Problem in Output of ants Registration interfaced via Nipype : FileNotFoundError output

[BastiaanvanTol]

Hello all,

I want to interface ANTs in Python using Nipype for cross-modal registration of MRI and PET images. I have compiled ANTs as suggested, and Nipype is working as well. I am able to import the Registration module via from nipype.interfaces.ants import Registration

However, when I define the transformation parameters and I try to run the registration, I have not been able to deal with a FileNotFoundError that keeps appearing. It says:

No such file or directory '/Users/bastiaanvantol/Desktop/Tau_PET/output_ants/output_Composite.h5' for output 'composite_transform' of a Registration interface

Can anybody help me where this problem may lie, and suggestions how I can fix this?
Any help would be greatly appreciated, thanks!

---

Here's is the example code from the Nipype website that I use to try out interfacing:

reg = Registration()
reg.inputs.fixed_image = pet_img
reg.inputs.moving_image = mr_img
reg.inputs.output_transform_prefix = "/Users/bastiaanvantol/Desktop/Tau_PET/output_ants/output_"
reg.inputs.transforms = ['Affine', 'SyN']
reg.inputs.transform_parameters = [(2.0,), (0.25, 3.0, 0.0)]
reg.inputs.number_of_iterations = [[1500, 200], [100, 50, 30]]
reg.inputs.dimension = 3
reg.inputs.write_composite_transform = True
reg.inputs.collapse_output_transforms = False
reg.inputs.initialize_transforms_per_stage = False
reg.inputs.metric = ['Mattes']*2
reg.inputs.metric_weight = [1]*2 # Default (value ignored currently by ANTs)
reg.inputs.radius_or_number_of_bins = [32]*2
reg.inputs.sampling_strategy = ['Random', None]
reg.inputs.sampling_percentage = [0.05, None]
reg.inputs.convergence_threshold = [1.e-8, 1.e-9]
reg.inputs.convergence_window_size = [20]*2
reg.inputs.smoothing_sigmas = [[1,0], [2,1,0]]
reg.inputs.sigma_units = ['vox'] * 2
reg.inputs.shrink_factors = [[2,1], [3,2,1]]
reg.inputs.use_estimate_learning_rate_once = [True, True]
reg.inputs.use_histogram_matching = [True, True] # This is the default
reg.inputs.output_warped_image = '/Users/bastiaanvantol/Desktop/Tau_PET/output_ants/warped_image.nii'
reg.cmdline
reg.run()  

[cookpa]

Can you please add all the output including reg.cmdline - so we can see what the ANTs command was

[BastiaanvanTol]

The output nifti image looks like this, I will add it zipped for visualisation purposes as well:

warped_image.nii.zip

[cookpa]

The inputs would be more informative than the output.

These kinds of issues are usually problems with the initial lack of overlap of physical space. The first thing antsRegistration does is to align the centers of mass of the images. If that puts them too far apart, the optimization cannot converge.

You can preview this interactively using ITK-SNAP's registration tool. Load the fixed image then add the moving image as an overlay, then go to registration and try to align COM.

[BastiaanvanTol]

Thank you Philip. Indeed the two images are far apart when I overlay them.

When I align them COM in ITK-SNAP, they fit very nicely on top of each other, thus solving the problem I had with my antsRegistration output. This is very good news! But it also means that ITK-SNAP did not have any problems with the COM alignment. Naturally I would like to have this step happen in my python pipeline as well, because I will be doing the coregistration for many images.

Is there a way for me to separately align the images COM with ANTs before I do the Rigid registration?

[cookpa]

Yes you can use antsAI

[BastiaanvanTol]

Thanks a lot for suggesting this function; I did not know I could use antsAI for this purpose, and it has been working well. Thank you!

[gdevenyi]

You didn't use an initial_moving_transform in your nipype call.

[BastiaanvanTol]

Thank you Gabriel! I defined initial_moving_transform using the trans.mat file coming from antsAI, and that solved my issue.

[cookpa]

There's also an option initial_moving_transform_com, which will do the COM alignment for you without the need to call antsAI.

[BastiaanvanTol]

Aha I see. Even more simple. I also found that using registrationSyNQuick does COM correction automatically, and for both my rigid & rigid-affine-SyN transformations, it performs very well. So, I think for my purposes, this is working great with increased computational efficiency.

[cookpa]

That's correct, in antsRegistration the option is -r [ fixed, moving, 1 ] and is done in the ants scripts. If omitted, it is assumed that the physical spaces of the images already overlap.