Hi Darren - lots of issues raised here!
First a note about the use of priors with FAST: note that if you turn on
priors, this is only used for the initialisation (ie setting up the
initial class means and to aid the initial segmentation) and is currently
NOT directly affecting the final iterated posteriors. It works by
registering the input image (assumed to be a BETted brain) to the standard
space brain image and using that transform to put the standard space
G/W/CSF priors images into the input image's space. It is _possible_ that
for all this to run you have to be working in the directory of the input
image rather than giving a full pathname.
> I noticed that FAST will not use all apriori images to do a 4 class
> segmentation, it complains that apriori can only be used with 3 classes.
> However, there is an apriori skull image in
> /usr/local/fsl/etc/standard/.
yes - the code would need to be changed to vary what prior images got
used.
>
> It is recommended that FAST is run after BET. However, I am using a
> "whole" T1 volume in FAST, before BET, to evaluate the possibility of
> segmenting the skull. Obviously a T1 is not optimal for this, but it's
> all I've got on these subjects. To facilitate the process, an apriori
> image could be good. One option is the BET skull image as an apriori,
> which requires no registration (ie, an identity registration matrix).
> Perhaps some avwmath functions need to be applied to this BET skull
> volume to transform it to a data type/range consistent with probability
> map volumes. Another option is the standard skull, which does require
> registration.
I think in order to find the skull you probably don't want to be directly
using FAST for the skull - in T1-weighted it would be hard to distinguish
skull from CSF etc. Maybe what you could do would be to run BET, save the
brain mask, dilate it a few times with avwmaths, apply that to the
original, and then run standard 3-class FAST with pve output turned on.
then take the CSF&skull pve output and * it by the standard space skull
prior - maybe that would work? you'd need to use FLIRT to get that into
input image space.
> Assuming this is a worthwhile exercise, it could have applications in
> sienna. As I understand it, sienna uses the skull images for
> registration from time1 to time2 to assist determination of changes in
> cortical volume. BET does a fairly good job of skull extraction, but I
> wonder if some combination of BET with a segmentation of the skull from
> FAST might prove to be useful.
I think in Siena the external skull surface is good enough for the simple
role of finding relative scaling - it seems to work ok....but maybe your
suggestions would be even better...
ttfn :)
Stephen M. Smith
Head of Image Analysis, FMRIB
Oxford University Centre for Functional MRI of the Brain
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
|