FIX: SDC-SyN ("fieldmap-less") overhaul #239
Conversation
|
Hello @oesteban, Thank you for updating! Cheers! There are no style issues detected in this Pull Request. 🍻 To test for issues locally, Comment last updated at 2021-10-06 13:31:44 UTC |
Salient points on this commit: - There's no need for intensity inversion if MI is used as cost function - Use of the T1w as reference (fixed) image increases reliability - Giving more flexibility to ``restrict_deformation`` with 0.1 instead of 1 makes optimization faster leveraging a minimal increase of dofs - To double-check things look good, use our internal transform object to generate the unwarped reference, and then generate the field, which should be fairly close to the one generated by ANTs. - Addressed LPS orientation bug in the function node that extracts the nonzero component of distortion. - Better registration parameters, with thanks to @yasseraleman for suggestions (in particular, the one about flexibilizing the gradient restriction above came from him). - Use of laplacians in a multi-channel registration framework Resolves: #186.
oesteban
force-pushed
the
fix/syn-revision
branch
from
ec1ee59
to
a75ee8a
Compare
oesteban
force-pushed
the
fix/syn-revision
branch
2 times, most recently
from
bfa2df9
to
816c894
Compare
oesteban
force-pushed
the
fix/syn-revision
branch
from
816c894
to
553a5c6
Compare
|
Okay, I think the registration parameters of both initialization and SyN steps, and the B-Spline approximation, are pretty good for the one subject I'm testing on from ds005. We probably want to merge this and maybe refine parameters as we test this on other datasets. |
There was a problem hiding this comment.
I think it makes sense. If I understand correctly, we're moving from histogram-based matching to finding a single deformation that optimally registers both intensity and the intensity-derived Laplacian (~edges) images?
Some comments (mostly readability) about the inverse functions.
Co-authored-by: Chris Markiewicz <markiewicz@stanford.edu> Co-authored-by: Mathias Goncalves <mathiasg@stanford.edu>
|
I believe the test failure is that |
| "number_of_iterations": [ [ 10, 0 ], [ 10, 5, 0 ], [ 2 ] ], | ||
| "metric": [ ["Mattes", "Mattes"], ["Mattes", "Mattes"] ], | ||
| "metric_weight": [ [0.5, 0.5], [0.5, 0.5] ], | ||
| "number_of_iterations": [ [ 20, 10 ], [ 2 ] ], |
There was a problem hiding this comment.
Same as sd_syn, but with factor of 0.1 and 0.2 applied, matching the previous factors 0.1, 0.5, 0.2.
|
Okay, I think all the comments have been addressed - I'm going to press the red (green) button now. |
In the context of nipreps/fmriprep#2530, this PR takes a deep revision on the SDC-SyN implementation. In passing, the problems about suboptimal registration (#85) are also addressed.