Dear Siew-Min,
and everyone else concerned with FA<0 or >1.
These values may seem disconcerting because we know that they are
"impossible", i.e. they come from voxels where the diffusion is
negative in some direction (indicating a black hole or some such thing
to which things can diffuse and disappear).
If instead you see it as having some set of data (your raw diffusion
data) affected by measurement error, and from these data you want to
estimate some parameters (let us say the diffusion in some direction).
The true value will be positive (though it can be quite small in the
case of hindered diffusion in white matter), and then our estimate
will have some finite precision that depends on the quality of the
data (SNR, # of directions etc). So, typically we will not calculate
the "true" value, but some value that is drawn from a distribution
around the true value (known as the "sampling distribution").
In some instances the true value is small (i.e. close to zero), and
then the sampling distribution (i.e. values that we may calculate)
will extend across zero. Then there is a chance/risk that we will
calculate/observe a negative diffusion value and subsequently an FA>1.
This might happen e.g. in highly anisotropic areas in the brain.
In other instances there isn't really a meaningful signal (e.g.
outside the brain) and the precision of our estimate will be very
poor, which again means that we may calculate/observe a diffusion < 0.
So, in short. We never get the "true" FA value. We get an estimate,
and that estimate can be more or less wrong (i.e. it can have better
or worse precision). Sometimes these "wrong" values are also
"impossible", but that doesn't really make them any more "wrong". It
just becomes a bit more obvious.
By increasing the quality of our data (higher SNR and/or more
directions) the errors become smaller and smaller, and the chance/risk
of calculating a negative diffusion becomes smaller. But it doesn't go
away. It just becomes so small that it is very unlikely that we
calculate/observe one given that we are only looking at a few tens of
thousands of voxels.
Therefore, I would suggest stop worrying about negative FA<0 or FA>1
and accept them for what they are. Noisy estimates of some unknown
value in the range 0-1.
Good luck Jesper
On 13 Mar 2009, at 15:11, Siew-Min Gan wrote:
> Hi all,
> I noticed my FA map output from DTIFIT has white matter brain
> voxels of 0 and >1. I ran the following procedures to get the FA
> map.
> i.eddycorrect on the original DTI
> ii.bet the eddycorrected data.nii.gz (or the 1st B0) to derive the
> nodif_brain_mask.
> iii.feed the nodif_brain_mask, data.nii.gz,bvecs and bvals into
> DTIFIT.
>
> If the original data is ok, could a problem with the bvecs file result
> some brain voxels with 0 intensity (or intensity>1) in the FA map
> output
> from the DTIFIT procedure?
>
> Initially, I thought it may be a problem with the original DTI data,
> however, it is a standard MGH 30 dir sequence, and the gradient file
> is
> from the diffusion toolkit. I ran the data on another DTI software
> and
> the FA map output doesnt seem to contain 0 intensity voxels, but it
> may be
> a scaling issue??
>
> The solution I tried with FSL is to do BET using the -F function
> (normally
> used for 4D FMRI). The resulting FA map from DTIFIT only then will not
> contain 0 intensity voxels, but still have voxels of intensity>1. I
> notice
> this is not the standard way to process (BET) DTI data.
>
> I am concern about this as I have significant positive results from
> the
> tbss FA voxel analysis. I wonder if these results would be
> independent of
> the fact that some voxels in the FA maps are >1. Also, I could not
> analyse
> the tbss using the FA maps from the standard processing pipeline due
> to
> some brain voxels having 0 intensity.
>
> May I ask if using BET with -F is ok? Has anyone has come across this
> problem before, and may explain why there may be voxel with 0
> intensity or
> intensity>1 in the FA maps output?
>
> Many Thanks for some kind advice.
>
> Siewmin
>
|