antsRegistration interface only using first items in fixed_image and moving_image input lists for all transforms

[sgiavasis]

Hey guys,

We noticed this error after a recent change in one of our registration workflows. We have a workflow where we send a skullstripped brain into antsRegistration for the first two transforms, and then a skull-on image to the third (it used to be that we used the same input and references for all three).

In the report.rst, the correct list is shown as the inputs for fixed_image and moving_image (i.e. [input1, input1, input2] and [reference1, reference1, reference2]). However, when the command string is formed, it produces the antsRegistration command with input1 and reference1 being used for all three transforms, instead of using input2 and reference2 for the third one.

Thanks in advance.

Steve

[sgiavasis]

Here's a snippet from the report.rst for reference:

Execution Inputs

• args :
• collapse_linear_transforms_to_fixed_image_header : False
• collapse_output_transforms : False
• convergence_threshold : [1e-08, 1e-08, 1e-09]
• convergence_window_size : [10, 10, 15]
• dimension : 3
• environ : {'NSLOTS': '1'}
• fixed_image : ['/usr/share/fsl/5.0/data/standard/MNI152_T1_2mm_brain.nii.gz', '/usr/share/fsl/5.0/data/standard/MNI152_T1_2mm_brain.nii.gz', '/usr/share/fsl/5.0/data/standard/MNI152_T1_2mm.nii.gz']
• fixed_image_mask :
• ignore_exception : False
• initial_moving_transform :
• initial_moving_transform_com : 0
• interpolation : Linear
• invert_initial_moving_transform :
• metric : ['MI', 'MI', 'CC']
• metric_item_trait :
• metric_stage_trait :
• metric_weight : [1.0, 1.0, 1.0]
• metric_weight_item_trait :
• metric_weight_stage_trait :
• moving_image : ['/home/sgiavasis/run/037/skullon/w/resting_preproc_2014113/anat_preproc_0/anat_brain_only/mprage_resample_calc.nii.gz', '/home/sgiavasis/run/037/skullon/w/resting_preproc_2014113/anat_preproc_0/anat_brain_only/mprage_resample_calc.nii.gz', '/home/sgiavasis/run/037/skullon/w/resting_preproc_2014113/anat_preproc_0/anat_reorient/mprage_resample.nii.gz']
• moving_image_mask :
• num_threads : 1
• number_of_iterations : [[1000, 500, 250, 100], [1000, 500, 250, 100], [100, 100, 70, 20]]
• output_inverse_warped_image :
• output_transform_prefix : transform
• output_warped_image : True
• radius_bins_item_trait :
• radius_bins_stage_trait :
• radius_or_number_of_bins : [32, 32, 4]
• sampling_percentage : [0.25, 0.25, None]
• sampling_percentage_item_trait :
• sampling_percentage_stage_trait :
• sampling_strategy : ['Regular', 'Regular', None]
• sampling_strategy_item_trait :
• sampling_strategy_stage_trait :
• shrink_factors : [[8, 4, 2, 1], [8, 4, 2, 1], [6, 4, 2, 1]]
• sigma_units :
• smoothing_sigmas : [[3.0, 2.0, 1.0, 0.0], [3.0, 2.0, 1.0, 0.0], [3.0, 2.0, 1.0, 0.0]]
• terminal_output : allatonce
• transform_parameters : [(0.1,), (0.1,), (0.1, 3.0, 0.0)]
• transforms : ['Rigid', 'Affine', 'SyN']
• use_estimate_learning_rate_once :
• use_histogram_matching : True
• winsorize_lower_quantile : 0.01
• winsorize_upper_quantile : 0.99
• write_composite_transform :

Runtime info

• command : antsRegistration --collapse-linear-transforms-to-fixed-image-header 0 --collapse-output-transforms 0 --dimensionality 3 --initial-moving-transform [ /usr/share/fsl/5.0/data/standard/MNI152_T1_2mm_brain.nii.gz, /home/sgiavasis/run/037/skullon/w/resting_preproc_2014113/anat_preproc_0/anat_brain_only/mprage_resample_calc.nii.gz, 0 ] --interpolation Linear --output [ transform, transform_Warped.nii.gz ] --transform Rigid[ 0.1 ] --metric MI[ /usr/share/fsl/5.0/data/standard/MNI152_T1_2mm_brain.nii.gz, /home/sgiavasis/run/037/skullon/w/resting_preproc_2014113/anat_preproc_0/anat_brain_only/mprage_resample_calc.nii.gz, 1, 32, Regular, 0.25 ] --convergence [ 1000x500x250x100, 1e-08, 10 ] --smoothing-sigmas 3.0x2.0x1.0x0.0 --shrink-factors 8x4x2x1 --use-histogram-matching 1 --transform Affine[ 0.1 ] --metric MI[ /usr/share/fsl/5.0/data/standard/MNI152_T1_2mm_brain.nii.gz, /home/sgiavasis/run/037/skullon/w/resting_preproc_2014113/anat_preproc_0/anat_brain_only/mprage_resample_calc.nii.gz, 1, 32, Regular, 0.25 ] --convergence [ 1000x500x250x100, 1e-08, 10 ] --smoothing-sigmas 3.0x2.0x1.0x0.0 --shrink-factors 8x4x2x1 --use-histogram-matching 1 --transform SyN[ 0.1, 3.0, 0.0 ] --metric CC[ /usr/share/fsl/5.0/data/standard/MNI152_T1_2mm_brain.nii.gz, /home/sgiavasis/run/037/skullon/w/resting_preproc_2014113/anat_preproc_0/anat_brain_only/mprage_resample_calc.nii.gz, 1, 4 ] --convergence [ 100x100x70x20, 1e-09, 15 ] --smoothing-sigmas 3.0x2.0x1.0x0.0 --shrink-factors 6x4x2x1 --use-histogram-matching 1
• duration : 484.53864

[satra]

antsRegistration only expects a single fixed and moving image. if you are giving these images, they must be at the level of a MapNode, in which case iterfield must also have fixed and moving images in them. each ants call should only have one pair.

[satra]

let me take back that statement. something is wrong here.

[satra]

@hjmjohnson - only the first image is being used here:

nipype/nipype/interfaces/ants/registration.py

Line 432 in 44f7566

fixed = self.inputs.fixed_image[0]

even though the example appears to give it multiple images.

[sgiavasis]

Not to rush you guys, but any progress on this?

[ischwabacher]

It looks like replacing the 0s with index in lines 432 and 434 of the file @satra linked should solve the problem, though I'm not sure there are no more problems waiting to crop up.

[sgiavasis]

I tried that, it doesn't. self.inputs.fixed_image and self.inputs.moving_image come in as one-item lists.

[ischwabacher]

_formatMetric works for me:

In [1]: import nipype.interfaces.ants as ants

In [2]: ants.Registration(fixed_image=['MNI152_T1.nii', 'MNI152_T2.nii'],
   ...:                   moving_image=['T1.nii', 'T2.nii'],
   ...:                   metric=['CC', 'MI'],
   ...:                   metric_weight=[1.0, 1.0],
   ...:                   radius_or_number_of_bins=[5, 32])._formatMetric(1)
Out[2]: ['MI[ MNI152_T2.nii, T2.nii, 1, 32 ]']

But then there's this, which looks problematic:

In [3]: ants.Registration(fixed_image=['MNI152_T1.nii', 'MNI152_T2.nii'],
   ...:                   moving_image=['T1.nii', 'T2.nii'],
   ...:                   metric=['CC', 'MI'],
   ...:                   metric_weight=[1.0, 1.0],
   ...:                   radius_or_number_of_bins=[5, 32],
   ...:                   smoothing_sigmas=[[4, 2, 1, 0]],
   ...:                   shrink_factors=[[8, 4, 2, 1]],
   ...:                   transforms=['SyN'],
   ...:                   number_of_iterations=[[10, 10, 10, 10]],
   ...:                   transform_parameters=[[0.5]]).cmdline
Out[3]: 'antsRegistration --collapse-linear-transforms-to-fixed-image-header 0 -
-collapse-output-transforms 0 --dimensionality 3 --interpolation Linear --output
 transform --transform SyN[ 0.5 ] --metric CC[ MNI152_T1.nii, T1.nii, 1, 5 ] --c
onvergence [ 10x10x10x10, 1e-06, 10 ] --smoothing-sigmas 4.0x2.0x1.0x0.0 --shrin
k-factors 8x4x2x1 --use-histogram-matching 1 --winsorize-image-intensities [ 0.0
, 1.0 ]  --write-composite-transform 0'
# We do indeed get a one-element list, so missing the second set of images

In [4]: ants.Registration(fixed_image=['MNI152_T1.nii', 'MNI152_T2.nii'],
   ...:                   moving_image=['T1.nii', 'T2.nii'],
   ...:                   metric=['CC', 'MI'],
   ...:                   metric_weight=[1.0, 1.0],
   ...:                   radius_or_number_of_bins=[5, 32],
   ...:                   smoothing_sigmas=[[4, 2, 1, 0]]*2,
   ...:                   shrink_factors=[[8, 4, 2, 1]]*2,
   ...:                   transforms=['SyN']*2,
   ...:                   number_of_iterations=[[10, 10, 10, 10]]*2,
   ...:                   transform_parameters=[[0.5]]*2).cmdline
Out[4]: 'antsRegistration --collapse-linear-transforms-to-fixed-image-header 0 -
-collapse-output-transforms 0 --dimensionality 3 --interpolation Linear --output
 transform --transform SyN[ 0.5 ] --metric CC[ MNI152_T1.nii, T1.nii, 1, 5 ] --c
onvergence [ 10x10x10x10, 1e-06, 10 ] --smoothing-sigmas 4.0x2.0x1.0x0.0 --shrin
k-factors 8x4x2x1 --use-histogram-matching 1 --transform SyN[ 0.5 ] --metric MI[
 MNI152_T2.nii, T2.nii, 1, 32 ] --convergence [ 10x10x10x10, 1e-06, 10 ] --smoot
hing-sigmas 4.0x2.0x1.0x0.0 --shrink-factors 8x4x2x1 --use-histogram-matching 1 
--winsorize-image-intensities [ 0.0, 1.0 ]  --write-composite-transform 0'
# Duplicating all the parameters gets the second pair of images, but duplicates the
# corresponding options as well

[sgiavasis]

What does _formatMetric do, exactly?

[ischwabacher]

IIRC it's the helper function that's actually responsible for generating the part of the command line you're interested in (as in Out[2]). What I know about it comes from looking at the source for it and for the cmdline property, which calls it via _formatRegistration and _format_arg.

[sgiavasis]

I was digging around, and after some testing, it seems that the interface for the ANTS command handles this properly:

From nipype/registration.py, line 110:

def _image_metric_constructor(self):
retval = []
intensityBased = ['CC', 'MI', 'SMI', 'PR', 'SSD', 'MSQ']
pointSetBased = ['PSE', 'JTB']
for ii in range(len(self.inputs.moving_image)):
if self.inputs.metric[ii] in intensityBased:
retval.append('--image-metric %s[ %s, %s, %g, %d ]' % (self.inputs.metric[ii],
self.inputs.fixed_image[ii],self.inputs.moving_image[ii],self.inputs.metric_weight[ii],self.inputs.radius[ii]))
elif self.inputs.metric[ii] == pointSetBased:
pass
return ' '.join(retval)

Could this be adapted to the antsRegistration interface? I am trying a hack of this now, but it seems it would need a few things to be changed. For example, this function returns a full string unlike antsRegistration's formatMetric, which returns only the metric descriptions.

[djarecka]

I think this was fixed in #1187 and #1196, but please reopen if this is still an issue.