I would like to better understand the projection methods used when one wants to project activations on a reconstructed or inflated cortical surface.
This is a central point for the correctness of the displayed results.
What does 'moyenne corrigée' and 'moyenne rehaussée' (in french in my version) exactly mean ?
Generally, I have a important lost of activation after projection. Is there any mask used on the data before projection ? Should data be in a specific range before projection ?
Is there any document available on the projection methods used ?
Thanks for your help
Michel
projection methods
-
riviere
- Site Admin
- Posts: 1361
- Joined: Tue Jan 06, 2004 12:21 pm
- Location: CEA NeuroSpin, Saint Aubin, France
- Contact:
Hi,
There seem to be no documentation at all on these fusion methods, you are right...
The projections methods use 2 distinct sets of modes: the way information is searched in the functional volume (geometrical sampling), and the "mixing" mode of information coming from these different points. They are set from the "fusion/control 3D fusion" menu of the 3D fusion object.
Geometrical sampling ("Method" parameter):
- point to point: the simplest: only the information coming from the voxel directly under the mesh vertex is used, directly
- point to point with depth offset: only one voxel is taken into account, but its position is shifted along the normal to the mesh (either inside the mesh or outside), for each mesh vertex
- line to point: information is taken along the normal line, both inside and outside, with a sampling (depth and step) specified by appropriate parameters
- inside/ouytside line to point: like line to point but only on one side of the mesh (still along the normal)
- sphere to point: a sampling into a sphere (depth and step parameters apply) is used to get locations in the 3D volume.
Mixing ("Submethod" parameter):
This only applies to sampling methods that are not single-voxel (point to point methods)
- max: the maximum value of all voxels of the functional volume at the sampled locations is mapped on the mesh
- min: the minimum is used
- mean: standard mean (sum of values divided by the number of locations)
- corrected mean / enhanced mean: only non-nul values are taken into account in the mean computation: this is more suitable for thresholded activation maps for instance to avoid blurring the mapped values.
- in the enhanced mean variant, a strange weighting of the final value is applied depending on the proportion of null values in the set of mixed values. I don't know exactly what it is supposed to do, it's a very old code that has survived through ages.
In any case, no specific range in the functional values is expected, and the only "masking" that is performed is the geomterical sampling method that takes values from voxels at specific locations.
You also have to be aware that all this is only a visualization toy and is not very robust: no real interpolation of the functional values is performed to get a continuous activation map along the mesh: especially the methods taking points along normals can produce inaccurate results on high curvature regions (produce discontinuities, map the same voxel value on several vertices etc). The sphere mode is more robust but involves an averaging (blurring) effet, and can take values outside the brain or grey matter...
Denis
There seem to be no documentation at all on these fusion methods, you are right...
The projections methods use 2 distinct sets of modes: the way information is searched in the functional volume (geometrical sampling), and the "mixing" mode of information coming from these different points. They are set from the "fusion/control 3D fusion" menu of the 3D fusion object.
Geometrical sampling ("Method" parameter):
- point to point: the simplest: only the information coming from the voxel directly under the mesh vertex is used, directly
- point to point with depth offset: only one voxel is taken into account, but its position is shifted along the normal to the mesh (either inside the mesh or outside), for each mesh vertex
- line to point: information is taken along the normal line, both inside and outside, with a sampling (depth and step) specified by appropriate parameters
- inside/ouytside line to point: like line to point but only on one side of the mesh (still along the normal)
- sphere to point: a sampling into a sphere (depth and step parameters apply) is used to get locations in the 3D volume.
Mixing ("Submethod" parameter):
This only applies to sampling methods that are not single-voxel (point to point methods)
- max: the maximum value of all voxels of the functional volume at the sampled locations is mapped on the mesh
- min: the minimum is used
- mean: standard mean (sum of values divided by the number of locations)
- corrected mean / enhanced mean: only non-nul values are taken into account in the mean computation: this is more suitable for thresholded activation maps for instance to avoid blurring the mapped values.
- in the enhanced mean variant, a strange weighting of the final value is applied depending on the proportion of null values in the set of mixed values. I don't know exactly what it is supposed to do, it's a very old code that has survived through ages.
In any case, no specific range in the functional values is expected, and the only "masking" that is performed is the geomterical sampling method that takes values from voxels at specific locations.
You also have to be aware that all this is only a visualization toy and is not very robust: no real interpolation of the functional values is performed to get a continuous activation map along the mesh: especially the methods taking points along normals can produce inaccurate results on high curvature regions (produce discontinuities, map the same voxel value on several vertices etc). The sphere mode is more robust but involves an averaging (blurring) effet, and can take values outside the brain or grey matter...
Denis