Hi Achille,

Great that you managed to solve the cropping issue.

Regarding the small lesions detection problem in the superior part of the brain (I guess you are looking at juxtacortical lesions?), hard to say without seeing the data. It might be a modeling problem (i.e., SAMSEG is not able to detect some of these lesions) and/or it might be a data quality problem. Also, note that the lesion version of SAMSEG was developed for MS patients, so there might be a bias towards segmenting MS lesions.

One thing that you can easily try is to tune the lesion threshold parameter. SAMSEG creates probability maps for each structure (including lesions). This lesion map is then thresholded using a predefined value (0.3 as default). It might be that these small lesions are detected by SAMSEG, but they have small probabilities, thus they are not included in the final segmentation. You can try to save these probability maps (--save-posteriors flag) and check the lesion probability in these small lesion areas. If these areas have zero probability, then SAMSEG is failing at detecting them (although it is hard to say why and how to solve this problem without looking at the data).

Hope it helps,

Stefano

From: achille.teillac@chu-lyon.fr <achille.teillac@chu-lyon.fr>
Sent: 13 September 2021 12:23
To: Stefano Cerri
Cc: freesurfer@nmr.mgh.harvard.edu
Subject: SAMSEG estimation error

Dear Stefano,

Sorry for the delayed answer and thank you for your advice. As you mentionned, it was probably a template registration issue because when I updated my FS version to the 7.2, the cropping of 3D T1 and T2 FLAIR images stopped and the segmentation performed really well.

Another concern I have is about the segmentation of the white matter hyperintensities (in the T2 FLAIR). It seems that SAMSEG does not succeed in finding the « small » lesions located in the most superior part of the brain. Do you have any feedback on that ?

Best,

Achille