FBA: Tracking recommendations

[Lestropie]

Still feels odd to me that the FBA recommendations involve homogeneous seeding, then SIFT. Current recommendation is 20M streamlines, reduced to 2M with SIFT. I did a quick experiment with dynamic seeding, 20M streamlines, and with 5M streamlines retained the fit to the FODs was better than at 2M streamlines with homogeneous seeding. Furthermore, given the RAM reductions for the fixel-fixel connectivity matrix, it might be possible to bump up the final streamline count slightly without severely influencing feasibility.

Could go to something like 15M generated with dynamic seeding, 2.5M retained?

It also may not be too difficult to modify fixelconnectivity to accept SIFT2 weights as input. A bit of templating involved, but no biggie.

[Lestropie]

Another quick test: Dynamic seeding w. 10M streamlines reaches an equivalent SIFT cost function as WM 20M -> 2M when it reaches 2.6M.

So could suggest:

1. Generating 10M using dynamic seeding rather than 20M with WM seeding, still filter to 2M, and have a better fit to the FOD data with fewer total streamlines generated;

2. Generate 10M using dynamic seeding, filter to 2.5M, have a comparable fit to the data but a more dense reconstruction. Memory usage of fixelconnectivity would be increased slightly, but currently in stats_enhancements the requirements are right on the limit of 24GB so if that were to increase slightly it'd still need a 32GB system regardless.

Other options are possible but would require more testing. E.g. Is there a benefit to having more streamlines with which to build the matrix? I don't know to what extent the reproducibility of such was tested historically. In the absence of such I'd probably go with 1.