It is a number that has been empirically set through usage in many different types of applications. For instance, for many major bundles (where the seed area comprises of many voxels, e.g. see AutoPtx scripts
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/AutoPtx), such a number is high enough to get reasonably converged results. In many cases, we see that even a smaller number (e.g. 1000 samples) is large enough.
Obviously the choice is a tradeoff between convergence and computation time, so if you have any reason to believe that in a particular case such number is not large enough, you can try e.g. with 10,000 samples and see if the probabilities/relative contrasts
change substantially.