uncbiag / mermaid Goto Github PK

The inverse map output is wrong when the initial map is not the identity map. It can be corrected by manually set the initial inverse map.

Must the moving image and the reference image have the same spaing?

si.register_images(I0, I1, spacing, model_name='lddmm_shooting_map',
nr_of_iterations=100,
use_multi_scale=False,
visualize_step=5,
optimizer_name='sgd',
learning_rate=0.1,
rel_ftol=1e-7,
json_config_out_filename=('test2d_tst.json',
'test2d_tst_with_comments.json'),
params='test2d_tst.json',
recording_step=None)
I0 and I1 must have the same spacing?Must I0 and I0 be resampled to the same spacing before registration?

Accidently I tried to register an image to itself using the multi_scale optimizer, but the code breaks. It should return an identity map. (It works when single scale optimizer is used.)

It would be good to have an option of doing interpolation that is better than linear interpolation. For example, b-splines would be good. This would be useful in particularly also for the low-res parameterizations.

The dimensions are not correct after ifft, because of the real fft in CUDA (for the Fourier filtering).
Fixed currently with a quick hack for map_low_res_factor which makes sure the last dimension is even. Needs a more proper solution.

Line 16 that is loading the json file does not match the file's name and directory. Could you please update it to:
params.load_JSON('../test/json/svf_momentum_base_config.json')

In ImageIO.read() the spacing was flipped to match the order of numpy array dimensions. However, it was not flipped back when writing into an image file using ImageIO.write().

LNCC output will be out of the range [0,1] when the images have large area with zero value.

Steps to reproduce:

1. Generate Data:

params = pars.ParameterDict()
params['square_example_images']['len_s'] = 10
params['square_example_images']['len_l'] = 30
I0,I1,spacing = EG.CreateSquares(dim=3, add_noise_to_bg=False).create_image_pair(np.array([64, 64, 64]),params=params)

2. Register with LNCC similarity.

params['model']['registration_model']['type'] = "svf_vector_momentum_map" 
params['model']['registration_model']['similarity_measure']['type'] = "lncc"
opt = MO.SimpleSingleScaleRegistration(I0,
                                       I1,
                                       spacing,
                                       sz,
                                       params,
                                       compute_inverse_map=False)
opt.register()

The log shows as follows:
0-Tot: E=099.4140 | simE=099.4140 | regE=000.0000 | optParE=000.0000 | relF= n/a |
0-Img: E=099.4140 | simE=099.4140 | regE=000.0000 |

python setup.py install does not install the settings folder to the python library

Hi,

How can I force the register_image on CPU? By default, example_simple_interface.py seems always run on GPU, even if CUDA_ON was set to false.

Thank you

Hi,

How can write the registration result to a nifti file?

Thank you

The atlas builder runs svf registration in a for loop and uses model properties from the previous registration for the current registration. A function that clears the model beforehand is needed.

variable dim is not defined in line 692, unless self.squeeze_image is true in line 616.

In lines 45/46 the attribute 'smooth_scalar_field' needs to be updated to the new name 'smooth'.

I am looking for Cross model image registration for some of my projects and I am wondering will it support cross-modal images.

Currently, the energies are divided by the batch-size, so that the energies stay the same independent of the size of the batch. But this is not a good idea, because then one requires a different learning rate for different batch sizes.

Needs to be fixed (i.e., remove division by batch-size for all the energy terms), but maybe add a display option that also displays energy normalized by batch if desired.

Zenodo enables creation of an open source scientific software citation, complete with a DOI. This enable recognition of the work and reproducibility through identification of computational dependencies.

An GitHub README badge with the citation that is automatically updated when a GitHub Release is created can be generated as documented here:

https://guides.github.com/activities/citable-code/

I've been getting this message from pip:

DEPRECATION: mermaid 0.3.2 has a non-standard dependency specifier torch>=1.7torchvision. pip 23.3 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of mermaid or contact the author to suggest that they release a version with a conforming dependency specifiers. Discussion can be found at pypa/pip#12063

I used this code to train my own data set. My data set came from ADNI. I used 240 pieces as the training set, 70 pieces as the verification set, and 34 pieces as the test set.I did not change other parameters, but the result of training test is worse than Voxelmorph. Could you please give me some opinion analysis.

Maps are currently all in [-1,1]^d and ignore the spacing. This is problematic because the Gaussian kernels for regularization rely on this spacing.

Instead the maps should be in the range of the spatial extent of an image (for a given spacing).

Requires a change of the core interpolation functions to do this mapping (as these rely on the canonical [-1,1] mapping).

After I modified the value of example_img_len in the script example_minimal_registration_without_simple_interface.py, the result seems to be wrong, the mesh does not change after iteration, and converges to the tolerance in step 29. When the image data size becomes 640*360, do I need to modify other parameters?

Hi,
Thank you for sharing such brilliant codes.

Could you provide more detailed documents for implementation, such as reference papers?
I am a beginner to LDDMM. I have read several papers about LDDMM. But I can not figure out what method you used in the codes.

Thank you very much.

The upsampling and downsampling is unnatural. As it uses (current_size-1)/(desired_size-1). This is strange. If one has a 4x4 image with spacing 1/3 it will become a spacing of 1 one upsampled to 2x2 instead of 2/3. To properly support this an origin should be introduced.

Hello,

I was wondering if there was something to describe what the consensus model is actually doing?
Is there a paper that could be used as a reference as well?

@marcniethammer any advice (as the contributor) would be appreciated!

Thank you!

when i run the jupyter --example_simple_interface.
the test2d_tst.json can't find in the repository.
and there's an error
Could not open file = test2d_tst.json; ignoring request.

In class NCCSimilarity, def compute_similarity has input parameters I0, I1 but tries to compute ncc with cI0, cI1. One of them needs to be renamed.

Code needs to be adapted to work with the newer pytorch. Currently only runs with pytorch 0.3.x

For some reason this code crashes
testRegistrationGenericMultiscaleCustom.py

MySVFNet (
)
Registering new similarity measure mySSD in factory
Optimizing for at most 20 iterations
Warning: optimizer name = lbfgs_ls specified, but ignored since optimizer was set explicitly
Traceback (most recent call last):
File "testRegistrationGenericMultiscaleCustom.py", line 160, in
mo.optimize()
File "/afs/cs.unc.edu/home/mn/private/research/programming/git/mermaid/pyreg/multiscale_optimizer.py", line 625, in optimize
ssOpt.optimize()
File "/afs/cs.unc.edu/home/mn/private/research/programming/git/mermaid/pyreg/multiscale_optimizer.py", line 446, in optimize
self.optimizer_instance.step(self._closure)
File "/playpen/anaconda2/lib/python2.7/site-packages/torch/optim/adam.py", line 39, in step
loss = closure()
File "/afs/cs.unc.edu/home/mn/private/research/programming/git/mermaid/pyreg/multiscale_optimizer.py", line 327, in _closure
self.rec_phiWarped = self.model(self.identityMap, self.ISource)
File "/playpen/anaconda2/lib/python2.7/site-packages/torch/nn/modules/module.py", line 224, in call
result = self.forward(*input, **kwargs)
TypeError: forward() takes exactly 2 arguments (3 given)

Hello @marcniethammer,
Thank you for making the code available, really interesting work.

I was looking into some of the code, in particular for the STN in 3D. I noticed that you are moving the channels around, here:

I have also another question that is not about your code in particular but it is in general about the Pytorch grid_sample function. As you can see from my question , I am still not sure about how to correctly use grid_sample in case of registration. Have you solved this problem yourself?