Thank you so much Mark! Very very helpful. I will absolutely check the registrations on all my subjects to make sure. I just ran one registration and viewed the transformed T1 brain overlapped with the standard brain in fslview...I did in fact see that the few small pieces of dura that were left on the BET'ed brain were a bit outside of the standard brain :)

And yes, I had my ROI'ed image (before BET'ing) titled identical to the BET'ed imaged except for "_brain" in the same folder when I was running FEAT before...so FSL is performing the registration correctly. Thanks so much for explaining!

Have a great day!
Dar

----------------

Dear Dar,

What you are doing with your data and the FEAT GUI sounds fine.
I'm not too concerned about a bit of dura, but please check the
registrations and see if they are affected or not.  That is, is the
dura overlapping with brain afterwards, or does the whole registration
look good.  This is the only way of being sure whether things are OK 
or not.

If it is only a bit of dura and not major and consistent bits of the brain, 
then I wouldn't worry about bias field - I just put that in as it could have 
been relevant, given that I didn't know much about the nature of the 
problems.  Also, bias field is an inhomogeneity in the B1 field (the RF 
field) and not the B0 field (which causes geometric distortion and signal 
loss instead).  These are different problems and solved in different ways.  
Normally the B1 field (bias field) issues are not so bad and nothing needs 
to be done specifically about them.  But B0 field issues are often quite 
significant and using fieldmaps is strongly recommended, so I'm glad you 
have them.

Finally, for using FNIRT in FEAT you can continue running with the GUI
but you must make sure that both the brain extracted and non-extracted
versions of the image are available.  This is stated in the web docs and
bubble help:

"FNIRT requires whole head (non-brain-extracted) input and reference images for optimal accuracy; if you turn on nonlinear registration, FEAT will look for the original non-brain-extracted structural and standard space images in the same directory as the brain-extracted images that you input into the GUI, and with the same filename except for the "_brain" at the end. It will complain if it can't find these, and if this is not corrected, nonlinear registration will run using the brain-extracted images, which is suboptimal."

So if you are not seeing a warning complaint then it is probably fine, but
just make sure that you have the two versions in the same place, with the
brain extracted one named with the same name, but ending in _brain.

All the best,
	Mark



On 20 Jul 2011, at 09:40, Dar Meshi wrote:

> Hi Mark, for some reason my reply did not include your previous answer, and it had some unusual characters in it...so I edited my text, and copied and pasted your previous answer below...
> 
> -----------------
> Hi Mark,
> 
> Thanks so much for the great answer! As far as the answers to the first two questions I asked, I completely understand. I am just using BET to extract the brain, no need for betsurf, hence no need for T2-weighted scans.
> 
> As far as the proper non-brain-to-brain ratio after BET, here are some thoughts and more questions...and I apologize for the long email:
> 
> 1. I just want to explicitly state that I'm performing a functional MRI experiment, and I took a T1 structural as well as a B0 fieldmap. I'm using the FEAT GUI to analyze functional MRI data, so I'm doing my registration there. If I understood things at the FSL course correctly, I'm registering my functional data to my BET'ed T1 with an affine transformation (6 DOF, FLIRT within FEAT), and I'm registering my BET'ed T1 to the MNI152 brain with a non-liner transformation (12 DOF, FNIRT within FEAT). Oh, and I'll also be use the B0 fieldmap in FUGUE within FEAT to correct for bias. Is this correct?
> 
> 2. Just clarify the problem; it's my concern over the non-brain-to-brain ratio after BET. It's not localized to one specific area (like the frontal cortex), but rather wherever there is a dura-to-brain transition. For example, depending on the parameters I've entered, after BET'ing I am never left with a 90% dura-free brain without having parts of cortex removed as well. So I need to settle for a 65-70% dura-free brain in order for the cortex to appear 98% intact. I'm just trying to find the right non-brain-to-brain ratio. Am I to understand that you think this ratio is relatively unimportant when doing FLIRT or FNIRT (in FEAT as I am doing it)? I mean, a little extra dura here, a little more brain over there...doesn't make a big difference because of the way the brain is transformed?
> 
> 3. Also, do the above-mentioned problems with BET (the non-brain-to-brain ratio when just looking at a brain post-BET, not checking registration, etc) imply that I have a field bias? My T1 structural scan looks uniformly bright, but how do I know? To note, I'm consistently having to provide a slightly negative value for the "threshold gradient" (around -0.1) in order to get the neck area (what is left after ROI'ing) removed. Does this mean my field has a bias? If so, you mentioned to use FAST, but I'm not sure how to use this, and also, if I'm not mistaken, I thought this was something that was done in FUGUE (or FUGUE within FEAT) using a B0 field map after I BET. Am I correct? Sorry, just got a little confused because I am unfamiliar with FAST :)
> 
> 4. Final question...you mentioned that FNIRT uses the whole head structural scan for the last few passes...does FEAT know to do this? I don't see where in the FEAT GUI I can enter the whole head T1 scan. If this is a problem, should I be doing all my registration outside of FEAT?
> 
> Thanks so much! I really appreciate your time and effort.
> 
> Have a great day,
> Dar
> 
> 
> 
> -----------------
> Hi,
> 
> The T2-weighted scan and "BET2" in this context is only concerned with 
> when you want to fit the "inner skull, outer skull and outer scalp" as it says
> in the slide.  If all you want to do is extract the brain (from non-brain)
> then a single T1-weighted image is fine and is the most common case.
> 
> You can't do the inner skull, outer skull and outer scalp estimation with
> the BET GUI - you need to run betsurf from the command line.
> 
> As for large/small errors in BET.  It is difficult to give any general rule of
> thumb because it depends on where the problem is and whether it is
> scatter or consistently in one area.  Are the problems you are having
> related to bias field by any chance?  You might benefit from running
> FAST to do bias correction (which does not need the initial BET result
> to be very accurate, as you will just use the bias field output which is
> very smooth and hence not unduly affected by small extra/missing regions).
> As for other tools - FLIRT is relatively insensitive because it only estimates
> 6 or 12 parameters from the whole brain, so as long as the erroneous
> parts make up a sufficient small region wrt the entire brain then it is unlikely
> to affect things a lot unless there is a very consistent change (like eroding
> the whole frontal part of the brain away by a few voxels - something like
> this would definitely cause problems).  For FNIRT it doesn't matter as it
> uses the non-brain extracted images for the final passes and hence should
> recover from any small errors introduced before.  The main one where there
> is a big effect is FAST when it is run for the purpose of extracting quantitative
> tissue volumes.  In this case either missing parts of the brain or extraneous
> non-brain material *will* affect the volume estimates, normally enough to make
> the measurements not very useful (since it is typical to look for small changes
> in volume, and so small errors matter).
> 
> In general, just try your best result and see what happens.  If the registration
> looks fine then you don't have to worry.  If you are doing something else, then
> check those results.  Unless it is a difficult thing to check visually, then it normally
> suffices to *LOOK AT YOUR RESULTS*!  :)
> 
> All the best,
> 	Mark
> 
> 
> 
> On 19 Jul 2011, at 12:55, Dar Meshi wrote:
> 
>> Hi Everyone,
>> 
>> I was at the latest FSL course in Montreal last month (thanks so much for the great course!) and I was just reviewing the slides on BET2. There is a slide on page 57 of the booklet which mentions BET2 "subsequently fits inner skull, outer skull and outer scalp using T1- and T2- weighted images." This leads me to ask a couple questions:
>> 
>> 1. I actually didn't collect a T2-weighted structural scan, but I think FSL is performing BET as it should (see below). Should I be concerned that I don't have a T2-weighted scan? How does BET2 work without the T2-weighted scan?
>> 
>> 2. In the future, if I was to collect a T2-weighted scan, how can I tell FSL which image file to use? Currently, I cannot find a way to link another input file in the BET GUI.
>> 
>> Also, I've been having problems with BET either leaving too many pieces of non-brain, or taking big pieces of brain out....very hard to find the sweet spot with BET, and this changes with each brain I analyze. On page 56 of the booklet you show an example where "leaving small pieces of non-brain is unimportant for most apps". Could you please describe why this is? If I'm registering functional data to this T1 brain and also registering this T1 brain to the MNI standard, wouldn't any extra pieces, or brain deformities cause problems? Oh, and also could you possibly give an example of a tolerable amount of non-brain pieces? In the picture in the booklet you have 2 small pieces, but I'm having problems getting it down to such a small amount without taking big chunks out...so how much is too much? :)
>> 
>> Thanks so much for your help!
>> 
>> Have a great day,
>> Dar
>> 
>