Dear Mishkin,
>
> Seems counter-intuitive. wouldn't the CSF Gaussian cross over with the
> GM Gaussian before it crosses over with the WM one? I'm using a T1w
> image so from hypo to hyper it goes CSF,GM,WM.
It makes sense because the CSF Gaussian has a very low amplitude in
general and so the cross over of the GM and WM occurs at an amplitude
well above the CSF distribution at that point, and so you typically get the
largest amplitude going from CSF (low intensities) to GM then to WM
and then back to CSF (when the WM has gone very low and the GM
is even lower, but the CSF - due to the much bigger variance - is still
larger).
>> This is obviously
>> not desirable, but is a consequence of the simple mathematical
>> model which is at the heart of FAST. One solution we've got
>> to work in the past is to increase the number of classes to 4,
>> which can result in the CSF getting modelled in two classes,
>> as often there are some very dark portions of the CSF and
>> some less-dark, and these are more effectively modelled
>> by two Gaussians, limiting the extent of their reach into the
>> high intensities.
>
> I tried 4 and 5 class but I still got the same misclassification.
That's a pity.
>>
>> So try this and let us know how you get on.
>> As a last resort you can always take the white matter mean
>> intensity and threshold the "CSF" component to exclude the
>> brighter voxels, but try the 4-class solution first.
>
> Could you elaborate as to where I would do the thresholding?
> On my original image I set all CSF voxels that had a value above the
> mean WM to 0 before I sent it into FAST, but this produced similar
> results. Probably because there are almost no CSF voxels that have an
> intensity higher than the mean WM value.
> Did you mean something else?
Sorry, I wasn't clear. I am assuming that the "holes" in your
WM are classified as CSF - correct? If that is the case then
you could reclassify the CSF by splitting it with a threshold
(at the WM mean) into low intensity (assumed true) CSF
and high intensity (assumed false) CSF, where the latter
can be added back to the WM segmentation. But this is
done *after* FAST has run, not before.
>
> Also, I'm trying to debug the mixeltype classes.
> Assuming we do a 3 class classification is this corect
> 0 PURE CSF/BACKGROUND
> 1 PURE GM
> 2 PURE WM
> 3 MIXTURE OF CSF AND ?
> 4 MIXTURE OF CSF AND WM?
> 5 MIXTURE OF GM and WM?
The codes are:
0 = CSF
1 = GM
2 = WM
3 = CSF & GM
4 = CSF & WM
5 = GM & WM
Hope some of this helps.
All the best,
Mark
|