I am running SAMSEG to segment MS lesions from T2 FLAIR scans. I wanted to create a lesion mask in order to obtain lesion volumes and lesion counts, so I used mri_binarize on seg.mgz to obtain all voxels with the value 99. However, upon visualising this mask I see that there are holes in the mask in regions that are quite clearly lesions based on the FLAIR image. I see in seg.mgz that these voxels are instead assigned incorrectly to a different tissue, as the lateral ventricle, for example.

I wondered if this misassignation of tissue was due to the chosen threshold (0.3) so I added the --save-probabilties flag to look at the probability maps too, and I see that actually the probability map thresholded at 0.3 does not contain holes. The probability map actually seems more accurate in general and picks up on more of the hyperintensities. This is important for my desired outputs (i.e. accurate lesion counts/volumes).

So I wondered if it was possible to explain a little why these holes might be occurring in the final segmentation but not the probability map?

Would it be incorrect to derive the lesion mask from the probability map instead?

Apologies I cannot attach the images of the data.

TIA,
Evie