Hi FreeSurfers,

Due to an accident and after several hundreds of hours of work I was finally able to completely (actually not completely, but it seems to work) process high resolution data with an isometric voxel size of 0.6mm without any manual intervention, while using the advantage of all atlases. I would like to share my experience with you and especially would like to know if and how some of my individual changes to the recon-all stream may affect the results.

The data I used was acquired at 7T (Siemens), so my volumes were not very homogenous. To get rid of most of the inhomogeneities we are using a method at our lab in which two volumes are acquired, a MPRAGE and 3D gradient echo. Both volumes are co-registered and afterwards the MPRAGE is divided by the 3D GE. Everything further has been done directly using FreeSurfer.

First you need to run recon-all –autorecon1 –cm –noskullstrip –s <your subject>. Using the –cm flag the volumes are not conformed to 1mm^3 voxel size and 256^3, but to the smallest voxel size (in my case 0.5990mm^3 and 385^3). It is crucial to not run –skullstrip using the high resolution data as this will not work! You need to go to your subject/mri folder and use mri_convert –cs 1 nu.mgz nu.conformed.mgz, this creates a conformed version of the nu.mgz. Now you should run mri_em_register –skull nu.conformed.mgz $FREESURFER_HOME/average/RB_all_withskull_2008-03-26.gca transforms/talairach_with_skull.lta.

This has to be done using a conformed nu.mgz as in the following stage mri_watershed seems to conform the input file (Will it be really conformed? The output volume doesn’t suggest this.) and if the talairach_with_skull.lta is created using a not conformed volume one gets an error stating that the WM intensity is lower than CSF intensity and the stage is exited. The nu.conformed.mgz is only used during this stage, so your high resolution volumes will never (Never? At least the output volumes are not.) be resampled, conformed or whatever to a lower resolution.

After this has been done, one can use recon-all –skullstrip –s <your subject> and take advantage of the atlas! Though the input volume (T1.mgz) seems to be conformed during mri_watershed, the output volume (brainmask.mgz) will have the correct resolution. If the brainmask.mgz doesn’t look well, try adding the –wsatlas switch first, as this improved the skull stripping if only really small bits of skull and dura were still left or small pieces of the temporal lobe or cerebellum were cut off. If this doesn’t help much adjust the watershed threshold using the –wsthresh <value> switch (but still use –wsatlas) like described in the wiki.

Now run recon-all –gcareg –canorm –careg –careginv –rmneck –skull-lta –calabel –s <your subject>. Even though the aseg.mgz should look perfectly fine, one needs to stop here, as the –normalization2 stage can be run only if the aseg.mgz is conformed (Why does it need to be conformed?). This is something you don’t want, so go to the subject/mri folder again and run mri_normalize –mask brainmask.mgz norm.mgz brain.mgz.

Right now I'm running the missing stages of -autorecon2 (being at the -fix stage which will take some time) and I think it should work totally fine. Maybe the following stages using the aseg.mgz might not run as those maybe want the aseg.mgz to be conformed as well, but I don't know yet. Additionally in my case this wouldn't be a problem as I just want to measure the cortical thickness.

I would highly appreciate if you could answer the questions in the brackets and how the results might be effected by the changes I have done.

By the way, I tried running mri_nu_correct.mni using the T1.mgz as the input volume and running mri_normalize again afterwards to improve the inhomogeneity correction. The result looks very good, but parts of the skull and dura are not erased during the skull stripping stage (as they seem to match the intensities of GM and WM). If I adjust the watershed threshold so these are erased, pieces of the cerebellum and temporal lobes are erased additionally. Would it be possible to run mri_nu_correct using the brainmask.mgz as input volume and mri_normalize afterwards, so the inhomogeneity correction can be further improved? Will mri_nu_correct.mni work fine on a volume without skull?

Thanks in advance,

Regards,

Falk