Hi,
If your lesions are very clear on the FLAIR with this segmentation then an easy way to fix this CSF class problem (which is a known issue from having a broad class in a mixture model, with dominant tails at both ends) is to do the following:
1 - take the grey matter segmentation (or the white matter) and calculate the mean intensity for this (either binarise the PVE or use the hard segmentation as a mask in fslstats)
2 - take the CSF segmentation (a binary mask from either the PVE or the hard segmentation) and multiply this by the original intensities, and then threshold using the mean intensity from step 1, and create a mask
This then gives you a mask for the high intensities in the CSF and you can use it to get just the low intensities too. From this you should be able to separate out your bright lesions from the dark CSF.
All the best,
Mark
Dear FSL group
I am trying to segment FLAIR images to automatically detect brain and lesion in some clinical MS scans. Just running FAST on brain-extracted FLAIR images (3 class segmentation on T1 images) yields some weird results with both CSF and lesions (opposite ends
of the intensity spectrum) being classified as CSF. Is this a product of FAST forcing high intensity voxels (presumably fat signal in T1 images) to be CSF (or more specifically "non-brain")? Is there a way I can get around this? These images have pretty good
differential intensities for CSF, brain and lesion so it looks like a relatively simple problem compared to grey/white matter segmentation.
cheers
Scott
---
Scott Kolbe
NHMRC Postdoctoral Fellow
Department of Anatomy and Neuroscience and
The Florey Institute
University of Melbourne
+61383441915 / +61432340471
