ants Registration parameters

[binarybottle]

I would like to obtain an affine (as well as nonlinear) transform from each image volume to a template (yet to be supplied; currently set to the MNI152 template). For this I have included nipype's antsRegistration wrapper: http://www.mit.edu/~satra/nipype-nightly/interfaces/generated/nipype.interfaces.ants.registration.html#examples,

but am having trouble setting the parameters according to: https://github.com/stnava/ANTs/blob/master/Scripts/newAntsExample.sh

in particular, i don't know how to set multiple metrics for syn: https://github.com/binarybottle/mindboggle/blob/master/mindboggle/pipeline.py#L637

and i have not been successful saving the affine transform.

[satra]

the nipype interface doesn't currently allow multiple metrics for a transform. i'm currently working on a cerebellar alignment evaluation and will post back what parameters work well for that.

also nick advised using the metrics in antsCorticalThickness.sh, so i tried those:

reg = Registration()
reg._cmd = '/software/ANTS/versions/130509/antsRegistration'
reg.inputs.transforms = ['Rigid', 'Affine', 'SyN']
reg.inputs.transform_parameters = [(0.1,), (0.1,), (0.1, 3.0, 0.0)]
reg.inputs.number_of_iterations = [[1000,500,250,100]]*2 + [[100,100,70,20]]
reg.inputs.dimension = 3
reg.inputs.write_composite_transform = True
reg.inputs.collapse_output_transforms = True
reg.inputs.metric = ['MI']*2 + ['CC']
reg.inputs.metric_weight = [1]*3 # Default (value ignored currently by ANTs)
reg.inputs.radius_or_number_of_bins = [32]*2 + [4]
reg.inputs.sampling_strategy = ['Regular']*2 + [None]
reg.inputs.sampling_percentage = [0.25]*2 + [None]
reg.inputs.convergence_threshold = [1.e-8]*2 + [1e-9]
reg.inputs.convergence_window_size = [10]*2 + [15]
reg.inputs.smoothing_sigmas = [[3,2,1,0]]*3
reg.inputs.sigma_units = ['mm']*3
reg.inputs.shrink_factors = [[8,4,2,1]]*2 + [[6,4,2,1]]
reg.inputs.use_estimate_learning_rate_once = [True, True, True]
reg.inputs.use_histogram_matching = [False]*2 + [True] # This is the default
reg.inputs.initial_moving_transform_com = True
reg.inputs.output_warped_image = True
reg.inputs.winsorize_lower_quantile = 0.01
reg.inputs.winsorize_upper_quantile = 0.99

takes about 6 hours with:

regnode = Node(reg, name='ANTSreg')
regnode.inputs.num_threads = 2

i can't get it to use more than 2 threads.

[binarybottle]

thank you, satra. these do differ quite a bit. i heard the new itk4 version would take longer, but 6 hours? well, i think it makes sense for now to use the above to compute and apply an affine transform to standard space, and worry about nonlinear volume deformation to label subcortical regions a little later. will the following save an affine transform?:

    reg = Node(name='Register_template', interface = Registration())
    mbFlow.add_nodes([reg])
    if do_input_nifti:
        mbFlow.connect([(niftiVol, reg, [('nifti_volume','moving_image')])])
    else:
        mbFlow.connect([(mgh2nifti, reg, [('out_file','moving_image')])])
    reg.inputs.fixed_image = [ants_template]
    reg.inputs.num_threads = 2
    reg.inputs.transforms = ['Rigid', 'Affine']
    reg.inputs.transform_parameters = [(0.1,), (0.1,)]
    reg.inputs.number_of_iterations = [[1000,500,250,100]]*2
    reg.inputs.dimension = 3
    reg.inputs.write_composite_transform = True
    reg.inputs.collapse_output_transforms = True
    reg.inputs.metric = ['MI']*2
    reg.inputs.metric_weight = [1]*2
    reg.inputs.radius_or_number_of_bins = [32]*2
    reg.inputs.sampling_strategy = ['Regular']*2
    reg.inputs.sampling_percentage = [0.25]*2
    reg.inputs.convergence_threshold = [1.e-8]*2
    reg.inputs.convergence_window_size = [10]*2
    reg.inputs.smoothing_sigmas = [[3,2,1,0]]*2
    reg.inputs.sigma_units = ['mm']*2
    reg.inputs.shrink_factors = [[8,4,2,1]]*2
    reg.inputs.use_estimate_learning_rate_once = [True, True]
    reg.inputs.use_histogram_matching = [False]*2
    reg.inputs.output_warped_image = True
    reg.inputs.winsorize_lower_quantile = 0.01
    reg.inputs.winsorize_upper_quantile = 0.99
    # Nipype output:
    reg.inputs.write_composite_transform = True
    mbFlow.connect([(reg, Sink,
                     [('composite_transform', 'transforms.@affine')])])

[binarybottle]

it seems to work only if i remove the line: reg.inputs.sigma_units = ['mm']*2

otherwise i get the error: "traits.trait_errors.TraitError: Cannot set the undefined 'sigma_units' attribute of a 'RegistrationInputSpec' object."

[satra]

you don't have nipype 0.8 - that was added in the current release.

[binarybottle]

thank you -- i just updated and am running, but i don't know what the output affine transform file name will be, so how do i connect it to the next node in?:

https://github.com/binarybottle/mindboggle/blob/master/mindboggle/pipeline.py#L1191

[satra]

since you are only using the affine transforms you can do:

wf.connect(regnode, 'composite_transform', at, 'transforms')

if in the future you only want to apply the affine part of a complex transform you would do:

wf.connect(regnode, ('forward_transforms', pickfirst), atl, 'transforms')

where

pickfirst = lambda x: x[:1]

[binarybottle]

thank you, satra. the resulting *GenericAffine.mat transform file is not a text file. is this because i chose write_composite_transform?: https://github.com/binarybottle/mindboggle/blob/master/mindboggle/pipeline.py#L646 how do i get it to output the transform as a text file?

also, i am now getting the following error (i got the first file but not the second): FileNotFoundError: File/Directory '['/Users/arno/Desktop/output/workspace/Mindboggle/_hemi_rh_subject_Twins-2-1/antsRegister_standard/affine_0GenericAffine.mat', '/Users/arno/Desktop/output/workspace/Mindboggle/_hemi_rh_subject_Twins-2-1/antsRegister_standard/affine_1InverseWarp.nii.gz']' not found for Registration output 'reverse_transforms'. Interface Registration failed to run.

[satra]

why do you need a text file? ants now stores the transforms as mat files or hdf5 files.

regarding the reverse transforms, how recent a version of ants are you using?

[binarybottle]

i need a text file for flexibility. for example, i am using it in the form of a numpy array to transform the coordinates within a vtk surface mesh.

i installed ants via github on april 26th.

[satra]

regarding the text file, you'll have to bring it up with the ants folks

can you paste the contents of the folder:

/Users/arno/Desktop/output/workspace/Mindboggle/_hemi_rh_subject_Twins-2-1/antsRegister_standard/

or put it in dropbox temporarily so i can take a look.

[binarybottle]

​i placed the directory in the dropbox.​

perhaps then i will revert to using the ANTS command instead of antsRegistration to get the text itk transform. is the ANTS command wrapped in nipype as well?

[satra]

it is wrapped under legacy.

but you can read the .mat file in python, so why do you need the text file?

also the reverse_transform error is a bug in nipype when no nonlinear transforms are specified. i won't have time to fix that immediately.

[binarybottle]

how do i read that .mat file in python?

until the bug is fixed i'll use flirt and read its -omat transform using np.loadtxt, or skip ahead to computing a nonlinear transform with antsRegistration.

[satra]

from scipy.io import loadmat

[binarybottle]

regarding a text transform option, nick tustison wrote to me:

"We currently don't have that option. There is an external program 'ConvertTransformFile' which will convert the binary format to text format."

[binarybottle]

unfortunately, i get the following error when i try to load the ants transform in our dropbox:

from scipy.io import loadmat transform = loadmat('/drop/MB/data/arno/mri/affine_to_template.mat') ValueError: Unknown mat file type, version 57, 56

[binarybottle]

i'll just go ahead and use the "ConvertTransformFile" program for now.