I do not see major differences in nonlinear vs. linear registrations (to TT_N27), via visual inspection, for a large number of subjects in my sample. Is there an overwhelming preference for using nonlinear registration for group analysis?
Thank you.
Hi-
I’d be curious to know more details about your data alignments-- what type of dset are you aligning to standard space? With what cost function? How are you evaluating it? I am just guessing that you are using 3dAllineate for linear affine and 3dQwarp for nonlinear; what commands are you using for each?
I would say that, yes, the overwhelming consensus would be that for aligning the brains of different subjects (such as a subject to standard space), you should be using nonlinear alignment.
–pt
The datasets are from task-based scans. I’m using the afni_proc.py script. In that script, nonlinear registration is indeed performed with 3dQwarp (when inlcuding the -tlrc_NL_warp flag).
Example for nonlinear registration:
afni_proc.py -subj_id ${SID} -script $TCSH_AFNI -out_dir $PWD_TASK \
-dsets $FN_NIFTI \
-blocks tshift align tlrc volreg blur mask scale regress \
-copy_anat SUMA/${SID}_SurfVol.nii \
-anat_has_skull yes \
-tcat_remove_first_trs 4 \
-volreg_align_e2a -volreg_tlrc_warp \
-align_opts_aea -big_move \
-tshift_opts_ts -tpattern alt+z2 \
-tlrc_NL_warp \
-blur_size 8 -volreg_align_to first \
....
Linear is the same except without the “-tlrc_NL_warp” flag and with additionally "-tlrc_opts_at -init_xform AUTO_CENTER "
Thank you.
Hi-
I think that 3dvolreg will be used for motion correction (aligning EPIs across time), then align_epi_anat.py (-> which builds a 3dAllineate command) for aligning the EPI to the subject’s anatomical and finally either 3dAllineate or 3dQwarp for aligning the subject’s anat to standard space.
Bob has designed the @snapshot_volreg function to help just the quality of subject-EPI → subject-anatomical, so that might be useful to look at. You could turn the anatomical that has been aligned to standard space into an edge-y view, using 3dedge3, and overlay the result on the standard space brain. That would highlight in more detail how important features (sulci/gyri) etc. look across the brain in each of of alignment.
–pt
Thank you for correcting my misunderstanding of 3dvolreg’s function. I misread the afni_proc.py documentation.
Thank you as well for pointing me to @snapshot_volreg. This is an amazing tool! I tried this out on one processed image that was result of linear affine and nonlinear transformations on the same scan.
I ran:
@snapshot_volreg anat_final.SUBJECT_ID_10000000+tlrc \
final_epi_vr_base+tlrc
The resulting jpegs are attached here. I wanted to ask how to interpret the images. The large lateral ventricles and sulci on superior parts of the brain both look fine (although the nonlinear is slightly better in my opinion).
But is there something fundamentally wrong with the registration? In the top row, I see lines (circles) outlines from the final_epi_vr_base+tlrc file. Is this typical?
This is from the documentation for @snapshot_volreg states:
"The edges from a typical EPI dataset are usually broken up and
do not completely outline sulci, ventricles, etc. In judging
the quality of alignment, I usually start by looking at the
outlines of the large lateral ventricles -- if those are very
wrong, the alignment is not good. After that, I look at the
sulci in the superior part of the brain -- if the EPI edges
there seem to be mostly aligned with the sulci, then I am
usually happy. The base of the brain, where lots of EPI
dropout happens, often does not not show good edge alignment
even when the rest of the brain alignment looks good."
It does not provide references to circles that can be seen in both the linear affine and nonlinear warps. At the same time, the circles are towards the inferior part of the brain in both cases, with documentation stating that EPI dropouts occurring at base of the brain.
Thank you again!
