always consider controlling for:
- Age: Older patients have worse outcomes independent of lesion
- Education: Affects cognitive test performance
- Time post-onset: If sample includes mixed acute/chronic patients
- Handedness: Affects lateralization of language
- Lesion hemisphere: If analyzing bilateral samples, hemisphere effects must be modeled
Software
| Software | Methods | Language | Source |
|---|---|---|---|
| NiiStat | VLSM, ROI analysis | MATLAB | Rorden et al., 2007 |
| VLSM2 | VLSM with permutation testing | MATLAB | Bates et al., 2003 |
| SVR-LSM toolbox | Multivariate SVR-LSM | MATLAB | Zhang et al., 2014 |
| BCBToolkit | Disconnection analysis, disconnectome | GUI/Python | Foulon et al., 2018 |
| LESYMAP | VLSM, SVR-LSM, SCCAN | R | Pustina et al., 2018 |
| ANTs | Registration with cost function masking | C++/Python | Avants et al., 2011 |
| FSL | Registration, lesion masking | Python/C++ | Jenkinson et al., 2012 |
Common Pitfalls
1. Not Controlling for Lesion Volume
The most frequent and most damaging error. Always include lesion volume as a covariate or use residualized behavioral scores (DeMarco & Turkeltaub, 2018).
2. Too Small a Sample
VLSM with N < 50 produces unreliable maps. With N = 30, false positive rates can exceed 50% in some simulations (Sperber, 2020). If your sample is small, use ROI-based approaches or descriptive methods.
3. Not Using Cost Function Masking During Registration
Registering lesioned brains to template without masking the lesion distorts the normalization, warping healthy tissue into the lesion cavity and misaligning the rest of the brain (Brett et al., 2001).
4. Using Uncorrected or Bonferroni Correction
Uncorrected thresholds produce massive false positives. Bonferroni is too conservative due to spatial correlation. Use permutation-based FWE (Kimberg et al., 2007).
5. Ignoring the Vascular Architecture
Interpreting a VLSM map as showing "the region responsible for function X" ignores that vascular territory collinearity may have driven the result. Consider supplementing with disconnection analysis or using multivariate methods (Sperber, 2020).
6. Mixing Acute and Chronic Patients Without Controlling for Time
Acute patients have larger effective lesions (edema) and worse behavior, confounding time-post-onset with lesion severity. Analyze chronic patients separately or include time as a covariate (Karnath et al., 2004).
Minimum Reporting Checklist
Based on Sperber (2020), DeMarco & Turkeltaub (2018), and Kimberg et al. (2007):
- Sample size and patient demographics (age, sex, education, handedness)
- Etiology (ischemic stroke, hemorrhagic, tumor, etc.) and affected hemisphere(s)
- Time post-onset for each patient (or range)
- Lesion segmentation method (manual, semi-automated, automated) and software
- Registration method and whether cost function masking was used
- Minimum lesion overlap threshold for voxel inclusion
- Statistical test used (t-test, Brunner-Munzel, regression, SVR)
- Multiple comparison correction method and parameters (number of permutations)
- Whether and how lesion volume was controlled
- Behavioral measure(s) and their psychometric properties
- Software and version for all analysis steps
- For disconnection analysis: normative tractography dataset and pipeline details
Key References
- 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(5), 448--450.
- Boes, A. D., Prasad, S., Liu, H., Liu, Q., Pascual-Leone, A., Caviness, V. S., & Fox, M. D. (2015). Network localization of neurological symptoms from focal brain lesions. Brain, 138(10), 3061--3075.
- Brett, M., Leff, A. P., Rorden, C., & Ashburner, J. (2001). Spatial normalization of brain images with focal lesions using cost function masking. NeuroImage, 14(2), 486--500.
- DeMarco, A. T., & Turkeltaub, P. E. (2018). A multivariate lesion symptom mapping toolbox and examination of lesion-volume biases and correction methods in lesion-symptom mapping. Human Brain Mapping, 39(11), 4169--4182.
- Foulon, C., Cerliani, L., Kinkingnehun, S., Levy, R., Rosso, C., Urbanski, M., Volle, E., & Thiebaut de Schotten, M. (2018). Advanced lesion symptom mapping analyses and implementation as BCBtoolkit. GigaScience, 7(3), giy004.
- Fox, M. D. (2018). Mapping symptoms to brain networks with the human connectome. New England Journal of Medicine, 379(23), 2237--2245.
- Kamnitsas, K., Ledig, C., Newcombe, V. F. J., Simpson, J. P., Kane, A. D., Menon, D. K., Rueckert, D., & Glocker, B. (2017). Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Medical Image Analysis, 36, 61--78.
- Karnath, H.-O., Fruhmann Berger, M., Kuker, W., & Rorden, C. (2004). The anatomy of spatial neglect based on voxelwise statistical analysis: A study of 140 patients. Cerebral Cortex, 14(10), 1164--1172.
- Kimberg, D. Y., Coslett, H. B., & Schwartz, M. F. (2007). Power in voxel-based lesion-symptom mapping. Journal of Cognitive Neuroscience, 19(7), 1067--1080.
- Nachev, P., Coulthard, E., Jager, H. R., Kennard, C., & Husain, M. (2008). Enantiomorphic normalization of focally lesioned brains. NeuroImage, 39(3), 1215--1226.
- Pustina, D., Coslett, H. B., Turkeltaub, P. E., Tustison, N., Schwartz, M. F., & Avants, B. (2016). Automated segmentation of chronic stroke lesions using LINDA: Lesion identification with neighborhood data analysis. Human Brain Mapping, 37(4), 1405--1421.
- Rorden, C., Karnath, H.-O., & Bonilha, L. (2007). Improving lesion-symptom mapping. Journal of Cognitive Neuroscience, 19(7), 1081--1088.
- Sperber, C. (2020). Rethinking causality and data complexity in brain lesion-behaviour inference and its implications for lesion-behaviour modelling. Cortex, 126, 49--62.
- Zhang, Y., Kimberg, D. Y., Coslett, H. B., Schwartz, M. F., & Wang, Z. (2014). Multivariate lesion-symptom mapping using support vector regression. Human Brain Mapping, 35(12), 5861--5876.
See references/vlsm-pipeline.md for step-by-step VLSM analysis workflow.
See references/disconnection-guide.md for detailed BCBToolkit and disconnection analysis procedures.