Hi FreeSurfers,

I’m still trying to process some high resolution data through the freesurfer stream, but it is constantly failing during several different stages. The biggest issue is the skull stripping and I’m still trying to get good results here.

Using the –hires switch during –autorecon1 normally leaves to much skull or cuts of too much tissue. Adjusting the parameters during mri_watershed (like changing the preflooding height or using the –atlas switch) doesn’t work too well in most cases.

I tried to manually edit the brainmask.mgz with tkmedit but I encountered another problem. I was using data with an isometric voxel size of 0.6mm and removed every bit of left skull from the coronal view. Afterwards I ran –autorecon2 but it did exit, so I viewed the brainmask in Slicer and noticed that probably every second slice was skipped in tkmedit.

I also tried using the –cm switch during –autorecon1 to create the .lta file for the skull stripping stage, which I was hoping to improve the results. But either mri_watershed exists with an error in which it states that the WM intensity is 0 and therefore below the csf intensity and that I should check the volume, though I can’t find anything unusual, or it exists with the error “segmentation fault”. I read that these “segmentation fault” errors seem to occur if the system runs out of memory, but as I’m having 32gb ram and 16gb swap memory I don’t think this is the reason.

Another approach I was using is to use either a brainmask created by a stream of Slicer and then importing it to freesurfer or to use a brainmask created by a conformed stream, then using mri_vol2vol to resample it to the original resolution and process the other stages. The first approach should work fine, but the volume I used wasn’t well processed during mri_fill. Most voxels were recognized as being part of the left hemisphere and only a few (about 150) were marked as belonging to the right hemisphere. And I was unable to fix this issue.

The second approach seems also to be a good idea, but the resulting brainmask is blurred due to the “upscaling”. I noticed that using the –hires switch the brainmask isn’t really used as a mask during normalization2 but is just normalized again and the output is written as brain.mgz. If I want to use the brainmask as a mask (as it should be intended) should I use it to mask the T1.mgz? Like mri_normalize –noconform –mask brainmask.mgz T1.mgz brain.mgz? In the original stream the brainmask is used here to mask the norm.mgz, which I quite don’t understand as the norm.mgz is already stripped. But maybe I just get something wrong.

Yesterday I was able get a good skull stripping by running mri_nu_correct again with T1.mgz as the input and using the –distance 25 –stop 0.0001 switches (which I read of in another mail here on the list to improve results using 3T data). But recon-all exists at the beginning of mris_fix_topology with the “segmentation fault” error on both hemispheres.

I apologize for the long text, but maybe someone can give me some input to try something else or help with existing issues.

Thanks in advance,

Regards,

Falk