Usage: 3dNwarpApply [options]
Program to apply a nonlinear 3D warp saved from 3dQwarp (or 3dNwarpCat, etc.)
to a 3D dataset, to produce a warped version of the source dataset.
The '-nwarp' and '-source' options are MANDATORY. For both of these options,
as well as '-prefix', the input arguments after the option name are applied up
until an argument starts with the '-' character, or until the arguments run out.
This program has been heavily modified [01 Dec 2014], including the following
major improvements:
(1) Allow catenation of warps with different grid spacings -- the functions
that deal with the '-nwarp' option will automatically deal with the grids.
(2) Allow input of affine warps with multiple time points, so that 3D+time
datasets can be warped with a time dependent '-nwarp' list.
(3) Allow input of multiple source datasets, so that several datasets can be
warped the same way at once. This operation is more efficient than running
3dNwarpApply several times, since the auto-regridding and auto-catenation
in '-nwarp' will only have to be done once.
* Specification of the output dataset names can be done via multiple
arguments to the '-prefix' option, or via the new '-suffix' option.
New Feature [28 Mar 2018]:
(4) If a source dataset contains complex numbers, then 3dNwarpApply will warp
the real and imaginary parts separately, combine them, and produce a
complex-valued dataset as output.
* Previously, the program would have warped the magnitude of the input
dataset and written out a float-valued dataset.
* No special option is needed to warp complex-valued datasets.
* If you WANT to warp the magnitude of a complex-valued dataset, you will
have to convert the dataset to a float dataset via 3dcalc, then use
3dNwarpApply on THAT dataset instead.
* You cannot use option '-short' with complex-valued source datasets!
More precisely, you can try to use this option, but it will be ignored.
* This ability is added for those of you who deal with complex-valued
EPI datasets (I'm looking at YOU, O International Man of Mystery).
OPTIONS:
--------
-nwarp www = 'www' is the name of the 3D warp dataset
(this is a mandatory option!)
++ Multiple warps can be catenated here.
-->> Please see the lengthier discussion below on this feature!
-->> Also see the help for 3dNwarpCat for some more information
on the formats allowed for inputting warp fields; for
example, warping in one direction only (e.g., 'AP') is
possible.
++ NOTE WELL: The interpretation of this option has changed somewhat,
as of 01 Dec 2014. In particular, this option is
generalized from the version in other programs, including
3dNwarpCat, 3dNwarpFuncs, and 3dNwarpXYZ. The major
change is that multi-line matrix files are allowed to
be included in the 'www' mixture, so that the nonlinear
warp being calculated can be time-dependent.
In addition, the warps supplied need not all be on the
same 3D grid -- this ability lets you catenate a warp
defined on the EPI data grid with a warp defined on the
structural data grid (e.g.).
-iwarp = After the warp specified in '-nwarp' is computed,
invert it. If the input warp would take a dataset
from space A to B, then the inverted warp will do
the reverse.
++ The combination "-iwarp -nwarp 'A B C'" is equivalent
to "-nwarp 'INV(C) INV(B) INV(A)'" -- that is, inverting
each warp/matrix in the list *and* reversing their order.
++ The '-iwarp' option is provided for convenience, and
may prove to be very slow for time-dependent '-nwarp' inputs.
-affter aaa = *** THIS OPTION IS NO LONGER AVAILABLE ***
See the discussion of the new '-nwarp' option above to see
how to do include time-dependent matrix transformations
in this program.
-source sss = 'sss' is the name of the source dataset.
++ That is, the dataset to be warped.
++ Multiple datasets can be supplied here; they MUST
all be defined over the same 3D grid.
-->>** You can no longer simply supply the source
dataset as the last argument on the command line.
-master mmm = 'mmm is the name of the master dataset.
++ Which defines the output grid.
++ If '-master' is not used, then output
grid is the same as the source dataset grid.
++ It is often the case that it makes more sense to
use the '-nwarp' dataset as the master, since
that is the grid on which the transformation is
defined, and is (usually) the grid to which the
transformation 'pulls' the source data.
++ You can use '-master WARP' or '-master NWARP'
for this purpose -- but ONLY if all the warps listed
in the '-nwarp' option have the same 3D grid structure.
++ In particular, if the transformation includes a
long-distance translation, then the source dataset
grid may not have a lot of overlap with the source
dataset after it is transformed -- in this case, you
really want to use this '-master' option -- or you
will end up cutting of a lot of the output dataset
since it will not overlap with the source dataset.
-newgrid dd = 'dd' is the new grid spacing (cubical voxels, in mm)
*OR = ++ This lets you resize the master dataset grid spacing.
-dxyz dd = for example, to bring EPI data to a 1 mm template, but at
a coarser resolution, use '-dxyz 2'.
++ The same grid orientation as the source is used if
the '-master' option is not given.
-interp iii = 'iii' is the interpolation mode
++ Default interpolation mode is 'wsinc5' (slowest, bestest)
++ Available modes are the same as in 3dAllineate:
NN linear cubic quintic wsinc5
++ The same interpolation mode is used for the warp
itself (if needed) and then for the data being warped.
++ The warp will be interpolated if the output dataset is
not on the same 3D grid as the warp itself, or if a warp
expression is used in the '-nwarp' option. Otherwise,
it won't need to be interpolated.
-ainterp jjj = This option lets you specify a different interpolation mode
for the data than might be used for the warp.
++ In particular, '-ainterp NN' would be most logical for
atlas datasets, where the data values being mapped are
integer labels.
-prefix ppp = 'ppp' is the name of the new output dataset
++ If more than 1 source dataset is supplied, then you
should supply more than one prefix. Otherwise, the
program will invent prefixes for each output, by
attaching the suffix '_Nwarp' to each source
dataset's prefix.
-suffix sss = If the program generates prefixes, you can change the
default '_Nwarp' suffix to whatever you want (within
reason) by this option.
++ His Holiness Emperor Zhark defines 'within reason', of course.
++ By using '-suffix' and NOT using '-prefix', the program
will generate prefix names for all output datasets in
a systematic way -- this might be useful for some people.
++ Note that only ONE suffix can be supplied even if many source
datasets are input -- unlike the case with '-prefix'.
-short = Write output dataset using 16-bit short integers, rather than
the usual 32-bit floats.
++ Intermediate values are rounded to the nearest integer.
No scaling is performed.
++ This option is intended for use with '-ainterp' and for
source datasets that contain integral values.
++ If the source dataset is complex-valued, this option will
be ignored.
-wprefix wp = If this option is used, then every warp generated in the process
of application will be saved to a 3D dataset with prefix 'wp_XXXX',
where XXXX is the index of the sub-brick being created.
For example, '-wprefix Zork.nii' will create datasets with names
'Zork_0000.nii', et cetera.
-quiet = Don't be verbose :-(
-verb = Be extra verbose :-)
SPECIFYING THE NONLINEAR WARP IN '-nwarp'
[If you are catenating warps, read this carefully!]
---------------------------------------------------
A single nonlinear warp (usually created by 3dQwarp) is an AFNI or NIfTI-1
dataset with 3 sub-bricks, holding the 3D displacements of each voxel.
(All coordinates and displacements are expressed in DICOM order.)
The '-nwarp' option is used to specify the nonlinear transformation used
to create the output dataset from the source dataset. For many purposes,
the only input needed here is the name of a single dataset holding the
warp to be used.
However, the '-nwarp' option also allows the catenation of a sequence of
spatial transformations (in short, 'warps') that will be combined before
being applied to the source dataset. Each warp is either a nonlinear
warp dataset or a matrix warp (a linear transformation of space).
A single affine (or linear) warp is a set of 12 numbers, defining a 3x4 matrix
a11 a12 a13 a14
a21 a22 a23 a24
a31 a32 a33 a34
A matrix is stored on a single line, in a file with the extension
'.1D' or '.txt', in this order
a11 a12 a13 a14 a21 a22 a23 a24 a31 a32 a33 a34
For example, the identity matrix is given by
1 0 0 0 0 1 0 0 0 0 1 0
This format is output by the '-1Dmatrix_save' options in 3dvolreg and
3dAllineate, for example.
If the argument 'www' following '-nwarp' is made up of more than one warp
filename, separated by blanks, then the nonlinear warp to be used is
composed on the fly as needed to transform the source dataset. For
example,
-nwarp 'AA_WARP.nii BB.aff12.1D CC_WARP.nii'
specifies 3 spatial transformations, call them A(x), B(x), and C(x) --
where B(x) is just the 3-vector x multiplied into the matrix in the
BB.aff12.1D file. The resulting nonlinear warp function N(x) is
obtained by applying these transformations in the order given, A(x) first:
N(x) = C( B( A(x) ) )
That is, the first warp A is applied to the output grid coordinate x,
then the second warp B to that results, then the third warp C. The output
coordinate y = C(B(A(x))) is the coordinate in the source dataset at which
the output value will be interpolated (for the voxel at coordinate x).
The Proper Order of Catenated Warps:
....................................
To determine the correct order in which to input the warps, it is necessary
to understand what a warp of the source dataset actually computes. Call the
source image S(x) = (scalar) value of source image at voxel location x.
For each x in the output grid, the warped result is S(N(x)) -- that is,
N(x) tells where each output location x must be warped to in order to
find the corresponding value of the source S.
N(x) does *NOT* tell to where an x in the source image must be moved to in
the output space -- which is what you might think if you mentally prioritize
the idea of 'warping the source image' or 'pushing the source image' -- DO NOT
THINK THIS WAY! It is better to think of N(x) as reaching out from x in the
output space to a location in the source space and then the program will
interpolate from the discrete source space grid at that location -- which
is unlikely to be exactly on a grid node. Another way to think of this is
that the warp 'pulls' the source image back to the coordinate system on which
the warp is defined.
Now suppose the sequence of operations on an EPI dataset is
(1) Nonlinearly unwarp the dataset via warp AA_WARP.nii (perhaps
from 3dQwarp -plusminus).
(2) Perform linear volume registration on the result from (1) (with
program 3dvolreg) to get affine matrix file BB.aff12.1D -- which
will have 1 line per time point in the EPI dataset.
(3) Linearly register the structural volume to the EPI dataset
(via script align_epi_anat.py). Note that this step transforms
the structural volume to match the EPI, not the EPI to match the
structural volume, so this step does not affect the chain of
transformations being applied to the EPI dataset.
(4) Nonlinearly warp the structural image from (3) to MNI space via
warp CC_WARP.nii (generated by 3dQwarp).
Finally, the goal is to take the original EPI time series dataset, and
warp it directly to MNI space, including the time series registration for
each sub-brick in the dataset, with only one interplation being used --
rather than the 3 interpolations that would come by serially implementing
steps (1), (2), and (4). This one-big-step transformation can be done
with 3dNwarpApply using the '-nwarp' option:
-nwarp 'CC_WARP.nii BB.aff12.1D AA_WARP.nii'
that is, N(x) = A( B( C(x) ) ) -- the opposite order to the sample above,
and with the transformations occurring in the opposite order to the sequence
in which they were calculated. The reason for this apparent backwardness
is that the 'x' being transformed is on the output grid -- in this case, in
MNI-template space. So the warp C(x) transforms such an output grid 'x' to
the EPI-aligned structural space. The warp B(x) then transforms THAT
coordinate from aligned spaced back to the rotated head position of the subject.
And the warp A(x) transforms THAT coordinate back to the original grid that had
to be unwarped (e.g., from susceptibility and/or eddy current artifacts).
Also note that in step (2), the matrix file BB.aff12.1D has one line for
each time point. When transforming a source dataset, the i-th time point
will be transformed by the warp computed using the i-th line from any
multi-line matrix file in the '-nwarp' specification. (If there are more
dataset time points than matrix lines, then the last line will be reused.)
In this way, 3dNwarpApply can be used to carry out time-dependent warping
of time-dependent datasets, provided that the time-dependence in the warp
only occurs in the affine (matrix) parts of the transformation.
Note that the now-obsolete option '-affter' is subsumed into the new way
that '-nwarp' works. Formerly, the only time-dependent matrix had to
be specified as being at the end of the warp chain, and was given via
the '-affter' option. Now, a time-dependent matrix (or more than one)
can appear anywhere in the warp chain, so there is no need for a special
option. If you DID use '-affter', you will have to alter your script
simply by putting the final matrix filename at the end of the '-nwarp'
chain. (If this seems too hard, please consider another line of work.)
The other 3dNwarp* programs that take the '-nwarp' option operate similarly,
but do NOT allow time-dependent matrix files. Those programs are built to
operate with one nonlinear warp, so allowing a time-dependent warp doesn't
make sense for them.
NOTE: If a matrix is NOT time-dependent (just a single set of 12 numbers),
it can be input in the .Xat.1D format of 3 rows, each with 4 values:
a11 a12 a13 a14 } 1 0 0 0
a21 a22 a23 a24 } e.g, identity matrix = 0 1 0 0
a31 a32 a33 a34 } 0 0 1 0
This option is just for convenience. Remember that the coordinates
are DICOM order, and if your matrix comes from Some other PrograM
or from a Fine Software Library, you probably have to change some
signs in the matrix to get things to work correctly.
RANDOM NOTES:
-------------
* At present, this program doesn't work with 2D warps, only with 3D.
(That is, each warp dataset must have 3 sub-bricks.)
* At present, the output dataset is stored in float format, no matter what
absurd data format the input dataset uses (but cf. the '-short' option).
* As described above, 3dNwarpApply allows you to catenate warps directly on
the command line, as if you used 3dNwarpCat before running 3dNwarpApply.
For example:
++ You have aligned dataset Fred+orig to MNI-affine space using @auto_tlrc,
giving matrix file Fred.Xaff12.1D
++ Then you further aligned from MNI-affine to MNI-qwarp via 3dQwarp,
giving warp dataset Fred_WARP+tlrc
++ You can combine the transformations and interpolate Fred+orig directly
to MNI-qwarp space using a command like
3dNwarpApply -prefix Fred_final \
-source Fred+orig \
-master NWARP \
-nwarp 'Fred_WARP+tlrc Fred.Xaff12.1D'
Note the warps to be catenated are enclosed in quotes to make a single
input argument passed to the program. The processing used for this
purpose is the same as in 3dNwarpCat -- see the help output for that
program for a little more information.
++ When you specify a nonlinear warp dataset, you can use the 'SQRT()' and
'INV()' and 'INVSQRT()' operators, as well as the various 1D-to-3D
displacement prefixes ('AP:' 'RL:' 'IS:' 'VEC:', as well as 'FAC:') --
for example, the following is a legal (and even useful) definition of a
warp herein:
'SQRT(AP:epi_BU_yWARP+orig)'
where the 'AP:' transforms the y-displacements in epi_BU_ywarp+orig to a
full 3D warp (with x- and z-displacments set to zero), then calculates the
square root of that warp, then applies the result to some input dataset.
+ This is a real example, where the y-displacement-only warp is computed between
blip-up and blip-down EPI datasets, and then the SQRT warp is applied to
warp them into the 'intermediate location' which should be better aligned
with the subject's anatomical datasets.
-->+ However: see also the '-plusminus' option for 3dQwarp for another way to
reach the same goal, as well as the unWarpEPI.py script.
+ See the output of 3dNwarpCat -help for a little more information on the
1D-to-3D warp prefixes ('AP:' 'RL:' 'IS:' 'VEC:').
++ You can scale the displacements in a 3D warp file via the 'FAC:' prefix, as in
FAC:0.6,0.4,-0.2:fred_WARP.nii
which will scale the x-displacements by 0.6, the y-displacements by 0.4, and
the z-displacments by -0.2.
+ So if you need to reverse the sign of x- and y-displacments, since in AFNI
+x=Left and +y=Posterior while another package uses +x=Right and +y=Anterior,
you could use 'FAC:-1,-1,1:Warpdatasetname' to apply a warp from that
other software package.
++ You can also use 'IDENT(dataset)' to define a "nonlinear" 3D warp whose
grid is defined by the dataset header -- nothing else from the dataset will
be used. This warp will be filled with all zero displacements, which represents
the identity warp. The purpose of such an object is to let you apply a pure
affine warp -- since this program requires a '-nwarp' option, you can use
-nwarp 'IDENT(dataset)' to define the 3D grid for the 'nonlinear' 3D warp and
then catenate the affine warp.
* PLEASE note that if you use the '-allineate' option in 3dQwarp, then the affine
warp is already included in the output nonlinear warp from 3dQwarp, and so it
does NOT need to be applied again in 3dNwarpApply! This mistake has been made
in the past, and the results were not good.
* When using '-allineate' in 3dQwarp, and when there is a large coordinate shift
between the base and source datasets, then the _WARP dataset output by 3dQwarp
will cover a huge grid to encompass both the base and source. In turn, this
can cause 3dNwarpApply to need a lot of memory when it applies that warp.
++ Some changes were made [Jan 2019] to reduce the size of this problem,
but it still exists.
++ We have seen this most often in source datasets which have the (0,0,0)
point not in the middle of the volume, but at a corner of the volume.
Since template datasets (such as MNI152_2009_template_SSW.nii.gz) have
(0,0,0) inside the brain, a dataset with (0,0,0) at a corner of the 3D
volume will need a giant coordinate shift to match the template dataset.
And in turn, the encompassing grid that overlaps the source and template
(base) datasets will be huge.
++ The simplest way to fix this problem is to do something like
@Align_Centers -base MNI152_2009_template_SSW.nii.gz -dset Fred.nii
which will produce dataset Fred_shft.nii, that will have its grid
center approximately lined up with the template (base) dataset.
And from then on, use Fred_shft.nii as your input dataset.
=========================================================================
* This binary version of 3dNwarpApply is compiled using OpenMP, a semi-
automatic parallelizer software toolkit, which splits the work across
multiple CPUs/cores on the same shared memory computer.
* OpenMP is NOT like MPI -- it does not work with CPUs connected only
by a network (e.g., OpenMP doesn't work across cluster nodes).
* For some implementation and compilation details, please see
https://afni.nimh.nih.gov/pub/dist/doc/misc/OpenMP.html
* The number of CPU threads used will default to the maximum number on
your system. You can control this value by setting environment variable
OMP_NUM_THREADS to some smaller value (including 1).
* Un-setting OMP_NUM_THREADS resets OpenMP back to its default state of
using all CPUs available.
++ However, on some systems, it seems to be necessary to set variable
OMP_NUM_THREADS explicitly, or you only get one CPU.
++ On other systems with many CPUS, you probably want to limit the CPU
count, since using more than (say) 16 threads is probably useless.
* You must set OMP_NUM_THREADS in the shell BEFORE running the program,
since OpenMP queries this variable BEFORE the program actually starts.
++ You can't usefully set this variable in your ~/.afnirc file or on the
command line with the '-D' option.
* How many threads are useful? That varies with the program, and how well
it was coded. You'll have to experiment on your own systems!
* The number of CPUs on this particular computer system is ...... 1.
* The maximum number of CPUs that will be used is now set to .... 1.
=========================================================================
++ Compile date = Dec 17 2024 {AFNI_24.3.10:linux_ubuntu_24_64}
