Hi,
Our current recommendation for this problem is to truncate the
top 5-8% of the images intensity values with fslmaths before
classification, see the thread "Fast - strange misclassification" for
more details.
Many Regards
Matthew
> Hello,
>
> I was wondering if there has been any update on this extreme dark/
> bright
> classification issue. I am trying segment WMH on FLAIR images. WMH are
> located around the CSF. I do a skull strip on T1 and mask the FLAIR
> image
> using this mask and feed the stripped image to FAST. The number of
> classes in
> FAST is set to 3. FAST tends to merge extreme dark (CSF) and extreme
> bright
> (WMH) and provide one class.
>
> I am using FSL version 4.1.2 and FAST version 4.1. I noticed there
> is already
> an update on FSL. If this issue has been addressed, I will update to
> the new
> version. If there is any work around, please let me know as well.
>
> Thank you very much..
>
> Deniz
>
>
> On Thu, 2 Apr 2009 10:15:58 +0100, Matthew Webster
> <[log in to unmask]> wrote:
>
>> Hi,
>> The thresh.txt needs 1 value for each class ( 4 in your case )
>> which are the _log_ values of the means ( we will be changing the
>> inputs back to normal intensities in the next release ). We're
>> currently looking into the extreme dark/bright classification
>> issue...
>>
>> Many Regards
>>
>> Matthew
>>
>>> Hi, I wanted to identify white matter hyperintensities in a
>>> carefully cleaned up flair image
>>> (Betted and Grey matter and ventricles pretty well removed).
>>> I was hoping that the -s (or --manualseg) option in fast would allow
>>> me to choose a starting value that characterized my wmh,
>>> but when I try that:
>>>
>>>> fast -s thresh.txt combo
>>>
>>> (thresh.txt contains a single number, 722 which is 3 SD above the
>>> mean)
>>>
>>> It seems to be doing something else (it lumps together very dark and
>>> very bright patches....even when I have 4 classes).
>>> Can anyone explain more about manualseg and how to use it? I've
>>> looked at the online materials and the forum, but I'm afraid I don't
>>> see much that is helpful.
>>>
>>> Thanks for any insights.
>>>
>>> -Dianne
>>>
>>> --
>>> Dianne Patterson, Ph.D.
>>> [log in to unmask]
>>> University of Arizona
>>> SLHS 328
>>> 621-5105
>>
>>
>
