This patch addresses half of the problem cited at #147 (comment)

The other half will be added in that patch (flipping vector field directions)

Hi @vfonov this fix is independent of the RAS<->LPS work, it estimates the inverse transform via a method similar to grid_inverse_transform_point_with_input_steps from grid_transform.c

1. How do you determine the domain and sampling where inverse transformation is calculated?
2. how long does it take now to open .xfm file that has .inverse flag set ?

Hi @vfonov,

The default for the filter is to create a domain and sampling identical to the input deformation field.

The time added for a 1mm iso nlin registration invert for the default 5 iterations is ~7ms.

well in my experiment, running InvertDisplacementFieldImageFilter with default parameters on a a 2x2x2 grid file produced by MINC longitudinal pipeline takes about 1 min. Is IterativeInverseDisplacementFieldImageFilter that much faster and still produces appropriate result?
Running minc-based inverse on a 2x2x2 grid file takes 20sec.

Sorry, follow-up here, a bit of a mixup on the timings, I was using the internal timing from the Print() functionality of the IterativeInverse function, which seems to be the mean of some loop within the code, rather than the true value. I annotated the code with a itk::TimeProbe start(); and stop(); for the filter->Update();

Here are the results for a 1x1x1 mm deformation field generated from antsRegistration between the GAD and MNI model mentioned in the RAS/LPS discussion, applied using antsApplyTransforms, this is specifically the time spent in the inversion filter.

Settings for the filters,
InterativeInverse (settings roughly copied from Invert filter):

inverter->SetNumberOfIterations(50);
inverter->SetStopValue(0.1);

Invert (settings based on the ITK example):

inverter->SetMaximumNumberOfIterations( 50 );
inverter->SetMeanErrorToleranceThreshold( 0.001 );
inverter->SetMaxErrorToleranceThreshold( 0.1 );
inverter->SetEnforceBoundaryCondition( false );

InverativeInverse: Time elapsed: 8.78565
InvertDisplacement: Time elapsed: 3.48849

minccmp comparing the transformed images made with the "true" inverse field (generated by antsRegistration) and the inverted fields:
Iterative:

file[0]:      true_inverse.mnc
mask file:    (null)
file[1]:      Iterative.mnc
ssq:          597264.9631
rmse:         0.2646113505
xcorr:        0.9999792877
zscore:       0.00321001647

Invert:

file[0]:      true_inverse.mnc
mask file:    (null)
file[1]:      Invert.mnc
ssq:          32444.84004
rmse:         0.06167338255
xcorr:        0.9999988759
zscore:       0.0008179832775

Happy to switch to the Invert filter based on it being faster and more accurate.

So, it looks like it works fast on a "good" deformation field from ANTs , how long would it take on a not so smooth field from minctracc?

Here,

this is output of CIVET pipeline on SAMIR-99 data sample. Let's see what the proposed code will do.
001_nlfit_It_inv.tar.gz

Thanks @vfonov

I generated nlpfit outputs (2x2x2) using the same GAD-MNI pairs and got the following results:
Invert: Time elapsed: 0.107751
IterativeInverse; Time elapsed: 0.836424

And generating inverse field files from the SAMIR-99 data you sent with antsApplyTransforms -o [output.mnc,1]
Invert: Time elapsed: 0.967698
Iterative: Time elapsed: 1.48106

Updated PR to use Invert filter instead of IterativeInverse filter.

is the result similar to mincresample ?

I don't have/couldn't find a copy of SAMIR-99 so I couldn't directly compare resamples, but for the GAD-MNI pair from nlpfit:

file[0]:      itk_Invert.mnc
mask file:    (null)
file[1]:      forward-mincresample.mnc
ssq:          1359805581
rmse:         12.62592949
xcorr:        0.9974990007
zscore:       0.01719338444

Its not quite as nice a comparison as the antsRegistration one since minctracc doesn't produce a forward/reverse pair that I can use to do a more-head-to-head comparison, but everything looks fine.

technically it is pretty trivial to see if the inverse is good - i.e X * inv(X) = identity
but here is the scan

001_t1_tal.tar.gz

Thanks @vfonov, an excellent suggestion re: identity.

Here, I concat (via added two transforms to the text of the .xfm) the forward and reverse transforms and apply those transforms with mincresample, and antsApplyTransforms with the itkInvertDisplacementField.

I then compare these outputs using minccmp against the original image since this operation should be the identity operation:
mincresample:

file[0]:      identity.mnc
mask file:    (null)
file[1]:      001_t1_tal_1.0.mnc
ssq:          1415726556
rmse:         14.11176214
xcorr:        0.9969889137
zscore:       0.01226894951

antsApplyTransforms:

file[0]:      001_t1_tal_1.0.mnc
mask file:    (null)
file[1]:      itk_invert_identity.mnc
ssq:          13572824.13
rmse:         1.381741078
xcorr:        0.9999711985
zscore:       0.002255303473

Look like the InvertDisplacement filter does a good job, and does so in 0.973168 seconds according to my instrumentation of the code.

ok, let's merge it into the mainline ITK.
Would be good to include identity test into the testing code.

+1 on a test. I'll look at how to add one and whether some of the existing example files could be used.

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

This code is valid, I still need to add tests