Bruce:
That's pretty interesting...
this is probably due to the partial volume correction we do when computing the volumes.
Is this algorithm described somewhere? I can see the intent, but your description doesn't completely illuminate how you decide which input voxels contribute to which output voxels, and how much.
Graham
Basically we assume that each voxel is a linear combination of the tissue at that voxel and the next most likely label that neighbors that voxel. The weighting is computed as mixture weight for the local class mean for each class that explains the voxel intensity. Our internal studies show that this improves test-retest reliability and presumably also accuracy.
cheers, Bruce
Hi Graham,
no, I think that was the only description. It's a small but significant effect. You'll be happy to hear that we're making good progress on the open source front thanks to Dennis Jen, Silvester Czanner and Nick Schmansky, and are hoping to post something before the end of the year.
cheers, Bruce
On Sat, 9 Sep 2006, Graham Wideman wrote:
Bruce:
That's pretty interesting...
this is probably due to the partial volume correction we do when computing the volumes.
Is this algorithm described somewhere? I can see the intent, but your description doesn't completely illuminate how you decide which input voxels contribute to which output voxels, and how much.
Graham
Basically we assume that each voxel is a linear combination of the tissue at that voxel and the next most likely label that neighbors that voxel. The weighting is computed as mixture weight for the local class mean for each class that explains the voxel intensity. Our internal studies show that this improves test-retest reliability and presumably also accuracy.
cheers, Bruce
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
Bruce Fischl -
Graham Wideman