I have T1w and T2*w brain volumes from 70 healthy, elderly subjects (75y). I extracted the brains, removed the bias field from each volume and co-registered them. Now I am using fast with both the T1w and T2*w volume for tissue segmentation:
fast -n 5 -S 2 -N -I 0 -o OUT T2SW_brain_restore T1W_brain_restore
1) There are slight differences (but differences) in the pve segmentation results (OUT_pveseg) when I swap the input volumes (e.g. T1W_brain_restore GRE_brain_restore, instead of the ordering above). I included an example which shows both white matter masks (PVE.jpg: yellow=agreement, green or red = one of the masks is bigger). I haven't found any documentation about this so I am wondering if that's intentional or a limitation. Additionally, the OUT_pve_* volumes have different ordering, which also depends on the ordering of the input volumes.
2) I need 5 classes (-n 5) to get the best segmentation result. I find that fast uses 2 classes for the gray matter for each of the 70 subjects. I used T2SW_brain_restore, T1W_brain_restore and the two gray matter masks of each subject to construct two histograms, one for each volume, which show the distribution of the volume intensities and the distribution of the gray matter intensities (Hist.jpg: cyan=gray matter). I noticed that I have two distinct gray matter distributions (Gaussian shape, I think) in case of T1W, but just one distribution (Gaussian shape, I think) in case of T2*. Are there any restrictions on the covariance matrix for each class (e.g. symmetric?). If there are than that would explain why I need two classes for the gray matter since the gray matter distribution of T1W_brain_restore seems among other thing a little broader than that of T2SW_brain_restore.
Any other suggestions are very welcome.
Thanks and best regards,