How do I train samseg with my own data

[ZauberZ]

In practical application, the segmentation effectiveness of a single MRI structure is not consistently robust in certain areas. Therefore, I intend to train a new probabilistic model tailored to my own data.
Thanks.

[ste93ste]

Which areas/brain structures are not well segmented? Before re-training the probabilistic atlas (which we do not support at the moment, and require a deep understanding of SAMSEG), you can try to play with the GMM parameters (--gmm flag in run_samseg). You can pass a txt file parameter like the one here, where you change which structure of interest shared the same intensity parameters. Look at the original SAMSEG paper for more information on how to change this file.

If you still want to build your atlas, you need to compile the gems library with the following flag: GEMS_BUILD_EXECUTABLES ON. Then use the command kvlBuildAtlasMesh. As said before, we currently do not support this functionality, so we cannot help you with the building of your atlas.

[ZauberZ]

Thank you very much. When using only T1w images, it's not easy to segment the entorhinal cortex from surrounding non-brain tissues accurately.