Dear Blake,

If you want the best registration, and you are working with structural images then our recommendations would be:
 - FLIRT with 6 DOF for images from the same subject
 - FLIRT with 12 DOF and then FNIRT for images from different subjects

In the first case you could use either brain extracted images or whole-head images.  If you can't get good brain extraction, then the latter is probably the best option.

In the second case, FNIRT uses the non-brain-extracted image to do its registration.  We recommend using brain extracted images for the initial 12 DOF FLIRT, but that isn't that important and you can also use the non-brain-extracted if that works better (or is easier).  FNIRT uses a standard-space brain mask to exclude non-brain material when it does its registration, so it is not unduly influenced by non-brain structures.

I hope this helps.
All the best,
	Mark



On 8 Apr 2014, at 13:23, "Dewey, Blake (NIH/NINDS) [F]" <[log in to unmask]> wrote:

> Thanks Mark for your detailed answer. I appreciate the commitment that you
> have for this. I have a followup question if you don't mind.
> 
> We normally perform a course registration with the skull on to align the
> images (-30 to 30 finesearch of 5) and interpolate the image to match
> size, FOV, etc. We then use a common brain mask as weight in the
> registration and do a finer registration (-15 to 15 finesearch of 1). In
> this way, we are hoping to achieve a very close match between images of
> the same patient and the closest alignment possible between patients. Is
> this not the proper way to achieve this? If not, what is the recommended
> way to achieve the best registration between two images, focusing on brain
> structures (ie. not biased by skull shape). Would it be better to create a
> weighted image that has non-zero values for the whole image, but has very
> small weights outside of the brain (for the mutual information cost
> function, which we have had the best results with).
> 
> Thanks again,
> Blake
> 
> On 4/8/14 2:15 AM, "Mark Jenkinson" <[log in to unmask]> wrote:
> 
>> Hi,
>> 
>> Yes, masking within an image (including zeros within the image) can
>> impose artificial boundaries that would bias registration.  Taking an ROI
>> by reducing the size of the image FOV (throwing away voxels, rather than
>> replace them with zeros) is fine and does not cause problems (as the
>> registration methods only use the overlapping FOV portions of the images
>> for all calculations).
>> 
>> What the problem is most likely to be caused by is the lack of brain to
>> non-brain boundary in the image.  We strongly discourage the use of the
>> brain mask as a weighting image, since it removes the information about
>> where the brain edge is (since the transition from brain to non-brain
>> never occurs, as all the non-brain voxels are outside of the mask and
>> hence ignored).  So I would just avoid masking like this.  If you need to
>> mask for some other reason (e.g. to avoid pathology) then that is fine,
>> although small masks can cause problems for the mutual information
>> measures, as they use 2D histograms to calculate their values, and these
>> histograms are more influenced by small numbers of voxels (giving poor
>> estimates of the values per bin) than for any of the other measures (that
>> use 1D histograms, or no histograms).
>> 
>> All the best,
>> 	Mark
>> 
>> 
>> 
>> On 7 Apr 2014, at 18:06, "Dewey, Blake (NIH/NINDS) [F]"
>> <[log in to unmask]> wrote:
>> 
>>> Wouldn't that impose an artificial boundary on the image for the
>>> registration?
>>> 
>>> Blake
>>> 
>>> From: David Grayson <[log in to unmask]<mailto:[log in to unmask]>>
>>> Reply-To: FSL - FMRIB's Software Library
>>> <[log in to unmask]<mailto:[log in to unmask]>>
>>> Date: Monday, April 7, 2014 1:02 PM
>>> To: "[log in to unmask]<mailto:[log in to unmask]>"
>>> <[log in to unmask]<mailto:[log in to unmask]>>
>>> Subject: Re: [FSL] FLIRT using Weighting Image and Mutual Information
>>> 
>>> Hi Blake,
>>> 
>>> For MI, you can try explicitly masking your images with fslmaths prior
>>> to registration.
>>> 
>>> Best,
>>> 
>>> David
>>> 
>>> 
>>> On Mon, Apr 7, 2014 at 8:05 AM, Blake Dewey
>>> <[log in to unmask]<mailto:[log in to unmask]>> wrote:
>>> I have been trying to use weighted registration for my structural image
>>> (using the brain mask as a weighting image). I have had success with
>>> each of the cost functions except for the mutual information and
>>> normalized mutual information functions. For these two cost function the
>>> registration fails spectacularly, often rotating the image out of the
>>> FOV. Is there a reason for this? What can I use if I have two very
>>> different contrasts to register in this way?
>>> 
>>> Blake