Matthew,
Thanks for the explanation. Since my inputs are multi-channels (T2W,
FLAIR, and PDW), I guess there is no easy way to enforce a consistent
numbering of the segmentation classes based on the mean class intensity
(what's its definition in the case of multi-channel inputs, btw?).
Gordon
On 09/15/2010 10:50 AM, Matthew Webster wrote:
> Hello,
> FAST has never assigned segmentation values randomly, all segmentation values are assigned via analysis of the image intensities. The numbering of the segmentation classes is in order of increasing mean class intensity ( reversed for T2 images ). However for some types of images ( e.g. ones with very high variance in the darker voxels ) FAST could appear to mislabel voxels at the other extreme of the intensity range. This is a limitation of the current segmentation algorithm, although performance can be improved by adding an additional segmentation class or applying some pre-segmentation thresholding to the input images.
>
> Many Regards
>
> Matthew
>
>> I'm trying to set up a segmentation pipeline for a very large (subject x time-point = ~ 70 x 20) set of sub-optimal structural data (clinical trial data: T1W, T2W, FLAIR, and PDW; no high-res MPRAGE).
>>
>> Most of the times the fast segmentation results are in the same order: e.g. GM (1), WM (2), and CSF (3). However, occasionally the order changes, which requires manual checking.
>>
>> I read in an earlier message that the values are assigned randomly. I think it doesn't hurt to ask if there is any unpublished fast option that can enforce the order of the integer value for the segmentation results or some clever way to detect the change?
>>
>> Regards
>> Gordon
>>
>
|