Hi,
The main difference is the partial volume processing.
This now (in FAST4) uses an MRF on the "mixel type" to regularise
the type of mixtures present in voxels. A "mixel" represents what
tissues occur in a mixture (e.g. WM/GM is one mixel, pure WM is
another, GM/CSF is another, etc.). We disallow the triple mixture
WM/GM/CSF by default, and put an MRF on the remaining mixels.
This means that if a voxel might be either a WM/GM mixel or a pure
WM mixel, but is surrounded mainly by pure WM mixels, and the
intensity data is not strong either way, then it will end up being
classified as a pure WM mixel. This has the effect, as you've
noticed, of making the transitions sharper and forcing more voxels
towards "pure" tissue, instead of letting them be 95% one tissue
and 5% another just due to the noise. This model should therefore
give a better segmentation and a better volume estimation - which is
why we replaced the old one, based on our testing.
I hope this is what you were wanting to know.
All the best,
Mark
On 17 Jan 2011, at 15:35, Keith Hulsey wrote:
> A few years back I ran some analysis using SIENAX in FSL version 4.0. The version of FAST which it used was 3.53. I am trying to modify my previous work, but I have found that the segmentation process has changed, as SIENAX now uses FAST version 4.1. Is there a document which explains the changes between fast3 and fast4 other than the conversion table for the program options? Looking at my data, it seems like results using fast4 have a quicker transition from white matter to gray matter at the cerebral cortex, giving fewer partial volume voxels. Would you expect that change given the differences between versions?
>
|