Hi AFNI expert,
I finally found the solution and I will probably be able to be more clear.
Two things were impacting the results of this analysis and fixing it has tremendously improved the quality of the coregistration:
1. Deobliquing the data prior to @afni_proc.py can be dangerous if like in my dataset the EPI and the anatomical images have different FOV orientations.
3drefit \
-deoblique \
${dset_bold_deob}
What I have done is to first read in the oblique transformation matrix from the anatomical image and make the cardinal dataset oblique to match the EPI
(It is the answer, explained with a good vocabulary of the question of this topic (2 images in the same space), we can probably change the title now =)):
3dWarp -card2oblique anat_image -prefix BOLD_anat_obliq_orig ${dset_bold_deob}
and then:
3drefit \
-deoblique \
${dset_bold_deob}
However, do not forget that EPI time series data should be time shifted with 3dTshift before rotating the volumes to a cardinal direction (https://afni.nimh.nih.gov/pub/dist/doc/program_help/3dWarp.html). In consequence, I removed Tshifte of @afni_proc.py and apply it directly to the raw EPI images prior to 3dWarp.
2. The second thing is that the quality of the images was worst than the previous group of monkeys. Changing the "-cost" in @afni_proc.py from "lpc+ZZ" to "mni" has greatly improved the coregistration.
I hope that this would help someone one-day =),
Dear AFNI expert sorry for the bad explanation at the beginning, I was clearly missing some "AFNI vocabulary" to explain my problem.
I hope that I didn't do anything wrong on the EPI images with these changes.
Thanks a lot to the AFNI community!
Clément