Replace registration to use ANTsPy instead of Nipype #90

teresamg · 2022-09-26T21:24:23Z

Addresses #89

arokem · 2022-09-26T21:45:30Z

@oesteban : I don't see circleci running the tests on this PR. Do we need to change the circleci config somehow to run on PRs?

arokem · 2022-09-26T21:51:56Z

src/eddymotion/estimator.py

-                            terminal_output="file",
-                            from_file=pkg_fn(
-                                "eddymotion",
-                                f"config/dwi-to-{reg_target_type}_level{i_iter}.json",


One question to mull over is how do we pass this config to ants.registration? Right now, it's no longer being passed, and it uses the default settings of that function instead.

As per discussion over email, we offline run registration.cmdline and the massage the output CLI to replace image paths with ants images.

arokem · 2022-09-26T21:56:25Z

src/eddymotion/estimator.py

-                            ),
-                            fixed_image=str(fixed.absolute()),
-                            moving_image=str(moving.absolute()),
-                            **align_kwargs,


Maybe we can use the align_kwargs variable to pass the default values in.

I'm leaning towards just storing one YAML file with the command lines that the estimator can pick.

arokem · 2022-09-26T21:58:39Z

src/eddymotion/estimator.py

-                            predicted,
-                            dwdata.affine,
-                            fixed,
-                            clip=reg_target_type == "dwi",


Looks like the clipping depends on the type of the fixed image (whether it's a b=0 or a diffusion-weighted volume). I guess we need to reincorporate clipping somehow.

Yes, it's great you guys are bringing this up to the forefront.

Basically, the question here is that I have come to the conclusion that, even though ANTs has histogram equalization, it really needs help from the outside.

In general, there are a bunch of problems/situations where ants doesn't work very well:

Presence of negative values (unless they are meaningufuly correct - e.g., if you are registering gradients)

Presence of long tails (blood vessels on T1w images typically add very bright, but few, pixels which extend the histogram)

Related to the former: concentration of quartiles at the beginning of the histogram. In other words, the cumulative distribution function goes up extremely quickly and then finds an elbow. This is particularly bad with the reference image.

ANTs prefers distribution without large range (e.g., normalize the intensities to fit in a uint8 or int16).

This clipping tries that, but I haven't played a lot to set something robust and relatively data-driven.

oesteban

solid progress... how can I help more?

oesteban · 2022-10-06T11:58:57Z

src/eddymotion/estimator.py

-                            predicted,
-                            dwdata.affine,
-                            fixed,
-                            clip=reg_target_type == "dwi",


Yes, it's great you guys are bringing this up to the forefront.

Basically, the question here is that I have come to the conclusion that, even though ANTs has histogram equalization, it really needs help from the outside.

In general, there are a bunch of problems/situations where ants doesn't work very well:

Presence of negative values (unless they are meaningufuly correct - e.g., if you are registering gradients)

Presence of long tails (blood vessels on T1w images typically add very bright, but few, pixels which extend the histogram)

Related to the former: concentration of quartiles at the beginning of the histogram. In other words, the cumulative distribution function goes up extremely quickly and then finds an elbow. This is particularly bad with the reference image.

ANTs prefers distribution without large range (e.g., normalize the intensities to fit in a uint8 or int16).

This clipping tries that, but I haven't played a lot to set something robust and relatively data-driven.

oesteban · 2022-10-06T11:59:41Z

src/eddymotion/estimator.py

-                            terminal_output="file",
-                            from_file=pkg_fn(
-                                "eddymotion",
-                                f"config/dwi-to-{reg_target_type}_level{i_iter}.json",


As per discussion over email, we offline run registration.cmdline and the massage the output CLI to replace image paths with ants images.

oesteban · 2022-10-06T12:00:36Z

src/eddymotion/estimator.py

-                            ),
-                            fixed_image=str(fixed.absolute()),
-                            moving_image=str(moving.absolute()),
-                            **align_kwargs,


I'm leaning towards just storing one YAML file with the command lines that the estimator can pick.

oesteban · 2022-10-06T12:07:02Z

src/eddymotion/utils.py

+    if isinstance(val, str):
+        if Path(val).exists() and (PurePosixPath(val).suffix == '.nii.gz' or PurePosixPath(val).suffix == '.nii'):
+            return ants.from_nibabel(nb.load(val))


unnecessary checks. Also, if there's a function to import from nibabel, we don't need to impose ".nii[.gz]" for the extension.

Suggested change

if isinstance(val, str):

if Path(val).exists() and (PurePosixPath(val).suffix == '.nii.gz' or PurePosixPath(val).suffix == '.nii'):

return ants.from_nibabel(nb.load(val))

val = Path(val)

if not val.exists():

raise FileNotFoundError(val)

return ants.from_nibabel(nb.load(val))

oesteban · 2022-10-06T12:08:35Z

test/test_utils.py

+import ants
+
+
+def test_antsimage_from_path():


Suggested change

def test_antsimage_from_path():

def test_antsimage_from_path(datadir, tmp_path):

oesteban · 2022-10-06T12:10:29Z

test/test_utils.py

+    # Generate and save test data to file
+    # Send through antsimage_from_path
+    # Check if ANTsImage, same as original data
+    assert isinstance(result, ants.ANTsImage)


Suggested change

assert isinstance(result, ants.ANTsImage)

fixed = tmp_path / "b0.nii.gz"

dwdata = DWI.from_filename(datadir / "dwi.h5")

b0nii = nb.Nifti1Image(dwdata.bzero, dwdata.affine, None)

b0nii.to_filename(fixed)

assert isinstance(antsimage_from_path(fixed), ants.ANTsImage)

oesteban · 2022-10-06T12:12:22Z

src/eddymotion/estimator.py

-                        # execute ants command line
-                        result = registration.run(cwd=str(tmpdir)).outputs
+                        #registration_kwargs = ants_utils._int_antsProcessArguments(registration_kwargs)
+                        registration_ants = ants.registration(**registration_kwargs)


Suggested change

registration_ants = ants.registration(**registration_kwargs)

ants_cli = ants.registration(**registration_kwargs).cmline

# Now split ants_cli and replace paths with image objects

# this is a bit more convoluted because some paths will come between brackets and commas

args = [load_if_image(token) for token in ants_cli.split(" ")]

eventually we will get rid of nipype, but this could be useful for now

Could you show us where the .cmline is coming from? Are you suggesting that we revert back to the nipype Registration() call instead of the current ants call?

arokem · 2022-11-24T02:32:17Z

After some more detailed profiling, we came to the conclusion that this does not provide speedup and potentially reduces flexibility relative to the nipype interface (because the Python API does not expose the full API of the CLI). Unless someone objects, I suggest that we close this PR.

teresamg added 3 commits September 26, 2022 14:19

update Nipype dependency

4f3a10b

replace registration with ANTsPy

bc5e447

flake8

45d12e8

Remove cache to tmp files

34ba5a5

arokem reviewed Sep 26, 2022

View reviewed changes

Added conversion from path to ANTsImage

92e02c5

oesteban reviewed Oct 6, 2022

View reviewed changes

Debugging

7586fa3

dPys mentioned this pull request Dec 13, 2022

Fix masking add composability #114

Closed

arokem closed this Dec 15, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Replace registration to use ANTsPy instead of Nipype #90

Replace registration to use ANTsPy instead of Nipype #90

teresamg commented Sep 26, 2022

arokem commented Sep 26, 2022

arokem Sep 26, 2022

oesteban Oct 6, 2022

arokem Sep 26, 2022

oesteban Oct 6, 2022

arokem Sep 26, 2022

oesteban Oct 6, 2022

oesteban left a comment

oesteban Oct 6, 2022

oesteban Oct 6, 2022

oesteban Oct 6, 2022

oesteban Oct 6, 2022

oesteban Oct 6, 2022

oesteban Oct 6, 2022

oesteban Oct 6, 2022

oesteban Oct 6, 2022

teresamg Oct 6, 2022

arokem commented Nov 24, 2022

	def test_antsimage_from_path():
	def test_antsimage_from_path(datadir, tmp_path):

-    assert isinstance(result, ants.ANTsImage)
+    fixed = tmp_path / "b0.nii.gz"
+    dwdata = DWI.from_filename(datadir / "dwi.h5")
+    b0nii = nb.Nifti1Image(dwdata.bzero, dwdata.affine, None)
+    b0nii.to_filename(fixed)
+    assert isinstance(antsimage_from_path(fixed), ants.ANTsImage)

-                        registration_ants = ants.registration(**registration_kwargs)
+                        ants_cli = ants.registration(**registration_kwargs).cmline
+                        # Now split ants_cli and replace paths with image objects
+                        # this is a bit more convoluted because some paths will come between brackets and commas
+                        args = [load_if_image(token) for token in ants_cli.split(" ")]

Replace registration to use ANTsPy instead of Nipype #90

Replace registration to use ANTsPy instead of Nipype #90

Conversation

teresamg commented Sep 26, 2022

arokem commented Sep 26, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

oesteban left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

arokem commented Nov 24, 2022