AFNI program: 3dZipperZapper
Output of -help
++ version: 2018_02_06
OVERVIEW ~1~
This is a basic program to help highlight problematic volumes in data
sets, specifically in EPI/DWI data sets with interleaved acquisition.
Intra-volume subject motion can be quite problematic, potentially
bad-ifying the data values in the volume so much that it is basically
useless for analysis. In FMRI analysis, outlier counts might be
useful to find ensuing badness (e.g., via 3dToutcount). However, with
DWI data, we might want to find it without aligning the volumes
(esp. due to the necessarily differing contrasts) and without tensor
fitting.
*Therefore*, this program will look through axial slices of a data set
for brightness fluctuations and/or dropout slices. It will build a
list of volumes indices that it identifies as bad, and the user can
then use something like the 'fat_proc_filter_dwis' program after to
apply the filtration to the volumetric dset *as well as* to any
accompanying b-value, gradient vector, b-matrix, etc., text files.
The program works by looking for alternating brightness patterns in
the data (again, specifically in axial slices, so if your data was
acquired differently, this program ain't for you! (weeellll, some
tricks with changing header info miiiight be able to work then)). It
should be run *before* any processing, particularly alignments or
unwarping things, because those could change the slice locations.
Additionally, it has mainly been tested on 3T data of humans; it is
possible that it will work equally well on 7T or non-humans, but be
sure to check results carefully in the latter cases (well, *always*
check your data carefully!).
Note that there is also 'fat_proc_select_vols' program for
interactively selecting out bad volumes, by looking at a sheet of
sagittal images from the DWI set. That might be useful for amending
or altering the output from this program, if necessary.
written by PA Taylor (started Jan, 2018)
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
USAGE ~1~
Input: + a 3D+time data set of DWI or EPI volumes,
+ a mask of the brain-ish region.
Output: + a mask of potentially bad slices across the input dset,
+ a 1D (text) file containing a list of the bad volumes,
+ a 1D file of the per-volume parameters used to detect
badness,
+ a 1D file of the slices within which calculations were made,
+ a text file with the selector string of *good* volumes
in the dset (for easy use with fat_proc_filter_dwis,
for example).
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
COMMAND ~1~
3dZipperZapper \
-input FFF {-mask MMM} \
-prefix PPP \
{-min_slice_nvox N} \
{-min_streak_len L} \
{-do_out_slice_param} \
{-no_out_bad_mask} \
{-no_out_text_vals} \
where:
-input FFF :input the 3D+time file of DWIs or EPIs.
-mask MMM :optional input of a single volume mask file, which
gets applied to the each volume in FFF. Otherwise,
the dataset is assumed to be masked already.
-prefix PPP :prefix for output file name. Any volumetric file
extension included here (e.g., '.nii.gz') is
propagated to any output volumetric dsets.
-do_out_slice_param
:output the map of slice parameters (not done by
default). Might be of interest for investigating
data. Output file name base will be: PPP_param.
-no_out_bad_mask
:do *not* output the mask of 'bad' slices that shows
which volumes are considered bad (is output by
default). Output file name base will be: PPP_badmask.
-no_out_text_vals
:do *not* output the 1D files of the slice parameter
values (are output by default). The list of slices
in the mask (file name: PPP_sli.1D) and the list of
values per slice per volume (file name: PPP_param.1D)
are output.
... and for having fine control of which drop criteria to use
(def: use all available, see listing in NOTES):
-dont_use_streak :
:several criteria are used to search for bad slices.
Using this opt, you elect to turn off the 'streak'
criterion. See the NOTES below for more description.
-dont_use_drop :
:several criteria are used to search for bad slices.
Using this opt, you elect to turn off the 'drop'
criterion. See the NOTES below for more description.
-dont_use_corr :
:several criteria are used to search for bad slices.
Using this opt, you elect to turn off the 'corr'
criterion. See the NOTES below for more description.
... and for having fine control of drop criteria parameters:
-disp_def_params
:display the defaults for each of the following parameters.
NB: the value for MIN_SLICE_NVOX will be '-1', meaning
that the number of voxels will be calculated from the
slice size---see the option help, below.
-min_slice_nvox N
:set the minimum number of voxels to be in the mask
for a given slice to be included in the calcs.
N must be >0 (and likely much more so, to be useful).
Default: use 10 percent of the axial slice's nvox.
-min_streak_len MSL
:set the minimum number of slices in a row to look for
fluctuations within (def: MSL=4). That is, if 'large
enough' fluctuations are found in L consecutive slices,
then the volume is flagged for motion. A larger MSL means
that more slices need to vary for a volume to be flagged
for 'brightness fluctuations'. NB: this does parameter
setting does not affect the search for dropout slices.
Part of 'streak' criterion; see NOTES for more details.
-min_streak_val MSV
:set the minimum magnitude of voxelwise relative diffs
to perhaps be problematic.
Part of 'streak' criterion; see NOTES for more details.
-min_drop_frac MDF
:set the minimum fraction for judging if the change in
'slice parameter' differences between neighboring slices
might be a sign of badness.
Part of 'drop' criterion; see NOTES for more details.
-min_drop_diff MDD
:set the minimum 'slice parameter' value within a single
slice that might be considered bad sign (e.g., of
dropout).
Part of 'drop' criterion; see NOTES for more details.
-min_corr_len MCL
:set the minimum number of slices in a row to look for
consecutive anticorrelations in brightness differences.
Part of 'corr' criterion; see NOTES for more details.
-min_corr_corr MCC
:set the threshold for the magnitude of anticorrelations
to be considered potentially bad.
Part of 'corr' criterion; see NOTES for more details.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
NOTES ~1~
Drop Criteria ~2~
At present, there are 3 distinct criteria used to search for bad slices,
by default. The list of bad slices from each method are combined through
a union operation, so that any slice identified as 'bad' by any of the
criteria is evaluated as 'bad' in the end. The set of criteria might
grow over time.
As of March 30, 2022, users have the option of turning of any of the
criteria, via the '-dont_use_*' options.
The current criteria are described by keyword as follows (see the next
section for definitions of slipar, slicorr, and other mysterious
quantities):
'streak' criterion
Walk upwards through slices in the volume. If the absolute value of
differences in slipar values stays high, you may have badness.
MIN_STREAK_VAL is the magnitude threshold for judging if differences
are high.
MIN_STREAK_LEN is the minimal number of consecutive slices that have to
have high differences to be a sign of badness.
'drop' criterion
If a particular slice has a very high slipar magnitude, you may have
badness.
MIN_DROP_FRAC is the threshold magnitude for that.
If the absolute difference in slipar between neighboring slices is very
high, you may have badness.
MIN_DROP_DIFF is the threshold for judging if the absolute difference
is large enough to be a sign of badness.
'corr' criterion
Walk upwards through slices in the volume. If slicorr values are
strongly anticorrelated for several slices in a row, you may have
badness.
MIN_CORR_CORR is the magnitude threshold for judging if anticorrelation
is high (the minus sign is applied internally).
MIN_CORR_LEN is the minimal number of consecutive slices that have to be
highly anticorrelated to be a sign of badness.
Underlying quantities for drop criteria ~2~
Many drop criteria depend on the calculated 'slice parameter' (slipar)
values. These are generated per slice as follows:
+ For each voxel in a slice, calculate its relative difference with its
'upstairs' neighbor:
reldiff(A, B) = 0.5*(A - B)/(abs(A) + abs(B)).
+ Calculate the number of times reldiff is positive in a slice, divide that
by the total number of voxels in the slice, and subtract 0.5 (to center
that quantity around 0). This is the slipar value per slice.
Separately, we also 'slice correlation' (slicorr) values of a slice with
its upstairs neighbor:
+ For each slice, make a time series by flattening the 2D array of slipar
values for voxels that exist in both that slice and its upstairs (call
that X).
+ Make a time series of flattening the matched upstairs neighbor slipar
values (call that Y).
+ The slicorr value per slices is the Pearson correlation value of X and Y.
So, slicorr tells you something about how correlated your slice's reldiff
patterns are with your upstairs neighbor.
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
EXAMPLES ~1~
1) All types of outputs:
3dZipperZapper \
-input AP.nii.gz \
-mask AP_mask.nii.gz \
-prefix ZZZ.nii.gz \
-do_out_slice_param
2) No volumetric outputs (only if speed/write time is super
important?):
3dZipperZapper \
-input AP.nii.gz \
-mask AP_mask.nii.gz \
-prefix ZZZ.nii.gz \
-no_out_bad_mask
# ------------------------------------------------------------------
# ------------------------------------------------------------------------
This page auto-generated on
Wed Apr 24 22:25:26 EDT 2024