ENH: Add ConvertToITKTransform example

[tbirdso]

Adds a Jupyter Notebook example demonstrating how to convert Elastix registration transform results into standard ITK transform format.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

dzenanz commented on 2023-08-28T17:13:48Z
----------------------------------------------------------------

Line #1.    itk_composite_transform = itk.CompositeTransform[itk.D, 2].New()

Why not use this, more convenient form:

py

comb_transform = registration_method.GetCombinationTransform()

itk_composite_transform = registration_method.ConvertToItkTransform(comb_transform)

thewtex commented on 2023-08-28T20:10:18Z
----------------------------------------------------------------

+1

[dzenanz]

comb_transform = registration_method.GetCombinationTransform()
itk_composite_transform = registration_method.ConvertToItkTransform(comb_transform)

ReviewNB messed up formatting.

[N-Dekker]

Thanks for this PR @tbirdso!

ConvertToItkTransform does indeed support converting a "single" transform (as returned by GetNthTransform) as well as a "combination transform" (returned by GetCombinationTransform()). It converts a combination transform to an ITK CompositeTransform (although it will only do so when the elastix combination transform did use composition, rather than addition, to combine the transforms).

[tbirdso]

@dzenanz @N-Dekker Thanks for the notes! I will add a cell to demonstrate GetCombinationTransform.

In practice I am not getting an itk.CompositeTransform from GetCombinationTransform, but instead some apparent Advanced representation of the last transform layer. Is this expected?

For the given example:

> print(registration_method.GetCombinationTransform())
AdvancedBSplineTransform (000001B7546E2960)
  RTTI typeinfo:   class elastix::AdvancedBSplineTransform<class elastix::ElastixTemplate<class itk::Image<float,2>,class itk::Image<float,2> > >
  Reference Count: 14
  Modified Time: 121552
  Debug: Off
  Object Name: 
  Observers: 
    none

The transform cannot be cast to a itk.CompositeTransform:

> itk.CompositeTransform[itk.D,2].cast(registration_method.GetCombinationTransform())
RuntimeError                              Traceback (most recent call last)
Cell In[17], line 1
----> 1 itk.CompositeTransform[itk.D,2].cast(registration_method.GetCombinationTransform())

RuntimeError: bad cast

However, calling TransformPoint on our manually composed itk.CompositeTransform and the result of GetCombinationTransform return the same results. So, I believe that internally represent the same conceptual state.

> print(registration_method.GetCombinationTransform().TransformPoint([0,0]))
itkPointD2 ([-0.944895, 1.38383])
> print(itk_composite_transform.TransformPoint([0,0]))
itkPointD2 ([-0.944895, 1.38383])

For demonstration purposes, I believe it is easier to understand what is happening in the itk.CompositeTransform where each transform stage is clearly enumerated.

EDIT: As another (perhaps more important) distinction, I am not able to serialize the result of GetCombinationTransform, whereas I am able to write/read the manually composed itk.CompositeTransform with itk.transformwrite/itk.transformread.

> itk.transformwrite([registration_method.GetCombinationTransform()],'testout_advanced.h5',compression=False)
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[16], line 1
----> 1 itk.transformwrite([registration_method.GetCombinationTransform()],'testout_advanced.h5',compression=False)

File C:\venvs\venv-itk\lib\site-packages\itk\support\extras.py:1447, in transformwrite(transforms, filename, compression)
   1445 for transform in transforms:
   1446     writer.AddTransform(transform)
-> 1447 writer.Update()

RuntimeError: C:\P\IPP\ITK-source\ITK\Modules\IO\TransformHDF5\src\itkHDF5TransformIO.cxx:462:
ITK ERROR: HDF5TransformIOTemplate(00000201B49C4B80): H5Fcreate failed

> itk.transformwrite([itk_composite_transform],'testout_composite_transform.h5',compression=False)
> itk.transformread('testout_composite_transform.h5')
[<itk.itkCompositeTransformPython.itkCompositeTransformD2; proxy of <Swig Object of type 'itkCompositeTransformD2 *' at 0x00000223FBBE81B0> >]

[dzenanz]

You need to convert the resulting thing into an ITK transform, like this:
itk_transform = registration_method.ConvertToItkTransform(registration_method.GetCombinationTransform())

[tbirdso]

You need to convert the resulting thing into an ITK transform, like this:

Thanks @dzenanz , this is the step I was missing. I will revise the notebook to prefer this approach 👍

[dzenanz]

Since you already did the work, you could leave the current thing as an alternative approach.

[tbirdso]

To avoid confusion between two approaches I've opted to remove the previous approach manually composing an itk.CompositeTransform from individual stages and instead use the expected type information for validation.

[thewtex]

+1

---

View entire conversation on ReviewNB

[thewtex]

Love this!! 📖 ❤️