Just to add to Rick's advice, there are a couple other choices for 'untalairaching.' 3dAllineate and 3dWarp both can invert the transformation of a single affine transformation like the one done by @auto_tlrc. For datasets that have been transformed by manual talairach transformation, you will have to use 3dfractionize for now. Each program has its own advantages. For example, 3dAllineate can transform multiple sub-bricks while 3dWarp is limited to a single sub-brick. Both 3dWarp and 3dAllineate transform the data much faster than 3dfractionize, but do not have the voting feature that is important if you are transforming multiple ROI datasets back to orig space.

Note if you don't know if a dataset has been auto-talairached or manually talairached, you can find that from the header with

3dAttribute WARP_DATA anat+tlrc

Auto-tlrc'ed data will produce 30 numbers while manually talairached data will produce 360 numbers.

3 ways to inverse talairach:

# 3dAllineate
cat_matvec -ONELINE anat+tlrc::WARP_DATA > tlrc.aff12.1D
3dAllineate -1Dmatrix_apply tlrc.aff12.1D -prefix invtlrc3dAl+orig -source anat+tlrc -master anat+orig

# 3dWarp
cat_matvec anat+tlrc::WARP_DATA > tlrc.1D
3dWarp -matvec_out2in tlrc.1D -prefix invtlrc_3dWarp+orig -gridset anat+orig anat+tlrc
3drefit -view orig invtlrc_3dWarp+tlrc.

# 3dfractionize - slow but useful voting option for multiple ROIs
3dfractionize -input anat+tlrc -warp anat+tlrc -preserve -prefix invtlrc_3dfrac -template anat+orig