Hi Frank,
On Tue, 26 Feb 2002, Frank Ezekiel wrote:
> I've been evaluating the SIENA package against BSI, and unlike
> some other sites have been getting significantly worse results with SIENA.
I presume you mean comparing BET against BSE? As Siena includes many steps
after BET, it makes more sense this way. Sure - BSE is pretty good,
although in the quantitative tests we carried out over a very wide range
of image types from a wide range of scanners, BET came out on top of BSE
and AFNI: http://www.fmrib.ox.ac.uk/analysis/research/bet/bet/node18.html
though, yes, of course, it won't win every time!
> 1) The BET function appears to include a LOT of non-brain in our coronal
> T1 studies, in spite of my attempts to tune the parameters. The appears
> to be particularly a problem in the anterior region around and below the
> eyes and optic nerve.
If you could email me directly an example image then I can play with BET -
that's probably the best way forward.
> 2) If I run SIENA with an expansion of our existing brain masks as the
> input, the results are much more closely correlated with BSI. This
> suggests to me that the failure of BET contributes substantially to the
> failure of SIENA.
that could well be right.
> Also, might it be possible to run SIENA with two inputs for each timepoint
> such that our own brain mask is used for the main analysis but the complete
> image is still provided for the skull estimate and corresponding voxel
> drift correction?
if necessary, yes, that should be possible - I'll have a look at the
images first, then let you know.
> I've also been looking over the FAST routine, and notice 2 things as well:
>
> 1) The bias field correction appears more anatomical than I would have
> expected, with changes at Gray/White boundaries. This tendency is
> reduced as I increase the number of iterations for bias field smoothing but
> is still disturbing.
this is very unusual - are you reducing the smoothing of the bias field?
the default smoothing normally prevents this kind of thing....
> 2) The final estimates appear to overestimate CSF by any reasonable
> standard.
>
> Questions:
> Can the bias field correction be disabled so that we can use our own method
> for this, or some other standard such as N3?
it's not an option, although you could effectively disable it by setting a
massive amount of smoothing (which would slow it down of course) although
other methods (including N3) do not work as well in general. again - it's
probably best if we take a look at these images.
> Can the algorithm be tweaked so as to increase the tissue percentage in the
> outcome?
There are many controls, such as number of tissue types, the initial
tissue means, and turning on prior probabilities......but, sorry to repeat
myself, we'd need to look at the images really, to see if there's anything
unusual about them.
Thanks for your feedback, Steve.
|