[ENH] Template-based masking of EPI boldrefs #1321

oesteban · 2018-10-15T18:49:01Z

Changes proposed in this pull request

This PR explores template-based masking of boldrefs to initialize the skull-stripping workflow (see issues related on #1000).

This PR is also highly related to #1050.

Documentation that should be reviewed

This is WIP, will update documentation when it is proven to work out.

oesteban · 2018-10-15T23:12:26Z

So these are the results:

I still need to investigate what happened here: https://2488-53608443-gh.circle-artifacts.com/0/tmp/data/reports/ds000216/sub-03_task-rest_echo-1_bold_masks.svg

oesteban · 2018-10-15T23:15:12Z

The problem with echo 1 of ds216 seems to have been introduced here (this PR vs latest master)

oesteban · 2018-10-16T04:22:03Z

The weird result in https://2488-53608443-gh.circle-artifacts.com/0/tmp/data/reports/ds000216/sub-03_task-rest_echo-1_bold_masks.svg does not happen when I test locally (???)

oesteban · 2018-10-16T14:32:58Z

Seems like the weird problem on ds216 is now fixed. At this point I would ask: do you think it is maybe worth changing the whole algorithm to the following?:

Initialization: ANTs' AI, using the template as target and the boldref as moving
Gross mask, applyTransform on the inverted transform of 1)
N4 on the boldref with the gross mask
First registration: very fast refinement (Rigid) of the gross mask using the template as target and the boldref as moving and initialized with 1)
N4 on the boldref with the less-gross mask
Second registration (Affine) initialized with 4)
Remaining boldref refinement with the final mask

effigies · 2018-10-16T16:18:08Z

Let's test on the subject from #1181, as well.

oesteban · 2018-10-16T23:59:18Z

Still testing on additional datasets, but this seems ready for review.
Artifacts in the latest commit without [skip ci] -> https://circleci.com/gh/oesteban/fmriprep/2525#artifacts/containers/0

Refs. #1000, #1050, #1321. Also includes a new filter of branches so builds other than tests are skipped if the name of the branch starts with ``tests?/``.

zhifangy · 2018-10-18T03:24:02Z

Hi, @oesteban. I did a quick test using this PR on my previous failed data. Generally, it gives pretty good results. But I think there are several issues we need to address.

The pre_mask's header should be fixed before N4 correction. I encountered Inputs do not occupy the same physical space! error on one of my test image.
The N4 correction has no effect on some of the unwrapped images (eg. sub-26 rest) in bold_bold_trans_wf. I tried to run N4 correction manually on the ref_image_valid.nii.gz and it yielded the same incorrect result.
What if the BOLD image only covers part of brain volume? I think the initial affine transformation would generate a quite large mask for N4 correction. I think we should include some partial volume scan (especially for scans only covers 30-40 slices parallels to the long axis of hippocampus) in the test battery to see how well the method works.

oesteban · 2018-10-18T16:16:01Z

Thanks a lot for your precious feedback! @ZhifangYe

The pre_mask's header should be fixed before N4 correction. I encountered Inputs do not occupy the same physical space! error on one of my test image.

Could you provide more information about the error (full trace) or even the dataset (privately is fine) where this is happening? This is a new caching procedure we are testing and it does not surprise me you got this kind of problem.

The N4 correction has no effect on some of the unwrapped images (eg. sub-26 rest) in bold_bold_trans_wf. I tried to run N4 correction manually on the ref_image_valid.nii.gz and it yielded the same incorrect result.

Cool, sub-26_task-rest is now within the regression tests, so I'll work on this one particularly.

What if the BOLD image only covers part of brain volume? I think the initial affine transformation would generate a quite large mask for N4 correction. I think we should include some partial volume scan (especially for scans only covers 30-40 slices parallels to the long axis of hippocampus) in the test battery to see how well the method works.

This question is also very pertinent. I'll dig up some partial FoV datasets from OpenNeuro and add them to the test battery.

Thanks a lot for your input, very much appreciated!

oesteban · 2018-10-19T04:17:13Z

Fixes #1000
Fixes #1181

oesteban · 2018-10-19T04:36:31Z

Ref #900

zhifangy · 2018-10-19T05:58:10Z

Here's the fMRIprep error log and the verbose log of N4 correction command.

I think it's due to the very small discrepancy between the qform matrix of pre_mask and ref_image.

The following link is the raw data to replicate this error. https://www.dropbox.com/s/akfyadrih5lgcbd/sub-30.zip?dl=0

oesteban · 2018-10-24T16:58:36Z

I think it's due to the very small discrepancy between the qform matrix of pre_mask and ref_image

Correct, there is a minimal difference in the qform.

oesteban · 2018-10-24T18:11:24Z

Alright, this ready for review (@ZhifangYe, @effigies). I've just fixed the issue of mismatched headers before N4.

Since we don't make any promises about very limited FoV (https://github.com/poldracklab/fmriprep/labels/partial%20FOV), I propose to merge this in and come back later to the FoV issue.

Regarding sub-26_task-rest, please check out the regression tests under the test_pytest build of Circle.

effigies

I'm not sure about the new interface.

effigies · 2018-10-24T18:17:52Z

fmriprep/interfaces/images.py

+        imghdr['dim_info'] = refhdr['dim_info']  # dim_info is lost sometimes
+
+        # Set qform
+        qform, qcode = refhdr.get_qform(coded=True)


Are all images guaranteed to have non-zero codes? If qcode == 0, qform == None.

https://github.com/nipy/nibabel/blob/d5494f3350967b65bd6de1abefd95ab2899044e0/nibabel/nifti1.py#L916-L917

effigies · 2018-10-24T18:18:29Z

fmriprep/interfaces/images.py

+        imghdr.set_qform(qform, int(qcode))
+
+        # Set sform
+        sform, scode = refhdr.get_sform(coded=True)


Also the case for sforms:

https://github.com/nipy/nibabel/blob/d5494f3350967b65bd6de1abefd95ab2899044e0/nibabel/nifti1.py#L1059-L1060

effigies · 2018-10-24T18:24:00Z

fmriprep/interfaces/images.py

+                                   newpath=runtime.cwd)
+
+        imgnii.__class__(imgnii.get_data(), imgnii.affine, imghdr).to_filename(
+            out_file)


I don't think this will work how you want. You may want to add tests.

If imgnii.affine does not match the result of imgnii.header.get_best_affine(), it will be pushed into the header, probably as the sform. If you've just put refhdr.get_sform() into the imghdr sform matrix, then these are unlikely to match, and you'll be pushing the original affine right back in.

A better approach would be:

imgnii.__class__(imgnii.get_data(), imghdr.get_best_affine(), imghdr).to_filename( out_file)

oesteban · 2018-10-24T22:57:47Z

This is sub-26_task-rest: https://2591-53608443-gh.circle-artifacts.com/0/tmp/data/reports/ds001240/sub-26_task-rest_bold_masks.svg which looks okay.

We are seeing ds005 fail due to a problem checking filenames that is also happening in master. I'm guessing that would be a trivial change to do in master.

Thumbs up?

oesteban · 2018-10-25T00:36:09Z

ds005 problems addressed here - #1344

If there are no further comments or voices against, I'll be merging later today and close #1050 as this PR supersedes that one.

Thanks to @ZhifangYe, this PR has been only possible with all the new data we have now for regression tests.

effigies

This looks good. I took another read through, and only saw one thing, which should have no impact one way or another, but would reduce the number of branches.

effigies · 2018-10-24T23:19:25Z

fmriprep/workflows/bold/util.py

+    if pre_mask:
+        workflow.connect([
+            (inputnode, enhance_and_skullstrip_bold_wf, [('bold_mask', 'inputnode.pre_mask')]),
+        ])


I think this should be okay to set unconditionally.

zhifangy

The test results look good to me. I'm really happy to see the issue is resolved. 👍

I read through the code and only find the doc in init_enhance_and_skullstrip_bold_wf line 185 doesn't match. I think it should be '7. Calculate a final mask as the intersection of 4) and 6).'.

[ENH] Template-based masking of EPI boldrefs

ec5e651

oesteban mentioned this pull request Oct 15, 2018

[ENH] Add "fMRIPrep" template, with new boldref template nipreps/niworkflows#255

Merged

oesteban added 3 commits October 15, 2018 11:52

fix flake8 error

3e344db

use nipreps/niworkflows@0946bad

2679533

update pinning to release

af0cd52

oesteban added 3 commits October 15, 2018 23:09

new registration parameters

c088fb6

pre-cache template getter

966d412

fix ds005 problem

be17090

oesteban mentioned this pull request Oct 16, 2018

[ENH/WIP] Robust EPI Mask for heavily nonuniform image #1050

Closed

oesteban added 2 commits October 16, 2018 12:46

1) shorten registration; 2) accept a pre_mask

76d0f70

[skip ci] update documentation

713132f

oesteban changed the title ~~[ENH,WIP] Template-based masking of EPI boldrefs~~ [ENH] Template-based masking of EPI boldrefs Oct 16, 2018

oesteban added 3 commits October 17, 2018 00:23

Merge remote-tracking branch 'upstream/master' into fix/1000

18fe6ca

Merge branch 'hotfix/reports-3' into fix/1000

4dc0908

Merge branch 'hotfix/reports-3' into fix/1000

30d7658

oesteban added a commit that referenced this pull request Oct 17, 2018

[TST] Add heavily-nonuniform boldrefs for regression tests

6215779

Refs. #1000, #1050, #1321. Also includes a new filter of branches so builds other than tests are skipped if the name of the branch starts with ``tests?/``.

oesteban mentioned this pull request Oct 17, 2018

[TST] Add heavily-nonuniform boldrefs for regression tests #1329

Merged

Merge remote-tracking branch 'upstream/master' into fix/1000

65e6fbc

Merge remote-tracking branch 'upstream/master' into fix/1000

8ba55fa

fix header mismatch problem

698e821

[skip ci] Pin niworkflows==0.4.4

e52394e

effigies reviewed Oct 24, 2018

View reviewed changes

address @effigies' comments on MatchHeader

2a09ae8

fix master

f86b543

effigies approved these changes Oct 25, 2018

View reviewed changes

effigies added this to the 1.2.0 milestone Oct 25, 2018

oesteban added 2 commits October 24, 2018 18:07

Merge branch 'ds005/expected-outputs' into fix/1000

9503750

remove one branch

ca32147

zhifangy approved these changes Oct 25, 2018

View reviewed changes

[skip ci] fix minimal error in docs

7535a08

oesteban merged commit 874c4ce into nipreps:master Oct 25, 2018

oesteban deleted the fix/1000 branch October 25, 2018 03:39

oesteban mentioned this pull request Oct 27, 2018

Reference brainmask failing for highly intensity nonuniform bold data #1000

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[ENH] Template-based masking of EPI boldrefs #1321

[ENH] Template-based masking of EPI boldrefs #1321

oesteban commented Oct 15, 2018

oesteban commented Oct 15, 2018 •

edited

Loading

oesteban commented Oct 15, 2018 •

edited

Loading

oesteban commented Oct 16, 2018

oesteban commented Oct 16, 2018

effigies commented Oct 16, 2018

oesteban commented Oct 16, 2018 •

edited

Loading

zhifangy commented Oct 18, 2018

oesteban commented Oct 18, 2018

oesteban commented Oct 19, 2018

oesteban commented Oct 19, 2018

zhifangy commented Oct 19, 2018

oesteban commented Oct 24, 2018

oesteban commented Oct 24, 2018

effigies left a comment

effigies Oct 24, 2018

effigies Oct 24, 2018

effigies Oct 24, 2018

oesteban commented Oct 24, 2018

oesteban commented Oct 25, 2018

effigies left a comment

effigies Oct 24, 2018

zhifangy left a comment

[ENH] Template-based masking of EPI boldrefs #1321

[ENH] Template-based masking of EPI boldrefs #1321

Conversation

oesteban commented Oct 15, 2018

Changes proposed in this pull request

Documentation that should be reviewed

oesteban commented Oct 15, 2018 • edited Loading

oesteban commented Oct 15, 2018 • edited Loading

oesteban commented Oct 16, 2018

oesteban commented Oct 16, 2018

effigies commented Oct 16, 2018

oesteban commented Oct 16, 2018 • edited Loading

zhifangy commented Oct 18, 2018

oesteban commented Oct 18, 2018

oesteban commented Oct 19, 2018

oesteban commented Oct 19, 2018

zhifangy commented Oct 19, 2018

oesteban commented Oct 24, 2018

oesteban commented Oct 24, 2018

effigies left a comment

Choose a reason for hiding this comment

effigies Oct 24, 2018

Choose a reason for hiding this comment

effigies Oct 24, 2018

Choose a reason for hiding this comment

effigies Oct 24, 2018

Choose a reason for hiding this comment

oesteban commented Oct 24, 2018

oesteban commented Oct 25, 2018

effigies left a comment

Choose a reason for hiding this comment

effigies Oct 24, 2018

Choose a reason for hiding this comment

zhifangy left a comment

Choose a reason for hiding this comment

oesteban commented Oct 15, 2018 •

edited

Loading

oesteban commented Oct 15, 2018 •

edited

Loading

oesteban commented Oct 16, 2018 •

edited

Loading