Integration and validation tests with b=0 only

[oesteban]

Two low-hanging fruit tests using only a b=0:

• Validation:
  • Unit tests of the TrivialB0Model implementation
  • Unit tests of ANTs' config for b=0.
• Integration:
  • Introduce a series of transforms (e.g., copy them from another dataset) and resample the b=0 through that motion series. Then run the whole estimator and check the proximity of the estimated transforms.

[oesteban]

Okay, I've tried to set up some basics, and implemented a unit-test for the TrivialB0Model sketching what we intend to do here:

eddymotion/eddymotion/tests/test_model.py

Lines 29 to 42 in 152e0d4

	def test_trivial_model():
	"""Check the implementation of the trivial B0 model."""
	# Should not allow initialization without a B0
	with pytest.raises(ValueError):
	model.TrivialB0Model(gtab=np.eye(4))
	_S0 = np.random.normal(size=(10, 10, 10))
	tmodel = model.TrivialB0Model(gtab=np.eye(4), S0=_S0)
	assert tmodel.fit() is None
	assert np.all(_S0 == tmodel.predict((1, 0, 0)))

Unit tests of ANTs' config for b=0.

I see this implemented in test_estimator.py. To do so:

• Upload an example B0 map (3D)
• Apply random (and small) rotations and translations on the image. These can be generated with nibabel.eulerangles.euler2mat (e.g., https://github.com/poldracklab/nitransforms/blob/master/nitransforms/tests/test_linear.py#L107). For the resampling, I would use nitransforms.linear.Affine.resample().
• Run the registration interface loading the parameters from config
• Check that the error of the affine is below some threshold when transforming points scattered across the volume.

Introduce a series of transforms (e.g., copy them from another dataset) and resample the b=0 through that motion series. Then run the whole estimator and check the proximity of the estimated transforms.

I see this implemented in test_integration.py. To do so:

• Upload an example B0 map (3D) - or reuse from the previous test
• Generate a series of moved B0's with the realignment matrices from another, real dataset.
• Initiate the TrivialB0Model with this B0 map and feed the dwi object with the moved data (simulate perhaps several transforms).

[oesteban]

Hi there! I've uploaded some data under https://github.com/nipreps/eddymotion/tree/main/eddymotion/tests/data, generated by @dPys with FiberFox.

[sebastientourbier]

@oesteban I am progressing! For now, I managed to:

• extract the b0 volume from the sub-01_dwi.nii.gz provided by Céline.
• run AFNI 3dvolreg on the fmri of the first run, first subject of ds000005 (240 time frames) on openneuro with the middle frame as reference. This gives me a text file where one line corresponds to the motion parameters for one time frame.
• make 20 copies of the b0 where I applied the motion parameters from the last 20 frames (with nitransforms), saved in temporary niftis.
• save a unique 4D nifti image that concatenates the original and the moving b0s

I haven't opened a PR yet but here is a quick preview of the motion introduced

What do think?

[dPys]

This looks so awesome.