Dear Julien,
Thanks for your very interesting email! One thing at the time:
At each voxel, the discrete segmentation is given by the label with the highest probability. Therefore, there is no fixed threshold. One voxel my require a probability of 0.51 to "win" (if another class has probability 0.49) whereas in another voxel 0.1 may suffice (if you had 18 other classes with probability 0.05; this is obviously an extreme example to illustrate the point). If you want to focus on a specific region (like the MGN or LGN), then using the individual posteriors may help.
Your second paragraph is correct. We did label a bunch of MGNs and LGNs manually to make the atlas reach further into them. We've compared the LGN segmentation with visual task fMRI and the average location is correct (the exact boundaries are hard to get due to the faint contrast). What is the exact problem with the MGN? Would you mind sharing some pictures?
Cheers,
/Eugenio
Juan Eugenio Iglesias Senior research fellow CMIC (UCL), MGH (HMS) and CSAIL (MIT) http://www.jeiglesias.com
On 6/16/20, 09:30, "freesurfer-bounces@nmr.mgh.harvard.edu on behalf of julienbesle" <freesurfer-bounces@nmr.mgh.harvard.edu on behalf of julienbesle@gmail.com> wrote:
External Email - Use Caution
Hello,
I'm interested in locating the MGN and LGN using segmentationThalamicNuclei, but there is a mismatch between their location in the discrete segmentation (ThalamicNuclei.v10.T1.mgz) and the underlying probability maps obtained by setting environment variable WRITE_POSTERIORS to 1. While the LGN in the segmentation has many voxels within the corresponding area of maximum probability, many high-probability voxels are missing from it. For the MGN, the segmentation volume is restricted to a small number of voxels in the most lateral/posterior/superior part of the area of max probability, but most high-probability voxels are absent from the segmentation. Any reason why this could be happening? Is there anything I can do (parameters I can change in the script) to fix this?
The defaut Freesurfer segmentation ("thalamus proper") is generally larger than the segmentation output by segmentThalamicNuclei, except in the area of the LGN, where the thalamic segmentation extends more laterally/inferiorly, as if it had been extended in order to include the LGN (although imperfectly). Unfortunately, this does not seem to be the case for the MGN (which is my primary area of interest).
To solve the issue, I'm thresholding the posterior probability maps at p>0.2 for both LGN and MGN and adding these voxels to the discrete segmentation, but that's not very satisfactory. Actually, it wasn't very clear to me from the published paper how the discrete segmentation is obtained from the posterior probability maps and if any cut-off value is used to decide whether to include any voxel in the final discrete segmentation?
(Note that I'm using T1-weighted MPRAGE data at a resolution of 0.7 mm isotropic)
Thanks
Julien
_______________________________________________ Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer