Dear Steve,
I've been trying SIENA on data for which I have manually segmented
brain masks (for one, or possibly both time-points). These masks trace
around the tissue that is either GM or WM; CSF is not included.
Obviously I don't have BET-style skull estimates for input to pairreg.
Following some work Bea Sneller did:
http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0703&L=FSL&P=R20435
I've been trying the following simple approach:
* let SIENA call BET as usual (getting skull estimates etc.)
* overwrite BET's _brain_mask images with my masks
- if only one mask is available, dilate this mask to provide an
over-generous mask for the second image
* it appears to me from siena_diff.cc that the new _brain_masks
should be handled correctly (transformed to the half-way space, etc.),
but I might be missing something?
The render images resulting from this have really nice looking edge
locations, without the non-brain tissue that occasionally gets
included by BET. So far so good...
However, the actual values of final PBVC don't seem as good as the BET
results (based on some simulated atrophy images with approximately
known volume loss) even though BET has not worked that perfectly on
some of the images.
So, am I missing something that is wrong with my approach? One thing
I'm wondering at the moment, is that the tight (no-CSF) brain masks
could give a sharp discontinuity at the edges (especially if the
manual segmentation actually went inside the brain slightly, by
mistake), which perhaps doesn't give the best results with
siena_diff's correlation of the intensity profiles. If this is the
case, I guess it would be better to dilate both brain-masks by some
amount. I'm intending to try this, but I thought I'd check first
whether there is anything else I might be missing...
Thanks in advance for your help,
Ged.
P.S. I'm under the impression that the SIENA pipeline (including BET,
FLIRT, etc.) is unchanged from FSL 3.3 to 4, is that correct?
|