Searchlight spheres exclude edge voxels (< instead of <=)

[alexeperon]

Hi all,

In the function _get_searchlight_neighbors , in searchlight.py, edge voxels are excluded from spheres by using exclusive less than. This means that for small spheres, fewer voxels are selected than you might intuitively expect. For a mask with a radius of 1, for example, only the centre voxel is included. For a radius of 2, this means choosing 27 instead of 33.

This might not be a big practical issue, but seems somewhat unintuitive.

To solve, I suggest changing these lines

x = x[abs(x - cx) < radius]
y = y[abs(y - cy) < radius]
z = z[abs(z - cz) < radius]

return tuple(data[distance <radius].T.tolist())

To the following:

x = x[abs(x - cx) <= radius]
y = y[abs(y - cy) <= radius]
z = z[abs(z - cz) <= radius]

return tuple(data[distance <=radius].T.tolist())

best wishes,
Alex

[JasperVanDenBosch]

hi @alexeperon

Thank you for this input. I think that this might be something that people have varying views on, and we couldn't remove the old option as that would break compatibility / replicability of existing projects.
However I can see how this could be a new option / parameter.

A PR would be very welcome, and alternatively I'll try to add this when I get around to a larger project on the searchlight functionality.

[JonathanShor]

I coincidentally just reviewed this design choice this past week, and felt the current < approach results in more "ball" like spaces. Using <= falls afoul of the aliasing we get when trying to draw a sphere in a grid.

An easy work-around for the current implementation is to simply pass a radius of, say, 3.01 instead of 3.
Perhaps the docs could emphasize that a float is a valid value for the parameter?

[alexeperon]

Thank you both! I think changing the docs to make the (open) circle voxel selection clear and indicating that a float is valid (right now in fact the docs ask for an integer, but a float works!) is a good way to go.

The key issue I think is replicability, both across rsatoolbox versions and different toolboxes.

I tend to agree that exclusion should result in a better shape - if the centre point of a voxel is r away and it is perpendicular to the centre, then most of the voxel's volume would be outside the sphere (this logic could in fact be codified and used for all voxels to get better spheres, but would slow things down a fair bit and only change things for big spheres).