tcksample sampling strategy

[Lestropie]

In tcksample, there are two possible ways in which to sample values from an underlying image:

• For each streamline vertex, perform trilinear interpolation upon the image values at that precise point; average these across all vertices in the streamline.

• Map a continuous representation of the discretely-sampled streamline; see Appendix 3. This produces, for each streamline, a set of voxels that are traversed, along with the length of the intersection between the streamline and each voxel. So from there you can perform a weighted mean, with voxels where the streamline intersection is longer contributing more to the mean calculation. (Activate using -precise option)

The latter may be a little more computationally expensive, but it may also be preferable to reduce interpolation effects.

[sina-mansour]

Adding the -precise function increased the total runtime from ~70 seconds to ~270 seconds (~3 extra minutes). I'll keep the -precise for now.

I'll report the exact runtimes from the first 100 UKB instances (using the finalized pipeline #21). We may then decide to revert this change in case the batch execution duration estimate does not meet our preferences.

[Lestropie]

Yes, -precise is usually ~ 3-4 times more expensive when I use it in tckmap. But, if the proposed modification to tcksample were implemented whereby the input image could be 4D and the output would be a matrix with one column per image volume, the determination of streamline-voxel intersections would only need to be computed once; the weighted mean per streamline could then be computed for all metrics since they all reside on the same voxel grid. So one proposal would be to switch to -precise in tcksample if I find the time to implement 4D image support in tcksample.