3dSkullStrip vs 3dAutomask for stripping EPI data

AFNI Message Board
1

Hey,

Is 3dAutomask (generally) more generous in it’s estimation of the brain when skullstripping EPI images than 3dSkullStrip?

I ask because in this screen shot https://www.dropbox.com/s/rk48mboy3eb1v9f/screenie.png?dl=0 (run on the same data from the same subject), on the left can be seen an invocation of align_epi_anat.py with 3dSkullStrip as the default and on the right is an invocation of aea with 3dAutomask as the skull stripper. Both invocations produce good alignments but the extent of the brain from the invocation involving 3dAutomask is clearly bigger.

Is there a reason to prefer 3dAutomask over 3dSkullStrip fro skullstripping EPI data?

LEFT AFNI:


align_epi_anat.py -anat2epi			\
		  -anat ${anatFile}		\
		  -epi ${epiFile}		\
		  -epi_base 0			\
		  -volreg off			\
		  -tshift off			\
		  -cost lpc			\
		  -multi_cost lpa lpc+ZZ mi

RIGHT AFNI:


align_epi_anat.py -anat2epi -anat bc002b.anat_unif+orig \
       -suffix _al_keep                                 \
       -epi vr_base+orig -epi_base 0                    \
       -epi_strip 3dAutomask                            \
       -cost lpc                                        \
       -volreg off -tshift off
2

Hi Colm,

I do not think we have really noticed much less studied
such an aspect. The original paper that align_epi_anat.py
represents used 3dSkullStrip for EPI data, though I do not
know that it was for any concrete reason. afni_proc.py has
been using 3dAutomask as the default for 4 years now,
mostly since it is faster and seems more robust (3dSkullStrip
was not written for EPI data, for that matter).

Precise stripping of the EPI data does not seem quite so
important, since registration with the anat is affine. The
place where accurate skull stripping is more important is
when applied to the anatomical dataset for non-linear
registration.

Sorry to still not answer your question…

  • rick
3

I can’t see the image, but I will echo Rick’s remarks. Personally, I prefer the 3dAutomask approach because it seems sufficient, and it is much faster. We haven’t tested the 3dAutomask option as much in align_epi_anat.py, but I expect it will work similarly. One of these days, we will change defaults for many AFNI programs, and this will probably be included. For partial coverage datasets, the skullstripping method doesn’t work; in that case, either using 3dAutomask or None (no stripping) works well.