I think your approach of correcting FLAIR + T1 volume is fundamentally
flawed.
Generally, fast will try to push all voxel intensities into 3 classes. So
contrast reduction for hyperintsity regions is an expected consequence.
Correcting FLAIR + T2 is what needs to be done since the intensity clusters
are overlapping.
pk
----- Original Message -----
From: "Erin Gibson" <[log in to unmask]>
To: <[log in to unmask]>
Sent: Thursday, March 08, 2007 2:20 PM
Subject: [FSL] mfast and FLAIR
Hello all,
I am trying to correct 3T FLAIR data affected by significant intensity
inhomogeneities. I have spent quite some time experimenting with mfast and
fast, and the various options (number of classes, priors, max iterations
etc.).
In general, what I have found is that while fast produces a decent
correction overall, certain areas in certain subjects can be left
under-corrected, causing problems for some of our other algorithms. Using
mfast with a coregistered T2 as secondary input produces excellent results
with little or no bias remaining in the image. Using mfast with a
coregistered T1 produces a good result overall, but in subjects with white
matter disease, white matter hyperintensities appear to be over-corrected,
such that the intensity of the hyperintensities is noticeably reduced
relative to normal tissue, as compared to the fast or mfast-with-T2 results,
and this too causes trouble for our other algorithms.
My problem is that while I will always have a coregistered T1 available, I
will not always have a coregistered T2. Therefore, I would like to find a
way to correct the FLAIR using only the T1.
I wondered if the sub-optimal results I obtained using mfast and the T1 was
related to the fact that FLAIR hyperintensities often appear hypointense on
T1. I tried multiplying the FLAIR image against the T1 image (FLxT1), so
that diseased hypointense areas on the T1 would appear hyperintense, as they
do on the FLAIR. I then used this FLxT1 image as secondary input to mfast.
This actually worked surprisingly well, giving me a result very similar to
the mfast-with-T2 result.
My question is whether using a FLxT1 image in this situation is reasonable.
I am satisfied with the results, but have I violated any underlying
assumption of the mfast algorithm that would somehow invalidate the
correction results?
Any suggestions or comments would be greatly appreciated.
Thanks,
Erin.
|