What is lesion to symptom mapping

Lesion-to-Symptom Mapping (LSM) is a method to map the areas of the brain that cause specific symptoms when lesioned.

What do you need?

Typically, you need data from a group of patients with brain lesions (i.e. stroke). For each patient you need some MRI or CT output which you will use to draw a lesion map. The lesion map is just an image (like the picture you take with your phone but in 3D). Using some software and good knowledge of brain anatomy, the lesion is simply drawn on the MRI image (i.e., using ITKsnap, check the video tutorial). On the other hand, each patient must undergo neuropsychological testing to evaluate cognitive performance in certain tasks (i.e., motor, language). These two pieces of information, lesion maps and cognitive performance, are put in relation to each other to find brain areas subserving specific functions.

I have lesion maps and cognitive scores, can I perform LSM?

No. Before you run the analysis you need to bring all the lesions in the same template space. If you don't do this, you cannot be sure that a voxel corresponds to the same brain area in every patient. Brain shapes differ a lot, and the scanner assigns each brain in a virtual space that may not overlap at all between patients. Thus, you need to register the brains of all patients together in a template space.

Since v0.0.0.9004, LESYMAP includes registerLesionToTemplate, a tool that can run template registrations for you. You can read the Registering lesions in template space wiki page to understand the registration pipeline.

How are the analyses done?

There are basically two approaches to perform analysis. The "voxel-wise" approach and the "multivariate" approach.

Voxel-wise LSM

In a voxel-wise lesion to symptom mapping, each voxel is considered on its own. Imagine that the voxel will be lesioned in some subjects, and healthy in others. This information is used to split the patients in two groups, those with a lesion at that voxel, and those with no lesion at that voxel. The behavioral scores of the two groups are compared with a test, typically t-test or (better) Brunner-Munzel tests. The result can be thresholded just like any other null-hypothesis statistical test (i.e., p < 0.05). The process is repeated independently at every voxel, and forms a full statistical map. This is also known as the "massive univariate approach". The biggest problem of this method is the number of tests (typically, hundreds of thousands). There will be inevitably false findings unless the results are corrected for multiple comparisons (i.e., false discovery rate, family-wise error correction, or Bonferroni). Independently of the logic behind correction methods, they all do the same thing: identify a critical statistical score below which the voxels are considered non-significant. The use of multiple comparison helps to avoid spurious findings. However, VLSM still carries an important conceptual pitfall: it considers each voxel as a functional unit that is worth testing on its own. In reality, the lesion of a single voxel has almost no impact on behavior. Deficits arise when multiple voxels are lesioned. Thus the analysis needs to consider which "group of voxels" constitute the functional brain unit that subserves the behavior of interest. The mass-univariate approach is not able to analyse voxel in groups, and, therefore, is severely limited. Moreover, there is evidence that VLSM results might be displaced in other areas of the brain (see Mah 2014, Sperber 2017).

Multivariate LSM

Multivariate analyses detect which group of voxels together contribute to the emergence of the behavioral deficit. Currently, only few tools can perform multivariate analyses, but there is a general agreement that this approach will evolve in the future because it is conceptually better at mapping functional brain areas. One proposed method uses support vector regression (SVR), which considers all voxels at once and identifies those voxels whose combined information form a relationship with the cognitive deficit (Zhang 2015). A rudimentary implementation of SVR is included in LESYMAP, although more testing is required (other Matlab implementations available). The main multivariate method in LESYMAP is sparse canonical correlations for neuroimaging analyses (SCCAN). SCCAN is an extension of a traditional statistical approach (Hotelling, 1936) but is designed to use the spatial information of neuroimaging data. SCCAN works by identifying voxel weights such that the application of weights to all voxels produces new hypothetical scores with maximal correlation with behavioral scores. Obviously, groups of voxels responsible for the deficit end up having larger weights, while other irrelevant voxels have zero weights. We have performed thousands of simulations of different brain-behavior scenarios and found much more accurate mapping with SCCAN compared to VLSM. This comparison constitutes the most thorough investigation of a LSM analysis method to date (see details in paper). Note that the specific use of SCCAN in LESYMAP can be considered a form of sparse regression because there is only a single behavioral score being analyzed. There is no comparison between SCCAN and SVR yet, but it may come soon. SCCAN can be used by selecting method='sccan' in LESYMAP.