3dQwarp¶
Usage: 3dQwarp [OPTIONS]¶
* Computes a nonlinearly warped version of source_dataset to match base_dataset.
++ Written by Zhark the Warped, so take nothing here too seriously.
++ The detail allowed in the warping is set by the '-minpatch' option.
++ The discrete warp computed herein is a representation of an underlying
piecewise polynomial C1 diffeomorphism.
++ See the OUTLINE OF WARP OPTIMIZATION METHOD section, far below, for details.
* Other AFNI programs in this nonlinear warping collection include:
++ 3dNwarpAdjust = adjust a set of nonlinear warps to remove any mean warp
++ 3dNwarpApply = apply a nonlinear warp to transform a dataset
++ 3dNwarpCat = catenate/compose two or more warps to produce a new warp
++ 3dNwarpFuncs = compute some functions of a nonlinear warp
++ 3dNwarpXYZ = apply a nonlinear warp to discrete set of (x,y,z) triples
++ @SSwarper = Script that combines 3dQwarp and 3dSkullStrip (SS) to
produce a brain volume warped to a template and with
the non-brain tissue ('skull') removed.
++ auto_warp.py = Python program to run 3dQwarp for you
++ unWarpEPI.py = Python program to unwarp EPI datasets, using
a reverse-blip reference volume
++ afni_proc.py = General AFNI pipeline for FMRI datasets, which can use
auto_warp.py and unWarpEPI.py along the way.
* 3dQwarp is where nonlinear warps come from (in AFNIland).
++ For the most part, the above programs either use warps from 3dQwarp,
or they provide easier ways to run 3dQwarp.
** NEVER use the obsolete '-nwarp' option to 3dAllineate. It is not
compatible with these other programs, and it does not produce
useful results.
* The simplest way to use 3dQwarp is via the @SSwarper script, for
warping a T1-weighted dataset to the (human brain) MNI 2009 template
dataset supplied with AFNI binaries (other templates also available).
* The next simplest way to use 3dQwarp is via the auto_warp.py program.
* You can use 3dQwarp directly if you want to control (or play with) the
various options for setting up the warping process.
* Input datasets must be on the same 3D grid (unlike program 3dAllineate)!
++ Or you will get a fatal error when the program checks the datasets!
++ However, you can use the '-allineate' option in 3dQwarp to do
affine alignment before the nonlinear alignment, which will also
resample the aligned source image to the base dataset grid.
++ OR, you can use the '-resample' option in 3dQwarp to resample the
source dataset to the base grid before doing the nonlinear stuff,
without doing any preliminary affine alignment. '-resample' is much
faster than '-allineate', but of course doesn't do anything but
make the spatial grids match. Normally, I would not recommend this!
++ UNLESS the base and source datasets are fairly close to each other ++
++ already, the '-allineate' option will make the process better. For ++
++ example, if the two datasets are rotated off 5 degrees, using ++
++ 3dQwarp alone will be less effective than using '3dQwarp -allineate'. ++
* 3dQwarp CAN be used on 2D images -- that is, datasets with a single
slice. How well it works on such datasets has not been investigated
much, but it DOES work (and quickly, since the amount of data is small).
++ You CAN input .jpg or .png files as the source and base images.
++ 3dQwarp will convert RGB images to grayscale and attempt to align those.
The output will still be in dataset format (not image format) and
will be in grayscale floating point (not color). To get the warped
image output in .jpg or .png format, you can open the output dataset
in the AFNI GUI and save the image -- after turning off crosshairs!
+ To get an RGB copy of a warped image, you have to apply the warp to
each channel (R, G, B) separately and then fuse the results.
Other approaches are possible, of course.
++ Applying this program to 2D images is entirely for fun; the actual
utility of it in brain imaging is not clear to Emperor Zhark.
(Which is why the process of getting a color warped image is so clumsy.)
* Input datasets should be reasonably well aligned already
(e.g., as from an affine warping via 3dAllineate).
++ The standard result from 3dAllineate will resample the affinely
aligned dataset to the same 3D grid as the -base dataset, so this
new dataset will be ready to run in 3dQwarp against the same base.
++ Again, the '-allineate' option can now do this for you, inside 3dQwarp.
* Input datasets should be 'alike'.
++ For example, if the '-base' dataset is skull stripped, then the '-source'
dataset should be skull stripped also -- e.g., via 3dSkullStrip.
+ Warping a skull-on dataset (source) to a skull-off dataset (base) will
sometimes work OK, but sometimes will fail in weird-looking ways.
++ If the datasets have markedly different contrasts (e.g., T1 and T2), then
using a non-standard matching function such as '-nmi' or '-hel' or '-lpa'
might work better than the default Pearson correlation matching function.
(This type of warping has not been tested much here at AFNI headquarters.)
+ Warping T2 to T1 would likely be best done by inverting the contrast of
the T2 dataset, via '3dUnifize -T2 -T2', to make it look like a T1 volume.
+ These non-standard methods are slower than the Pearson correlation default.
* If the input datasets do NOT overlap reasonably well (please look at them *
* them in AFNI), or when the source is in scanner space and the base is in a *
* template space (e.g., MNI), then you need to use '-allineate', or you will *
* probably get *
* (a) a very bad result (or a program crash) *
* (b) that takes a long time and a lot of memory to compute. *
* 'Overlap well' means that the datasets match well in coordinate space. *
* In some cases, datasets may match well voxel-wise, but the xyz coordinates *
* defined in the dataset headers do not match -- in such a case, 3dQwarp *
* will fail. This is why Zhark urges you to LOOK at the overlap in AFNI, *
* which uses coordinates for display matching, not voxel indexes. Or use *
* the '-allineate' option to get 3dAllineate to line up the dataset by *
* brute force, just to be safe (at the cost of a some extra CPU time). *
* Outputs of 3dQwarp are the warped dataset and the warp that did it.
++ These datasets are stored in float format, no matter what the
data type of the source dataset.
++ MANY other optional outputs are described later.
* Simple example:
3dQwarp -allineate -blur 0 3 \
-base ~/abin/MNI152_2009_template_SSW.nii.gz \
-source sub637_T1.nii \
-prefix sub637_T1qw.nii
which will produce a dataset warped to match the MNI152 T1 template
at a 1 mm resolution. Since the MNI152 template is already somewhat
blurry, the amount of blurring applied to it is set to zero, while
the source dataset (presumably not blurry) will be Gaussian blurred
with a FWHM of 3 mm.
* Matching uses the 'clipped Pearson' method by default, and
can be changed to 'pure Pearson' with the '-pear' option.
++ The purpose of 'clipping' is to reduce the impact of outlier values
(small or large) on the correlation.
++ For the adventurous, you can also try these matching functions:
'-hel' for Hellinger distance
'-mi' for Mutual Information
'-nmi' for Normalized Mutual Information
These options have NOT been extensively tested for usefulness,
and should be considered experimental at this infundibulum.
++ The 'local' correlation options are also now available:
'-lpc' for Local Pearson minimization (i.e., EPI-T1 registration)
'-lpa' for Local Pearson maximization (i.e., T1-FA registration)
However, the '+ZZ' modifier is not available for these cost functions,
unlike in program 3dAllineate :(
These advanced cost options will slow 3dQwarp down significantly.
** For aligning EPI to T1, the '-lpc' option can be used; my advice
would be to do something like the following:
3dSkullStrip -input SUBJ_anat+orig -prefix SUBJ_anatSS
3dbucket -prefix SUBJ_epiz SUBJ_epi+orig'[0]'
align_epi_anat.py -anat SUBJ_anat+orig \
-epi SUBJ_epiz+orig -epi_base 0 -partial_axial \
-epi2anat -master_epi SUBJ_anat+orig \
-big_move
3dQwarp -source SUBJ_anatSS+orig.HEAD \
-base SUBJ_epiz_al+orig \
-prefix SUBJ_anatSSQ \
-lpc -maxlev 0 -verb -iwarp -blur 0 3
3dNwarpApply -nwarp SUBJ_anatSSQ_WARPINV+orig \
-source SUBJ_epiz_al+orig \
-prefix SUBJ_epiz_alQ
* Zeroth, the T1 is prepared by skull stripping and the EPI is prepared
by extracting just the 0th sub-brick for registration purposes.
* First, the EPI is aligned to the T1 using the affine 3dAllineate, and
at the same time resampled to the T1 grid (via align_epi_anat.py).
* Second, it is nonlinearly aligned ONLY using the global warping -- it is
futile to try to align such dissimilar image types precisely.
* The EPI is used as the base in 3dQwarp so that it provides the weighting,
and so partial brain coverage (as long as it covers MOST of the brain)
should not cause a problem (fondly do we hope).
* Third, 3dNwarpApply is used to take the inverse warp from 3dQwarp to
transform the EPI to the T1 space, since 3dQwarp transformed the T1 to
EPI space. This inverse warp was output by 3dQwarp using '-iwarp'.
* Someday, this procedure may be incorporated into align_epi_anat.py :-)
** It is vitally important to visually look at the results of this process! **
* In humans, the central structures usually match a template very well,
but smaller cortical gyri can match well in some places and not match
in others.
* In macaques, where there is less inter-animal variation, cortical
matching will be better than humans (but not perfect).
* For aligning T1-weighted anatomical volumes, Zhark recommends that
you use the 3dUnifize program to (approximately) spatially uniformize
and normalize their intensities -- this helps in the matching process,
especially when using datasets from different scanners.
++ Skull stripping a la 3dSkullStrip is also a good idea (prior to 3dUnifize),
even if you are registering datasets from the same subject; see the
SAMPLE USAGE section below for an example.
+ But if you are matching to a skull-on template as the -base dataset,
then keeping the skull on the -source dataset is necessary, since the
goal of the program is to distort the source to 'look like' the base,
and if major parts of the two datasets cannot be made to look like
each other, the poor poor program will get lost in warp-land.
++ If you ultimately want a non-3dUnifize-d transformed dataset, you can use
the output WARP dataset and 3dNwarpApply to transform the un-3dUnifize-d
source dataset; again, see the SAMPLE USAGE section below.
++ Some people prefer to nonlinearly align datasets with the 'skull' left on.
You are free to try this, of course, but we have not tested this method.
+ We give you tools; you build things with them (hopefully nice things).
++ Note again the script @SSwarper, which is for skull-stripping and warping
a T1-weighted dataset to a template volume; AFNI provides such a template
volume for the MNI152 space.
* If for some deranged reason you have datasets with very non-cubical voxels,
they should be resampled to a cubical grid before trying 3dQwarp. For example,
if you have acquired 1x1x4 mm T1-weighted structural volumes (why?), then
resample them to 1x1x1 mm before doing any other registration processing.
For example:
3dAllineate -input anatT1_crude+orig -newgrid 1.0 \
-prefix anatT1_fine -final wsinc5 \
-1Dparam_apply '1D: 12@0'\'
This operation will be done using the '-allineate' or '-resample'
options to 3dQwarp, if the -base dataset has cubical voxels.
** Please note that this program is very CPU intensive, and is what computer
scientists call a 'pig' (i.e., run time from 10s of minutes to hours).
SAMPLE USAGE¶
* For registering a T1-weighted anat to a mildly blurry template at about
a 1x1x1 mm resolution (note that the 3dAllineate step, to give the
preliminary alignment, will also produce a dataset on the same 3D grid
as the TEMPLATE+tlrc dataset, which 3dQwarp requires):
3dUnifize -prefix anatT1_U -input anatT1+orig
3dSkullStrip -input anatT1_U+orig -prefix anatT1_US -niter 400 -ld 40
3dAllineate -prefix anatT1_USA -base TEMPLATE+tlrc \
-source anatT1_US+orig -twopass -cost lpa \
-1Dmatrix_save anatT1_USA.aff12.1D \
-autoweight -fineblur 3 -cmass
3dQwarp -prefix anatT1_USAQ -blur 0 3 \
-base TEMPLATE+tlrc -source anatT1_USA+tlrc
You can then use the anatT1_USAQ_WARP+tlrc dataset to transform other
datasets (that were aligned with the input anatT1+orig) in the same way
using program 3dNwarpApply, as in
3dNwarpApply -nwarp 'anatT1_USAQ_WARPtlrc anatT1_USA.aff12.1D' \
-source NEWSOURCE+orig -prefix NEWSOURCE_warped
For example, if you want a warped copy of the original anatT1+orig dataset
(without the 3dUnifize and 3dSkullStrip modifications), put 'anatT1' in
place of 'NEWSOURCE' in the above command.
Note that the '-nwarp' option to 3dNwarpApply has TWO filenames inside
single quotes. This feature tells that program to compose (catenate) those
2 spatial transformations before applying the resulting warp. See the -help
output of 3dNwarpApply for more sneaky/cunning ways to make the program warp
datasets (and also see the example just below).
** PLEASE NOTE that if you use the '-allineate' option in 3dQwarp, to **
** do the 3dAllineate step inside 3dQwarp, then you do NOT catenate **
** the affine and nonlinear warps as in the 3dNwarpApply example above, **
** since the output nonlinear warp will ALREADY have be catenated with **
** the affine warp -- this output warp is the transformation directly **
** between the '-source' and '-base' datasets (as is reasonable IZHO). **
If the NEWSOURCE+orig dataset is integer-valued (e.g., anatomical labels),
then you would use the '-ainterp NN' with 3dNwarpApply, to keep the program
from interpolating the voxel values.
* If you use align_epi_anat.py to affinely transform several EPI datasets to
match a T1 anat, and then want to nonlinearly warp the EPIs to the template,
following the warp generated above, the procedure is something like this:
align_epi_anat.py -anat anatT1+orig -epi epi_r1+orig \
-epi_base 3 -epi2anat -big_move \
-child_epi epi_r2+orig epi_r3+orig
3dNwarpApply -source epi_r1+orig \
-nwarp 'anatT1_USAQ_WARP+tlrc anatT1_USA.aff12.1D' \
-affter epi_r1_al_reg_mat.aff12.1D \
-master WARP -newgrid 2.0 \
-prefix epi_r1_AQ
(mutatis mutandis for 'child' datasets epi_r2, epi_r3, etc.).
The above procedure transforms the data directly from the un-registered
original epi_r1+orig dataset, catenating the EPI volume registration
transformations (epi_r1_al_reg_mat.aff12.1D) with the affine anat to
template transformation (anatT1_USA.aff12.1D) and with the nonlinear
anat to template transformation (anatT1_USAQ_WARP+tlrc). 3dNwarpApply
will use the default 'wsinc5' interpolation method, which does not blur
the results much -- an important issue for statistical analysis of the
EPI time series.
Various functions, such as volume change fraction (Jacobian determinant)
can be calculated from the warp dataset via program 3dNwarpFuncs.
COMMAND LINE OPTIONS (too many of them)¶
++++++++++ Input and Outputs +++++++++++++
-base base_dataset = Alternative way to specify the base dataset.
-source source_dataset = Alternative way to specify the source dataset.
* You can either use both '-base' and '-source',
OR you can put the base and source dataset
names last on the command line.
* But you cannot use just one of '-base' or '-source'
and then put the other input dataset name at the
end of the command line!
*** Please note that if you are using 3dUnifize on one
dataset (or the template was made with 3dUnifize-d
datasets), then the other dataset should also be
processed the same way for better results. This
dictum applies in general: the source and base
datasets should be pre-processed the same way,
as far as practicable.
-prefix ppp = Sets the prefix for the output datasets.
* The source dataset is warped to match the base
and gets prefix 'ppp'. (Except if '-plusminus' is used.)
* The final interpolation to this output dataset is
done using the 'wsinc5' method. See the output of
3dAllineate -HELP
(in the "Modifying '-final wsinc5'" section) for
the lengthy technical details.
* The 3D warp used is saved in a dataset with
prefix '{prefix}_WARP' -- this dataset can be used
with 3dNwarpApply and 3dNwarpCat, for example.
* To be clear, this is the warp from source dataset
coordinates to base dataset coordinates, where the
values at each base grid point are the xyz displacements
needed to move that grid point's xyz values to the
corresponding xyz values in the source dataset:
base( (x,y,z) + WARP(x,y,z) ) matches source(x,y,z)
Another way to think of this warp is that it 'pulls'
values back from source space to base space.
* 3dNwarpApply would use '{prefix}_WARP' to transform datasets
aligned with the source dataset to be aligned with the
base dataset.
** If you do NOT want this warp saved, use the option '-nowarp'.
-->> (But: This warp is usually the most valuable possible output!)
* If you want to calculate and save the inverse 3D warp,
use the option '-iwarp'. This inverse warp will then be
saved in a dataset with prefix '{prefix}_WARPINV'.
* This inverse warp could be used to transform data from base
space to source space, if you need to do such an operation.
* You can easily compute the inverse later, say by a command like
3dNwarpCat -prefix Z_WARPINV 'INV(Z_WARP+tlrc)'
or the inverse can be computed as needed in 3dNwarpApply, like
3dNwarpApply -nwarp 'INV(Z_WARP+tlrc)' -source Dataset.nii ...
-nowarp = Do not save the _WARP file.
* By default, the {prefix}_WARP dataset will be saved.
-iwarp = Do compute and save the _WARPINV file.
* By default, the {prefix}_WARPINV file is NOT saved.
-nodset = Do not save the warped source dataset (i.e., if you only
need the _WARP).
* By default, the warped source dataset {prefix} is saved.
-awarp = If '-allineate' is used, output the nonlinear warp that
transforms from the 3dAllineate-d affine alignment of
source-to-base to the base. This warp (output {prefix}_AWARP)
combined with the affine transformation {prefix}.aff12.1D is
the same as the final {prefix}_WARP nonlinear transformation
directly from source-to-base.
* The '-awarp' output is mostly useful when you need to have
this incremental nonlinear warp for various purposes; for
example, it is used in the @SSwarper script.
* '-awarp' will not do anything unless '-allineate' is also
used, because it doesn't have anything to do!
* By default, this {prefix}_AWARP file is NOT saved.
-inwarp = This option is for debugging, and is only documented here
for completenes.
* It causes an extra dataset to be written out whenever a warp
is output. This dataset will have the string '_index' added
to the warp dataset's prefix, as in 'Fred_AWARP_index.nii'.
* This extra dataset contains the 'index warp', which is the
internal form of the warp.
* Instead of displacements between (x,y,z) coordinates, an
index warp stores displacements between (i,j,k) 3D indexes.
* An index warp dataset has no function outside of being
something to look at when trying to figure out what the hell
the program did.
++++++++++ Preliminary affine (linear transformation) alignment ++++++++++
-allineate = This option will make 3dQwarp run 3dAllineate first, to align
*OR* the source dataset to the base with an affine transformation.
-allin It will then use that alignment as a starting point for the
*OR* nonlinear warping.
-allinfast * With '-allineate', the source dataset does NOT have to be on
the same 3D grid as the base, since the intermediate output
of 3dAllineate (the substitute source) will be on the grid
as the base.
* If the datasets overlap reasonably already, you can use the
option '-allinfast' (instead of '-allineate') to add the
option '-onepass' to the 3dAllineate command line, to make
it run faster (by avoiding the time-consuming coarse pass
step of trying lots of shifts and rotations to find an idea
of how to start). But you should KNOW that the datasets do
overlap well before using '-allinfast'. (This fast option
does include center-of-mass correction, so it will usually
work well if the orientations of the two volumes are close
-- say, within 10 degrees of each other.)
-->>** The final output warp dataset is the warp directly between
the original source dataset and the base (i.e., the catenation
of the affine matrix from 3dAllineate and the nonlinear warp
from the 'warpomatic' procedure in 3dQwarp).
-->>** The above point means that you should NOT NOT NOT use the
affine warp output by the '-allineate' option in combination
with the nonlinear warp output by 3dQwarp (say, when using
3dNwarpApply), since the affine warp would then be applied
twice -- which would be WRONG WRONG WRONG.
-->>** The final output warped dataset is warped directly from the
original source dataset, NOT from the substitute source.
* The intermediate files from 3dAllineate (the substitute source
dataset and the affine matrix) are saved, using 'prefix_Allin'
in the filenames.
*** The following 3dQwarp options CANNOT be used with -allineate:
-plusminus -inilev -iniwarp
* The '-awarp' option will output the computed warp from the
intermediate 3dAllineate-d dataset to the base dataset,
in case you want that for some reason. This option will
only have meaning if '-allineate' or '-allinfast' is used.
The prefix of the '-awarp' output will have the string
'_AWARP' appended to the {prefix} for the output dataset.
-allineate_opts '-opt ...'
*OR* * This option lets you add extra options to the 3dAllineate
-allopt command to be run by 3dQwarp. Normally, you won't need
to do this.
* Note that the default cost functional in 3dAllineate is
the Hellinger metric ('-hel'); many people prefer '-lpa+ZZ',
and so should use something like this:
-allopt '-cost lpa+ZZ'
to ensure 3dAllineate uses the desired cost functional.
-> Note that if you use '-lpa' in 3dQwarp, then 3dAllineate
will automatically be supplied with '-cost lpa+ZZ'.
* If '-emask' is used in 3dQwarp, the same option will be
passed to 3dAllineate automatically, so you don't have to
do that yourself.
*** Do NOT attempt to use the (obsolescent) '-nwarp' option in
3dAllineate from inside 3dQwarp -- bad things will probably
happen, and you won't EVER get any birthday presents again!
-resample = This option simply resamples the source dataset to match the
*OR* base dataset grid. You can use this if the two datasets
-resample mm overlap well (as seen in the AFNI GUI), but are not on the
same 3D grid.
* If they don't overlap very well, use '-allineate' instead.
* As with -allineate, the final output dataset is warped
directly from the source dataset, not from the resampled
source dataset.
* The reampling here (and with -allineate) is done with the
'wsinc5' method, which has very little blurring artifact.
* If the base and source datasets ARE on the same 3D grid,
then the -resample option will be ignored.
* You CAN use -resample with these 3dQwarp options:
-plusminus -inilev -iniwarp
In particular, '-iniwarp' and '-resample' will work
together if you need to re-start a warp job from the
output of '-allsave'.
* Unless you are in a hurry, '-allineate' is better.
*** After '-resample', you can supply an affine transformation
matrix to apply during the resampling. This feature is
useful if you already have the affine transformation
from source to base pre-computed by some other program
-- for example, from 3dAllineate.
The command line argument that follows '-resample',
if it does not start with a '-', is taken to be
a filename with 12 values in one row: the usual
affine matrix representation from 3dAllineate and
other AFNI programs (in DICOM order coordinates);
for example '-resample ZharkRules.aff12.1D'
You can also use the following form to supply the
matrix directly on the command line:
'1D: 1 2 3 4 5 6 7 8 9 10 11 12'
where the numbers after the initial '1D: ' are
to be replaced by the actual matrix entries!
-aniso = Before aligning, do a little bit of anisotropic smoothing
(see 3danisosmooth) on the source dataset.
* Note that the final output dataset is warped directly
from the input dataset, NOT this smoothed dataset.
If you want the warped output dataset to be from the
smoothed dataset, you'll have to use 3danisosmooth
separately before 3dQwarp, and supply that result
as the source dataset.
* The purpose of '-aniso' is just to smooth out the noise
a little before other processing, and maybe make things
work a little betterer.
* Anisotropic smoothing comes before 3dAllineate, if both
are used together.
++++++++++ Computing the 'cost' functional = how datasets are matched ++++++++++
** If '-allineate' is used, AND one of these options is given, then the **
** corresponding option is also passed to 3dAllineate for its optimization. **
** Otherwise, 3dAllineate will use its default optimization cost functional. **
-pcl = clipped Pearson correlation [default method]; clipping reduces
the impact of outlier values.
-pear = Use strict Pearson correlation for matching.
* Not usually recommended, because without the clipping-ness
used by '-pcl', then outliers can have more effect.
* No partridges or trees are implied by this option.
-hel = Hellinger metric
-mi = Mutual information
-nmi = Normalized mutual information
-lpc = Local Pearson correlation (signed).
-lpa = Local Pearson correlation (absolute value)
* These options mirror those in 3dAllineate.
* In particular, nonlinear warping of low resolution EPI
data to T1 data is a difficult task, and can introduce
more distortions to the result than it fixes.
* If you use one of these 5 options, and also use '-allineate' or
'-allinfast', then the corresponding option is passed to
3dAllineate: '-hel' => '-cost hel'
'-mi' => '-cost mi'
'-nmi' => '-cost nmi'
'-lpc' => '-cost lpc+ZZ'
'-lpa' => '-cost lpa+ZZ'
'-pcl' or -pear' => '-cost ls'
-noneg = Replace negative values in either input volume with 0.
-zclip * If there ARE negative input values, and you do NOT use -noneg,
then strict Pearson correlation will be used, since the
'clipped' method only is implemented for non-negative volumes.
* '-noneg' is not the default, since there might be situations
where you want to align datasets with positive and negative
values mixed.
* But, in many cases, the negative values in a dataset are just
the result of interpolation artifacts (or other peculiarities),
and so they should be ignored. That is what '-noneg' is for.
* Therefore, '-noneg' is recommended for most applications.
-nopenalty = Don't use a penalty on the cost functional; the goal
of the penalty is to reduce grid distortions.
* If there penalty is turned off AND you warp down to
a fine scale (e.g., '-minpatch 11'), you will probably
get strange-looking results.
-penfac ff = Use the number 'ff' to weight the penalty.
The default value is 1. Larger values of 'ff' mean the
penalty counts more, reducing grid distortions,
insha'Allah; '-nopenalty' is the same as '-penfac 0'.
-warpscale f = This option allows you to downsize the scale of the warp
displacements for smaller patch sizes. In some applications,
the amount of displacement allowed is overly aggressive at
small patch sizes, but larger displacements at large patch
sizes are needed to get the overall shapes of the base and
template to match. The factor 'f' should be a number between
0.1 and 1.0 (inclusive), and indicates the amount the max
displacement should shrink when the patch size shrinks by
a factor of 10. I suggest '-warpscale 0.5' as a starting
point for experimentation.
* This option is currently [Feb 2020] for experimenting
only, and in the future it may change! In particular,
the equivalent of '-warpscale 0.5' may become the default.
-useweight = With '-useweight', each voxel in the base automask is weighted
by the intensity of the (blurred) base image. This makes
white matter count more in T1-weighted volumes, for example.
-->>* [24 Mar 2014] This option is is now the default.
-wtgaus G = This option lets you define the amount of Gaussian smoothing
applied to the base image when creating the weight volume.
The default value of G is 4.5 (FWHM voxels). See the 'WEIGHT'
section (far below) for details on how the automatic
weight volume is calculated. Using '-wtgaus 0' means that
no Gaussian blurring is applied in creating the weight.
* [15 Jan 2020] This option is really just for fooling around.
-noweight = If you want a binary weight (the old default), use this option.
That is, each voxel in the base volume automask will be
weighted the same in the computation of the cost functional.
-weight www = Instead of computing the weight from the base dataset,
directly input the weight volume from dataset 'www'.
* Useful if you know what over parts of the base image you
want to emphasize or de-emphasize the matching functional.
-wball x y z r f =
Enhance automatic weight from '-useweight' by a factor
of 1+f*Gaussian(FWHM=r) centered in the base image at
DICOM coordinates (x,y,z) and with radius 'r'. The
goal of this option is to try and make the alignment
better in a specific part of the brain.
* Example: -wball 0 14 6 30 40
to emphasize the thalamic area (in MNI/Talairach space).
* The 'r' parameter must be positive (in mm)!
* The 'f' parameter must be between 1 and 100 (inclusive).
* '-wball' does nothing if you input your own weight
with the '-weight' option :(
* '-wball' does change the binary weight created by
the '-noweight' option.
* You can only use '-wball' once in a run of 3dQwarp.
*** The effect of '-wball' is not dramatic. The example
above makes the average brain image across a collection
of subjects a little sharper in the thalamic area, which
might have some small value. If you care enough about
alignment to use '-wball', then you should examine the
results from 3dQwarp for each subject, to see if the
alignments are good enough for your purposes.
-wmask ws f = Similar to '-wball', but here, you provide a dataset 'ws'
that indicates where to increase the weight.
* The 'ws' dataset must be on the same 3D grid as the base
dataset.
* 'ws' is treated as a mask -- it only matters where it
is nonzero -- otherwise, the values inside are not used.
* After 'ws' comes the factor 'f' by which to increase the
automatically computed weight. Where 'ws' is nonzero,
the weighting will be multiplied by (1+f).
* As with '-wball', the factor 'f' should be between 1 and 100.
* You cannot use '-wball' and '-wmask' together!
-wtprefix p = Saves auto-computed weight volume to a dataset with prefix 'p'.
If you are sufficiently dedicated, you could manually edit
this volume, in the AFNI GUI, in 3dcalc, et cetera. And then
use it, instead of the auto-computed default weight, via the
'-weight' option.
* If you use the '-emask' option, the effects of the exclusion
mask are NOT shown in this output dataset!
-emask ee = Here, 'ee' is a dataset to specify a mask of voxels
to EXCLUDE from the analysis -- all voxels in 'ee'
that are NONZERO will not be used in the alignment.
* The base image is always automasked -- the emask is
extra, to indicate voxels you definitely DON'T want
included in the matching process, even if they are
inside the brain.
-->>* Note that 3dAllineate has the same option. Since you
usually have to use 3dAllineate before 3dQwarp, you
will probably want to use -emask in both programs.
[ Unless, of course, you are using '-allineate', which ]
[ will automatically include '-emask' in the 3dAllineate ]
[ phase if '-emask' is used here in 3dQwarp. ]
* Applications: exclude a tumor or resected region
(e.g., draw a mask in the AFNI Drawing plugin).
-->>* Note that the emask applies to the base dataset,
so if you are registering a pre- and post-surgery
volume, you would probably use the post-surgery
dataset as the base. If you eventually want the
result back in the pre-surgery space, then you
would use the inverse warp afterwards (in 3dNwarpApply).
-inedge = Enhance interior edges in the base and source volumes, to
make the cost functional give more weight to these edges.
* This option MIGHT produce slightly better alignments, but
its effect is usually small.
* The output transformed source dataset will NOT have these
enhanced edges; the enhancement is done internally on the
volume image copies that are being matched.
*** This option has been disabled, until problems with it
can be resolved. Sorry .... 01 Apr 2021 [not a joke].
++++++++++ Blurring the inputs (avoid trying to match TOO much detail) +++++++++
-blur bb = Gaussian blur the input images by 'bb' (FWHM) voxels before
doing the alignment (the output dataset will not be blurred).
The default is 2.345 (for no good reason).
* Optionally, you can provide 2 values for 'bb', and then
the first one is applied to the base volume, the second
to the source volume.
-->>* e.g., '-blur 0 3' to skip blurring the base image
(if the base is a blurry template, for example).
* A negative blur radius means to use 3D median filtering,
rather than Gaussian blurring. This type of filtering will
better preserve edges, which might be important in alignment.
* If the base is a template volume that is already blurry,
you probably don't want to blur it again, but blurring
the source volume a little is probably a good idea, to
help the program avoid trying to match tiny features.
-pblur = Use progressive blurring; that is, for larger patch sizes,
the amount of blurring is larger. The general idea is to
avoid trying to match finer details when the patch size
and incremental warps are coarse. When '-blur' is used
as well, it sets a minimum amount of blurring that will
be used. [06 Aug 2014 -- '-pblur' may be the default someday].
* You can optionally give the fraction of the patch size that
is used for the progressive blur by providing a value between
0 and 0.25 after '-pblur'. If you provide TWO values, the
the first fraction is used for progressively blurring the
base image and the second for the source image. The default
parameters when just '-pblur' is given is the same as giving
the options as '-pblur 0.09 0.09'.
* '-pblur' is useful when trying to match 2 volumes with high
amounts of detail; e.g, warping one subject's brain image to
match another's, or trying to match a detailed template.
* Note that using negative values with '-blur' means that the
progressive blurring will be done with median filters, rather
than Gaussian linear blurring.
-->>*** The combination of the -allineate and -pblur options will make
the results of using 3dQwarp to align to a template somewhat
less sensitive to initial head position and scaling.
-nopblur = Don't use '-pblur'; equivalent to '-pblur 0 0'.
++++++++++ Restricting the warp directions ++++++++++
-noXdis = These options let you specify that the warp should not
-noYdis = displace in the given direction. For example, combining
-noZdis = -noXdis and -noZdis would mean only warping along the
y-direction would be allowed.
* Here, 'x' refers to the first coordinate in the dataset,
which is usually the Right-to-Left direction. Et cetera.
* Note that the output WARP dataset(s) will have sub-bricks
for the displacements which are all zero; every WARP dataset
has 3 sub-bricks.
++++++++++ Controlling the warp calculation process in detail ++++++++++
-iniwarp ww = 'ww' is a dataset with an initial nonlinear warp to use.
* If this option is not used, the initial warp is the identity.
* You can specify a catenation of warps (in quotes) here, as in
program 3dNwarpApply.
* You can scale a 3D warp's displacements by prefixing the dataset
name by 'FAC:a,b,c:Warpdatasetname' where a b c are numbers
by which to scale the x- y- z-displacments.
* As a special case, if you just input an affine matrix in a .1D
file, that also works -- it is treated as giving the initial
warp via the string "IDENT(base_dataset) matrix_file.aff12.1D".
* -iniwarp is usually used with -inilev to re-start 3dQwarp from
a previous stopping point, or from the output of '-allsave'.
* In particular, '-iniwarp' and '-resample' will work
together if you need to re-start a warp job from the
output of '-allsave'.
-inilev lv = 'lv' is the initial refinement 'level' at which to start.
* The combination of -inilev and -iniwarp lets you take the
results of a previous 3dQwarp run and refine them further:
3dQwarp -prefix Q25 -source SS+tlrc -base TEMPLATE+tlrc \
-minpatch 25 -blur 0 3
3dQwarp -prefix Q11 -source SS+tlrc -base TEMPLATE+tlrc \
-inilev 7 -iniwarp Q25_WARP+tlrc -blur 0 2
Note that the source dataset in the second run is the SAME as
in the first run. If you don't see why this is necessary,
then you probably need to seek help from an AFNI guru.
-->>** Also see the script @toMNI_Qwarpar for the use of this option
in creating a template dataset from a collection of scans from
different subjects.
-minpatch mm = Set the minimum patch size for warp searching to 'mm' voxels.
*OR* * The value of mm should be an odd integer.
-patchmin mm * The default value of mm is 25.
* For more accurate results than mm=25, try 19 or 13.
* The smallest allowed patch size is 5.
* OpenMP parallelization becomes inefficient for patch sizes
smaller than about 15x15x15 -- which is why running 3dQwarp
down to the minimum patch level of 5 can be very slow.
* You may want stop at a larger patch size (say 7 or 9) and use
the -Qfinal option to run that final level with quintic warps,
which might run faster and provide the same degree of warp
detail.
* Trying to make two different brain volumes match in fine detail
is usually a waste of time, especially in humans. There is too
much variability in anatomy to match gyrus to gyrus accurately,
especially in the small foldings in the outer cerebral cortex.
For this reason, the default minimum patch size is 25 voxels.
Using a smaller '-minpatch' might try to force the warp to
match features that do not match, and the result can be useless
image distortions -- another reason to LOOK AT THE RESULTS.
-maxlev lv = Here, 'lv' is the maximum refinement 'level' to use. This
is an alternate way to specify when the program should stop.
* To only do global polynomial warping, use '-maxlev 0'.
* If you use both '-minpatch' and '-maxlev', then you are
walking on the knife edge of danger.
* Of course, I know that you LIVE for such thrills.
-gridlist gl = This option provides an alternate way to specify the patch
grid sizes used in the warp optimization process. 'gl' is
a 1D file with a list of patches to use -- in most cases,
you will want to use it in the following form:
-gridlist '1D: 0 151 101 75 51'
* Here, a 0 patch size means the global domain. Patch sizes
otherwise should be odd integers >= 5.
* If you use the '0' patch size again after the first position,
you will actually get an iteration at the size of the
default patch level 1, where the patch sizes are 75% of
the volume dimension. There is no way to force the program
to literally repeat the sui generis step of lev=0.
* You cannot use -gridlist with: -plusminus :(
-allsave = This option lets you save the output warps from each level
*OR* of the refinement process. Mostly used for experimenting.
-saveall * Cannot be used with: -nopadWARP :(
* You could use the saved warps to create different versions
of the warped source datasets (using 3dNwarpApply), to help
you visualize how the warping process makes progress.
* The saved warps are written out at the end of each level,
before the next level starts computation. Thus, they could
be used to re-start the computation if the program crashed
(by using options '-inilev' and '-iniwarp').
* If '-allsave' is used with '-plusminus', the intermediate
saved warps are the "PLUS" half-warps (which are what the
program is optimizing).
-duplo = *** THIS OPTION IS NO LONGER AVAILABLE ***
-workhard = Iterate more times, which can help when the volumes are
hard to align at all, or when you hope to get a more precise
alignment.
* Slows the program down (possibly a lot), of course.
* Combined with '-lite', takes about the same amount of time
as '-nolite' without '-workhard' :)
* For finer control over which refinement levels work hard,
you can use this option in the form (for example)
-workhard:4:7
which implies the extra iterations will be done at levels
4, 5, 6, and 7, but not otherwise.
* You can also use '-superhard' to iterate even more, but
this extra option will REALLY slow things down.
-->>* Under most circumstances, you should not need to use either
-workhard or -superhard.
-->>* If you use this option in the form '-Workhard' (first letter
in upper case), then the second iteration at each level is
done with quintic polynomial warps.
-Qfinal = At the finest patch size (the final level), use Hermite
quintic polynomials for the warp instead of cubic polynomials.
* In a 3D 'patch', there are 2x2x2x3=24 cubic polynomial basis
function parameters over which to optimize (2 polynomials
dependent on each of the x,y,z directions, and 3 different
directions of displacement).
* There are 3x3x3x3=81 quintic polynomial parameters per patch.
* With -Qfinal, the final level will have more detail in
the allowed warps, at the cost of yet more CPU time.
* However, no patch below 7x7x7 in size will be done with quintic
polynomials.
* This option is also not usually needed, and is experimental.
(((........... Also see the section 'The warp polynomials' below ...........)))
-cubic12 = Use 12 parameter cubic polynomials, instead of 24 parameter
polynomials (the current default patch warps are 24 parameter).
* '-cubic12' will be faster than '-cubic24' and combining
it with '-workhard' will make '-cubic12' run at about the
same speed as the 24 parameter cubics.
* Is it less accurate than '-cubic24'? That is very hard
to say accurately without more work. In priniple, No.
* This option is now the default.
-cubic24 = Use 24 parameter cubic Hermite polynomials.
* This is the older set of basis functions [pre-2019], and
would normally be used only for backwards compatibility or
for testing.
-Qonly = Use Hermite quintic polynomials at all levels.
* Very slow (about 4 times longer than cubic).
* Will produce a (discrete representation of a) C2 warp.
-Quint81 = When quintic polynomials are used, use the full 81 parameter
set of basis functions.
* This is the older set of basis functions [pre-2019], and
would normally be used only for backwards compatibility or
for testing.
-Quint30 = Use the smaller 30 parameter set of quintic basis functions.
* These options ('-Quint81' and '-Quint30') only change
the operation if you also use some other option that
implies the use of quintic polynomials for warping.
-lite = Another way to specify the use of the 12 parameter cubics
and the 30 parameter quintics.
* This option now works with the '-plusminus' warping method :)
* THIS OPTION IS NOW THE DEFAULT * [Jan 2019]
-nolite = Turn off the '-lite' warp functions and use the 24 parameter
cubics *and* the 81 parameter quintics.
* This option is present for if you wish to have backwards
warping compatibility with older versions of 3dQwarp.
-nopad = Do NOT use zero-padding on the 3D base and source images.
[Default == zero-pad as needed]
* The underlying model for deformations goes to zero at the
edge of the volume being warped. However, if there is
significant data near an edge of the volume, then it won't
get displaced much, and so the results might not be good.
* Zero padding is designed as a way to work around this potential
problem. You should NOT need the '-nopad' option for any
reason that Zhark can think of, but it is here to be
symmetrical with 3dAllineate.
++ If the base dataset is closely cropped, so that the edges of
its 3D grid come close to the significant part of the volume,
using '-nopad' may cause poor fitting of the source to the
base, as the distortions required near the grid edges will
not be available in the restricted model. For this reason,
Zhark recommends that you do NOT use '-nopad'.
* Note that the output (warped from source) dataset will be on
the base dataset grid whether or not zero-padding is allowed.
However, unless you use option '-nopadWARP', allowing zero-
padding (i.e., the default operation) will make the output WARP
dataset(s) be on a larger grid (also see '-expad' below).
**** When grid centers of the base and source dataset are far apart
in (x,y,z) coordinates, then a large amount of zero-padding
is required to make the grid spaces overlap. This situation can
cause problems, and most often arises when the (x,y,z)=(0,0,0)
point in the source grid is in a corner of the volume instead
of the middle. You can fix that problem by using a command
like
@Align_Centers \
-base MNI152_2009_template_SSW.nii.gz \
-dset Fred.nii
and then using dataset Fred_shft.nii as your input file for all
purposes (including afni_proc.py).
++ One problem that happens with very large spatial shifts (from
3dAllineate initial alignment) is that the warp dataset can
be very huge. Not only does this cause a large file on output,
it also uses a lot of memory in the 3dQwarp optimization - so
much memory in some cases to cause the program to crash.
* A warning message will be output to the screen if very large
amounts of zero-padding are required.
* Intermediate between large amounts of padding and no padding
is the option below:
-Xpad = Puts an upper limit on the amount of padding, to prevent huge
warp datasets from being created.
-nopadWARP = If you do NOT use '-nopad' (that is, you DO allow zero-padding
during the warp computations), then the computed warp will often
be bigger than the base volume. This situation is normally not
an issue, but if for some reason you require the warp volume to
match the base volume, then use '-nopadWARP' to have the output
WARP dataset(s) truncated.
* Note that 3dNwarpApply and 3dNwarpAdjust will deal with warps
that are defined over grids that are larger than the datasets
to which they are applied; this is why Zhark says above that
a padded warp 'is normally not an issue'.
* However, if you want to use an AFNI nonlinear warp in some
external non-AFNI program, you might have to use this option :(
-expad EE = This option instructs the program to pad the warp by an extra
'EE' voxels (and then 3dQwarp starts optimizing it).
* This option is seldom needed, but can be useful if you
might later catenate the nonlinear warp -- via 3dNwarpCat --
with an affine transformation that contains a large shift.
Under that circumstance, the nonlinear warp might be shifted
partially outside its original grid, so expanding that grid
can avoid this problem.
* Note that this option perforce turns off '-nopadWARP'.
-ballopt = Normally, the incremental warp parameters are optimized inside
a rectangular 'box' (e.g., 24 dimensional for cubic patches, 81
for quintic patches), which limits the amount of distortion
allowed at each step. Using '-ballopt' switches these limits
to be applied to a 'ball' (interior of a hypersphere), which
can allow for larger incremental displacements. Use this
option if you think things need to be able to move farther.
* Note also that the '-lite' polynomial warps allow for
larger incremental displacements than the '-nolite' warps.
-boxopt = Use the 'box' optimization limits instead of the 'ball'
[this is the default at present].
* Note that if '-workhard' is used, then ball and box
optimization are alternated in the different iterations at
each level, so these two options have no effect in that case.
++++++++++ Meet-in-the-middle warping - Also know as '-plusminus' +++++++++
-plusminus = Normally, the warp displacements dis(x) are defined to match
base(x) to source(x+dis(x)). With this option, the match
is between base(x-dis(x)) and source(x+dis(x)) -- the two
images 'meet in the middle'.
* One goal is to mimic the warping done to MRI EPI data by
field inhomogeneities, when registering between a 'blip up'
and a 'blip down' down volume, which will have opposite
distortions.
* Define Wp(x) = x+dis(x) and Wm(x) = x-dis(x). Then since
base(Wm(x)) matches source(Wp(x)), by substituting INV(Wm(x))
wherever we see x, we have base(x) matches
source(Wp(INV(Wm(x))));
that is, the warp V(x) that one would get from the 'usual' way
of running 3dQwarp is V(x) = Wp(INV(Wm(x))).
* Conversely, we can calculate Wp(x) in terms of V(x) as follows:
If V(x) = x + dv(x), define Vh(x) = x + dv(x)/2;
then Wp(x) = V(INV(Vh(x)))
*** Also see the '-pmBASE' option described below.
-->>* Alas: -plusminus does not work with: -allineate :-(
++ If a prior linear alignment is needed, it will have
to be done "manually" using 3dAllineate, and then use
the output of that program as the '-source' dataset for
3dQwarp.
++ -plusminus works well if the base and source datasets
are reasonably well-aligned to start with. By this, I
mean that they overlap well, are not wildly rotated from
each other, and need some 'wiggling' to make them aligned.
-->>++ This option is basically meant for unwarping EPI data,
as described above.
* However, you can use -iniwarp with -plusminus :-)
-->>* The outputs have _PLUS (from the source dataset) and _MINUS
(from the base dataset) in their filenames, in addition to
the {prefix}. The -iwarp option, if present, will be ignored.
* If you use '-iniwarp' with '-plusminus', the warp dataset to
provide with '-iniwarp' is the '_PLUS' warp. That is, you can't
use a "full base-to-source warp" for the initial warp
(one reason '-allineate' doesn't work with '-plusminus').
-pmNAMES p m = This option lets you change the PLUS and MINUS prefix appendages
alluded to directly above to something else that might be more
easy for you to grok. For example, if you are warping EPI
volumes with phase-encoding in the LR-direction with volumes
that had phase-encoding in the RL-direction, you might do
something like
-base EPI_LR+orig -source EPI_RL+orig -plusminus -pmNAMES RL LR -prefix EPIuw
recalling that the PLUS name goes with the source (RL) and the
MINUS name goes with the base (RL). Then you'd end up with
datasets
EPIuw_LR+orig and EPIuw_LR_WARP+orig from the base
EPIuw_RL+orig and EPIuw_RL_WARP+orig from the source
The EPIuw_LR_WARP+orig file could then be used to unwarp (e.g.,
using 3dNwarpApply) other LR-encoded EPI datasets from the same
scanning session.
-pmBASE = With '-plusminus', computes the V(x) warp (source to base)
from the plusminus half-warp, and writes it to disk.
Also writes out the source dataset warped to base space,
in addition to the Wp(x) '_PLUS' and Wm(x) '_MINUS' results
* Sneaky aside: if you want potentially larger displacements
than 'normal' 3dQwarp, use '-plusminus', since the meet-in-the-
middle approach will allow the full-size displacements in EACH
of the half-warps, so that the overall displacement between
base and source can be larger. The use of '-pmBASE' will let
you get the source-transformed-to-base result at the end.
If you don't want the plusminus 'in-the-middle' outputs,
just delete them later.
++++++++++ How 'LOUD' do you want this program to be? ++++++++++
-verb = Print out very verbose progress messages (to stderr) :-)
-quiet = Cut out most of the fun fun fun progress messages :-(
INTERRUPTING the program gracefully¶
If you want to stop the program AND have it write out the results up to
the current point, you can do so with a Unix command like
kill -s QUIT processID
where 'processID' is the process identifier number (pid) for the 3dQwarp
program you want to terminate. A command like
ps aux | grep 3dQwarp
will give you a list of all your processes with the string '3dQwarp' in
the command line. For example, at the moment I wrote this text, I would
get the response
rwcox 62873 693.8 2.3 3274496 755284 p2 RN+ 12:36PM 380:25.26 3dQwarp -prefix ...
rwcox 6421 0.0 0.0 2423356 184 p0 R+ 1:33PM 0:00.00 grep 3dQwarp
rwcox 6418 0.0 0.0 2563664 7344 p4 S+ 1:31PM 0:00.15 vi 3dQwarp.c
so the processID for the actual run of 3dQwarp was 62873.
(Also, you can see that Zhark is a 'vi' acolyte, not an 'emacs' heretic.)
The program will 'notice' the QUIT signal at the end of the optimization
of the next patch, so it may be a moment or two before it actually saves
the output dataset(s) and exits.
Of course, if you just want to kill the process in a brute force way, with
nothing left behind to examine, then 'kill processID' will work.
Using 'kill -s QUIT' combined with '-allsave' might be useful in some
circumstances. At least to get some idea of what happened before you
were forced to stop 3dQwarp.
CLARIFICATION about the very confusing forward and inverse warp issue¶
An AFNI nonlinear warp dataset stores the displacements (in DICOM mm) from
the base dataset grid to the source dataset grid. For computing the source
dataset warped to the base dataset grid, these displacements are needed,
so that for each grid point in the output (warped) dataset, the corresponding
location in the source dataset can be found, and then the value of the source
at that point can be computed (interpolated).
That is, this forward warp is good for finding where a given point in the
base dataset maps to in the source dataset. However, for finding where a
given point in the source dataset maps to in the base dataset, the inverse
warp is needed. Or, if you wish to warp the base dataset to 'look like' the
source dataset, then you use 3dNwarpApply with the input warp being the
inverse warp from 3dQwarp.
STORAGE of 3D warps in AFNI¶
AFNI stores a 3D warp as a 3-volume dataset (NiFTI or AFNI format), with the
voxel values being the displacements in mm (32-bit floats) needed to
'reach out' and bring (interpolate) another dataset into alignment -- that is,
'pulling it back' to the grid defined in the warp dataset header. Thus, the
identity warp is all zero. These 3 volumes I refer to as ‘xd’, ‘yd’, and ‘zd’
in the internal comments, and they store (delta-x,delta-y,delta-z)
respectively (duh).
There is no provision in the warping software for 2D-only warps; that is,
warping one 2D image to another will still result in a 3D warp, with the zd
brick being chock full of zeros. This happenstance rarely occurs, since Zhark
believes he is the only person who actually has run the AFNI warping program
on 2D images.
In AFNI, (xd,yd,zd) are stored internally in DICOM order, in which +x=Left,
+y=Posterior, +z=Superior (LPS+); thus, DICOM x and y are sign-reversed from
the customary 'neuroscience order' RAS+. Note that the warp dataset grid need
not be stored in this same DICOM (x,y,z) order, which is sometimes confusing.
In the template datasets to which we nonlinearly warp data, we always use
DICOM order for the grids, so in practice warps generated in AFNI are usually
also physically ordered in the DICOM way -- but of course, someone can run our
warping software any which way they like and so get a warp dataset whose grid
order is not DICOM. But the (xd,yd,zd) entries will be in DICOM order.
On occasion (for example, when composing warps), the math will want the
displacement from a location outside of the warp dataset’s grid domain.
Originally, AFNI just treated those ghost displacements as zero or as equal
to the (xd,yd,zd) value at the closest edge grid point. However, this
method sometimes led to unhappy edge effects, and so now the software
linearly extrapolates the (xd,yd,zd) fields from each of the 6 faces of the
domain box to allow computation of such displacements. These linear
coefficients are computed from the warp displacement fields when the warp
dataset is read in, and so are not stored in the warp dataset header.
Inverse warps are computed when needed, and are not stored in the same
dataset with the forward warp. At one time, I thought that I’d always
keep them paired, but that idea fell by the wayside. AFNI does not make
use of deformation fields stored in datasets; that is, it does not
store or use datasets whose components are (x+xd,y+yd,z+zd). Such
a dataset could easily be computed with 3dcalc, of course.
There is no special header code in an AFNI warp dataset announcing that
'I am a warp!' By AFNI convention, 3D warp datasets have the substring
'_WARP' in their name, and inverse warps '_WARPINV'. But this is just a
convention, and no software depends on this being true. When AFNI warps
2 input datasets (A and B) together to 'meet in the middle' via the
'-plusminus' option (vs. bringing dataset A to be aligned directly to B),
two warp files are produced, one with the warp that brings A to the middle
'point' and one which brings 'B' to the middle point. These warps are
labeled with '_PLUS_WARP' and '_MINUS_WARP' in their filenames, as in
'Fred_PLUS_WARP.nii'. ('PLUS' and 'MINUS' can be altered via the
'-pmNAMES' option to 3dQwarp.)
If one is dealing only with affine transformation of coordinates, these
are stored (again referring to transformation of coordinates in DICOM
order) in plain ASCII text files, either with 3 lines of 4 numbers each,
(with the implicit last row of the matrix being 0 0 0 1, as usual).
or as all 12 numbers catenated into a single line (first 4 numbers are
the first row of the matrix, et cetera). This latter format is
always used when dealing with time-dependent affine transformations,
as from FMRI time series registration. A single matrix can be stored in
either format. At present, there is no provision for storing time-dependent
nonlinear warp datasets, since the use case has not arisen. When catenating
a time-dependent affine transform and a nonlinear warp (e.g., for direct
transformation from original EPI data to MNI space), the individual nonlinear
warp for each time point is computed and applied on-the-fly. Similarly, the
inverse warp can be computed on-the-fly, rather than being stored permanently.
Such on-the-fly warp+apply calculations are done in program 3dNwarpApply.
OUTLINE of warp optimization method¶
Repeated composition of incremental warps defined by Hermite cubic basis
functions, first over the entire volume, then over steadily shrinking and
overlapping patches at increasing 'levels': the patches shrink by a factor
of 0.75 at each level. Patches at levels 1 and higher have a 50% overlap.
NOTE: Internally, warps are stored as 'index warps', which are displacements
between 3D (i,j,k) grid indexes rather than between (x,y,z) coordinates.
The reason for this approach is that indexes are what is needed to
find the location in a dataset that a warp maps to. On output and on
input, the (x,y,z) displacements are converted from/to (i,j,k)
displacements. The '-inwarp' option allows you to output an 'index warp'
dataset, but this dataset has no function other than looking at it in
order to understand what the program was working with internally.
At 'level 0' (1 patch over the entire volume), Hermite quintic basis functions
are also employed, but these are not used at the more refined levels -- unless
one of the '-Qxxx' options is used. All basis functions herein are (at least)
continuously differentiable, so the discrete warp computed can be thought of
as a representation of an underlying C1 diffeomorphism. The basis functions
go to zero at the edge of each patch, so the overall warp will decay to the
identity warp (displacements=0) at the edge of the base volume. (However, use
of '-allineate' can make the final output warp be nonzero at the edges; the
programs that apply warps to datasets linearly extrapolate warp displacements
outside the 3D box over which the warp is defined.)
NOTE: * Option '-Qfinal' will use quintic polynomials at the final (smallest)
patch level.
* Option '-Qonly' will use quintic polynomials at all patch levels.
* Option '-Workhard' will run optimization on each patch twice,
first using cubic polynomials and later using quintic polynomials.
For this procedure to work, the source and base datasets need to be reasonably
well aligned already (e.g., via 3dAllineate, if necessary), as the nonlinear
optimization can only deal with relatively small displacements -- fractions of
a patch size.. Multiple warps can later be composed and applied via program
3dNwarpApply and/or 3dNwarpCat.
Note that it is not correct to say that the resulting warp is a piecewise cubic
(or quintic) polynomial. The first warp created (at level 0) is such a warp;
call that W0(x). Then the incremental warp W1(x) applied at the next iteration
is also a cubic polynomial warp (say), and the result is W0(W1(x)), which is
more complicated than a cubic polynomial -- and so on. The incremental warps
aren't added, but composed, so that the mathematical form of the final warp
would be very unwieldy to express in polynomial form. Of course, the program
just keeps track of the displacements, not the polynomial coefficients, so it
doesn't 'care' much about the underlying polynomials at all
One reason for incremental improvement by composition, rather than by addition,
is the simple fact that if W0(x) is invertible and W1(x) is invertible, then
W0(W1(x)) is also invertible -- but W0(x)+W1(x) might not be. The incremental
polynomial warps are kept invertible by simple constraints on the magnitudes
of their coefficients (i.e., the maximum size of incremental displacements).
The penalty is a Neo-Hookean elastic energy function, based on a combination of
bulk and shear distortions: cf. http://en.wikipedia.org/wiki/Neo-Hookean_solid
The goal is to keep the warps from becoming too 'weird' (doesn't always work).
By perusing the many options above, you can see that the user can control the
warp optimization in various ways. All these options make using 3dQwarp seem
pretty complicated. The reason there are so many options is that many different
cases arise, and we are trying to make the program flexible enough to deal with
them all. The SAMPLE USAGE section above is a good place to start for guidance.
*OR* you can use the @SSwarper or auto_warp.py scripts.
-————- The warp polynomials: ‘-lite’ and ‘-nolite’ —————-¶
The '-nolite' cubics have 8 basis functions per spatial dimension, since they
are the full tensor product of the 2 Hermite cubics H0 and H1:
H0(x)*H0(y)*H0(z) H1(x)*H0(y)*H0(z) H0(x)*H1(y)*H0(z) H0(x)*H0(y)*H1(z)
H1(x)*H1(y)*H0(z) H1(x)*H0(y)*H1(z) H0(x)*H1(y)*H1(z) H1(x)*H1(y)*H1(z)
and then there are 3 sets of these for x, y, and z displacements, giving
24 total basis functions for a cubic 3D warp patch. The '-lite' cubics
omit any of the tensor product functions whose indexes sum to more than 1,
so there are only 4 basis functions per spatial dimension:
H0(x)*H0(y)*H0(z) H1(x)*H0(y)*H0(z) H0(x)*H1(y)*H0(z) H0(x)*H0(y)*H1(z)
yielding 12 total basis functions (again, 3 of each function above for each
spatial dimension). The 2 1D basis functions, defined over the interval
[-1,1], and scaled to have maximum magnitude 1, are
H0(x) = (1-abs(x))^2 * (1+2*abs(x)) // H0(0) = 1 H0'(0) = 0
H1(x) = (1-abs(x))^2 * x * 6.75 // H1(0) = 0 H1'(0) = 6.75 H1(1/3) = 1
These functions and their first derivatives are 0 at x=+/-1, which is apparent
from the '(1-abs(x))^2' factor they have in common. The functions are also
continuous and differentiable at x=0; thus, they and their unit translates
can serve as a basis for C1(R): the Banach space of continuously differentiable
functions on the real line.
One effect of using the '-lite' polynomial warps is that 3dQwarp runs faster,
since there are fewer parameters to optimize for each patch. Accuracy should
not be imparied,as the approximation quality (in the mathematical sense) of
the '-lite' polynomials is the same order as the '-nolite' full tensor product.
Another effect is that the upper limits on the displacements by any individual
warp patch are somewhat larger than for the full basis set, which may be useful
in some situations.
Similarly, the '-nolite' quintics have 27 basis functions per spatial
dimension, since they are the tensor products of the 3 Hermite quintics
Q0, Q1, Q2. The '-lite' quintics omit any tensor product whose indexes sum
to more than 2. Formulae for these 3 polynomials can be found in function
HQwarp_eval_basis() in AFNI file mri_nwarp.c. For each monomial Qi(x),
Qi(+/-1)=Qi'(+/-1)=Qi''(+/-1) = 0;
these functions are twice continuously differentiable, and can serve as
a basis for C2(R).
-——– Why is it ‘cost functional’ and not ‘cost function’ ??? ——–¶
In mathematics, a 'functional' is a function that maps an object in an infinite
dimensional space to a scalar. A typical example of a functional is a function
of a function, such as I(f) = definite integral from -1 to 1 of f(x) dx.
In this example, 'f' is a function, which is presumed to be integrable, and thus
an element of the infinite dimensional linear space denoted by L1(-1,1).
Thus, as Zhark was brutally trained in his mathematics bootcamp, the value
being optimized, being a number (scalar) that is calculated from a function
(warp), the 'machine' that calculates this value is a 'functional'. It also
gets the word 'cost' attached as it is something the program is trying to
reduce, and everyone wants to reduce the cost of what they are buying, right?
(AFNI does not come with coupons :-)
WEIGHT construction¶
The cost functional is computed giving (potentially) different weights to
different voxels. The default weight 3D volume is constructed from the
base dataset as follows (i.e., this is what '-useweight' does):
(0) Take absolute value of each voxel value.
(1) Zero out voxels within 4% of each edge
(i.e., 10 voxels in a 256x256x256 dataset).
(2) Define L by applying the '3dClipLevel -mfrac 0.5' algorithm
and then multiplying the result by 3. Then, any values over this
L 'large' value are reduced to L -- i.e., spikes are squashed.
(3) A 3D median filter over a ball with radius 2.25 voxels is applied
to further squash any weird stuff. (This radius is fixed.)
(4) A Gaussian blur of FWHM '-wtgaus' is applied (default = 4.5 voxels).
(5) Define L1 = 0.05 times the maximum of the result from (4).
Define L2 = 0.33 times '3dClipLevel -mfrac 0.33' applied to (4).
Define L = max(L1,L2).
Create a binary mask of all voxels from (4) that are >= L.
Find the largest contiguous cluster in that mask, erode it
a little, and then again find the largest cluster in what remains.
(The purpose of this to is guillotine off any small 'necks'.)
Zero out all voxels in (4) that are NOT in this surviving cluster.
(6) Scale the result from (5) to the range 0..1. This is the weight
volume.
(X) Remember you can save the computed weight volume to a dataset by
using the '-wtprefix' option.
Where did this scheme come from? A mixture of experimentation, intuition,
and plain old SWAG.
***** This program is experimental and subject to sudden horrific change! *****
((((( OK, it's less experimental now, and so sudden changes will be mild. )))))
----- AUTHOR = Zhark the Grotesquely Warped -- Fall/Winter/Spring 2012-13 -----
----- (but still strangely handsome) -----
* This binary version of 3dQwarp 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.
* Tests show that using more 12-16 CPUs with 3dQwarp doesn't help much.
If you have more CPUs on one system, it's faster to run two or three
separate registration jobs in parallel than to use all the CPUs on
one 3dQwarp task at a time.
