# 3dTproject¶

```
Usage: 3dTproject [options]
This program projects (detrends) out various 'nuisance' time series from each
voxel in the input dataset. Note that all the projections are done via linear
regression, including the frequency-based options such as '-passband'. In this
way, you can bandpass time-censored data, and at the same time, remove other
time series of no interest (e.g., physiological estimates, motion parameters).
--------
OPTIONS:
--------
-input dataset = Specifies the input dataset.
-prefix ppp = Specifies the output dataset, as usual.
-censor cname = As in 3dDeconvolve.
-CENSORTR clist = As in 3dDeconvolve.
-cenmode mode = 'mode' specifies how censored time points are treated in
the output dataset:
++ mode = ZERO ==> put zero values in their place
==> output datset is same length as input
++ mode = KILL ==> remove those time points
==> output dataset is shorter than input
++ mode = NTRP ==> censored values are replaced by interpolated
neighboring (in time) non-censored values,
BEFORE any projections, and then the
analysis proceeds without actual removal
of any time points -- this feature is to
keep the Spanish Inquisition happy.
** The default mode is KILL !!!
-concat ccc.1D = The catenation file, as in 3dDeconvolve, containing the
TR indexes of the start points for each contiguous run
within the input dataset (the first entry should be 0).
++ Also as in 3dDeconvolve, if the input dataset is
automatically catenated from a collection of datasets,
then the run start indexes are determined directly,
and '-concat' is not needed (and will be ignored).
++ Each run must have at least 9 time points AFTER
censoring, or the program will not work!
++ The only use made of this input is in setting up
the bandpass/stopband regressors.
++ '-ort' and '-dsort' regressors run through all time
points, as read in. If you want separate projections
in each run, then you must either break these ort files
into appropriate components, OR you must run 3dTproject
for each run separately, using the appropriate pieces
from the ort files via the '{...}' selector for the
1D files and the '[...]' selector for the datasets.
-noblock = Also as in 3dDeconvolve, if you want the program to treat
an auto-catenated dataset as one long run, use this option.
++ However, '-noblock' will not affect catenation if you use
the '-concat' option.
-ort f.1D = Remove each column in f.1D
++ Multiple -ort options are allowed.
++ Each column will have its mean removed.
-polort pp = Remove polynomials up to and including degree pp.
++ Default value is 2.
++ It makes no sense to use a value of pp greater than
2, if you are bandpassing out the lower frequencies!
++ For catenated datasets, each run gets a separate set
set of pp+1 Legendre polynomial regressors.
++ Use of -polort -1 is not advised (if data mean != 0),
even if -ort contains constant terms, as all means are
removed.
-dsort fset = Remove the 3D+time time series in dataset fset.
++ That is, 'fset' contains a different nuisance time
series for each voxel (e.g., from AnatICOR).
++ Multiple -dsort options are allowed.
-passband fbot ftop = Remove all frequencies EXCEPT those in the range
*OR* -bandpass fbot..ftop.
++ Only one -passband option is allowed.
-stopband sbot stop = Remove all frequencies in the range sbot..stop.
++ More than one -stopband option is allowed.
++ For example, '-passband 0.01 0.10' is equivalent to
'-stopband 0 0.0099 -stopband 0.1001 9999'
-dt dd = Use time step dd for the frequency calculations,
*OR* -TR rather than the value stored in the dataset header.
-mask mset = Only operate on voxels nonzero in the mset dataset.
*OR* ++ Use '-mask AUTO' to have the program generate the
-automask mask automatically (or use '-automask')
++ Voxels outside the mask will be filled with zeros.
++ If no masking option is given, then all voxels
will be processed.
-blur fff = Blur (inside the mask only) with a filter that has
width (FWHM) of fff millimeters.
++ Spatial blurring (if done) is after the time
series filtering.
-norm = Normalize each output time series to have sum of
squares = 1. This is the LAST operation.
-quiet = Hide the super-fun and thrilling progress messages.
-verb = The program will save the fixed ort matrix and its
singular values into .1D files, for post-mortems.
It will also print out more progress messages, which
might help with figuring out what's happening when
problems occur.
------
NOTES:
------
* The output dataset is in floating point format.
* Removal of the various undesired components is via linear regression.
In particular, this method allows for bandpassing of censored time
series.
* If you like technical math jargon (and who doesn't?), this program
performs orthogonal projection onto the null space of the set of 'ort'
vectors assembled from the various options '-polort', '-ort',
'-passband', '-stopband', and '-dsort'.
* If A is a matrix whose column comprise the vectors to be projected
out, define the projection matrix Q(A) by
Q(A) = I - A psinv(A)
where psinv(A) is the pseudo-inverse of A [e.g., inv(A'A)A' -- but
the pseudo-inverse is actually calculated here via the SVD algorithm.]
* If option '-dsort' is used, each voxel has a different matrix of
regressors -- encode this extra set of regressors in matrix B
(i.e., each column of B is a vector to be removed from its voxel's
time series). Then the projection for the compound matrix [A B] is
Q( Q(A)B ) Q(A)
that is, A is projected out of B, then the projector for that
reduced B is formed, and applied to the projector for the
voxel-independent A. Since the number of columns in B is usually
many fewer than the number of columns in A, this technique can
be much faster than constructing the full Q([A B]) for each voxel.
(Since Q(A) only need to be constructed once for all voxels.)
A little fun linear algebra will show you that Q(Q(A)B)Q(A) = Q([A B]).
* A similar regression could be done via the slower 3dTfitter program:
3dTfitter -RHS inputdataset+orig \
-LHS ort1.1D dsort2+orig \
-polort 2 -prefix NULL \
-fitts Tfit
3dcalc -a inputdataset+orig -b Tfit+orig -expr 'a-b' \
-datum float -prefix Tresidual
3dTproject should be MUCH more efficient, especially when using
voxel-specific regressors (i.e., '-dsort'), and of course, it also
offers internal generation of the bandpass/stopband regressors,
as well as censoring, blurring, and L2-norming.
* This version of the program is compiled using OpenMP for speed.
* Authored by RWCox in a fit of excessive linear algebra [summer 2013].
++ Compile date = Sep 17 2020 {AFNI_20.2.18:linux_ubuntu_16_64}
```