Reproducibility of the transformation

[maeinss]

Hi, I am trying to save a RandomElasticDeformation transformation such that it can be applied for future use. The problem I have is I am unable to find a way to do so. Here is a minimal reproducible example of a cube where I tried to see the shape of the deformed cube within a few iterations. (For the given example my aim is to get the same deformed structure with each iteration) Thank you.

  import numpy as np
  import torch
  import torchio as tio
  import SimpleITK as sitk

  #%% trial3D object
  trialCube = np.zeros((60, 60, 60))
  cube_size = 30 
  x_start = (trialCube.shape[0] - cube_size) // 2
  y_start = (trialCube.shape[1] - cube_size) // 2
  z_start = (trialCube.shape[2] - cube_size) // 2
  trialCube[x_start:x_start+cube_size, y_start:y_start+cube_size, z_start:z_start+cube_size] = 1
  
  #%% transformation
  seed = 1234
  torch.manual_seed(seed)
  transformation = tio.RandomElasticDeformation(
  num_control_points =12, locked_borders=0, 
                                      image_interpolation = "nearest", label_interpolation = "nearest", max_displacement = 7
                                      )
  
  
  #%% created subject
  subject = tio.Subject(
      image = tio.ScalarImage(tensor = np.array([trialCube]))
      )
  subject.plot()

  #%% transformed Matrix and Plotting
  for __ in range(3):
      transformed = transformation(subject)
      transformed.plot()

[romainVala]

hi
well, a random transform will by definition give a different (random) set of parameter
beside here you transform the original subject 3 times

subject= transformation(subject)

if you want to assign the same deformation at each step then you have to use

subject.history and setup and torchio.ElasticDeformation

[maeinss]

Thank you romainVala, could you help me set up a elastic deformation with the same example, and use subject history to reproduce. Also, is there any way I can set up control points in the regions I want. I apologize I am relatively new to this but it is a very nice tool.
Update:
I tried changing the transformation to RandomElastic, but now I have an error doing so. First, it will throw error if I put a single value for control_points or max_displacement so put the values in standard tuple form.

transformation = tio.ElasticDeformation(control_points =(2,2,2),
                              image_interpolation="nearest", label_interpolation="nearest", max_displacement = (2,2,2))

using this trasformation on the subject I get this error now: AttributeError: 'tuple' object has no attribute 'copy' which comes from the get_bspline_transform line 241.

[romainVala]

suj_elast = transformation(subject)
suj_elast.plot()

used_transformed = suj_elast.history[0]

for __ in range(2):
  suj_elast = used_transformed(suj_elast)
  suj_elast.plot()

history contains the choosen transform (after random parameter choice)
you can set up the transform from the parameter (so you'll understand how to play with it)

used_transformed = tio.ElasticDeformation(control_points=suj_elast.history[0].control_points, max_displacement=suj_elast.history[0].max_displacement)

so control point is not a tuple, but a grid

[maeinss]

Thank you for clarifying, one last thing here for the control_points. I am unable to figure out what type of grid structure are we expecting for the control_points. In this Class description of ElasticDeformation, control points doesn't have description apart from it's a ndarray, thus I assumed it's similar to the what we have from the the get_params function described earlier. But It doesn't work either. Here are the steps:

class ElasticDeformation(SpatialTransform):
r"""Apply dense elastic deformation.

Args:
    control_points:
    max_displacement:
    image_interpolation: See :ref:`Interpolation`.
    label_interpolation: See :ref:`Interpolation`.
    **kwargs: See :class:`~torchio.transforms.Transform` for additional
        keyword arguments.
"""

The get_params function was used to get control_points grid for apply_transforms

def get_params(
    num_control_points,
    max_displacement,
    num_locked_borders: int,
) -> np.ndarray:
    grid_shape = num_control_points
    num_dimensions = 3
    coarse_field = torch.rand(*grid_shape, num_dimensions)  # [0, 1)
    coarse_field -= 0.5  # [-0.5, 0.5)
    coarse_field *= 2  # [-1, 1]
    for dimension in range(3):
        # [-max_displacement, max_displacement)
        coarse_field[..., dimension] *= max_displacement[dimension]

    # Set displacement to 0 at the borders
    for i in range(num_locked_borders):
        coarse_field[i, :] = 0
        coarse_field[-1 - i, :] = 0
        coarse_field[:, i] = 0
        coarse_field[:, -1 - i] = 0

    return coarse_field.numpy()
def apply_transform(self, subject: Subject) -> Subject:
    subject.check_consistent_spatial_shape()
    control_points = self.get_params(
        self.num_control_points,  # type: ignore[arg-type]
        self.max_displacement,  # type: ignore[arg-type]
        self.num_locked_borders,
    )

So now I define my control points as from there and try to use it on the transformation. I still get error.

ctrlPnts = get_params((1,1,1), (2,2,2), 0)  
transformation = tio.ElasticDeformation(control_points =val,
                          image_interpolation="nearest", label_interpolation="nearest", max_displacement = 2)

suj_elast = transformation(subject) 

Thank you for your help

[romainVala]

I do not get any error ...
may be you should replace
tio.ElasticDeformation(control_points =val,
by
tio.ElasticDeformation(control_points =ctrlPnts,

[maeinss]

Yes, it works now. Thank you. However, ability to specify the control points positions in a structure would be quite a help if there is any way for that. I appreciate your help

[fepegar]

What do you mean by "in a structure"?

[maeinss]

With structure I meant the 3D volume. In this example it's the trialCube. Let's say I want to put 7 control points in the certain index locations of the trialCube. At this time, I only have an option to select the number of the control points but not the position of those points about which the deformation will take place. Thank you.

[fepegar]

I might be wrong, but I don't think that's possible using SimpleITK.

[maeinss]

I understand, thank you for your help.