0) Your interpretation seems good, except for viewing
those voxels that do not make the full_mask cut as being
worthless. There is not such a presumption with the
full_mask dataset.
1) One can apply the full_mask using -mask_apply. But
it is not done by default because we prefer to see what
results come out across the entire volume, rather than
to blind ourselves to something that might be important
(including problematic artifacts).
Also, masks do not align perfectly across subjects, so
it makes some sense to mask at the group level, instead.
2) The TSNR volume is masked because it is created by a
division, and division by small numbers is not generally
a good idea. The TSNR volume is really a diagnostic tool
to show how good the signals at "good voxels" are. It is
not used to find signal dropout, since one can see that
plainly, or with the full_mask dataset.
- rick