aseg.presurf edits are undone or incomplete when brainmask.mgz edits are involved - Freesurfer

16 Jun 2022


      External Email - Use Caution
Hi FreeSurfer Team,
Our group is in the process of switching to Freesurfer 7.1.1 from 5.3. We have been taking advantage of Freesurfer 7's aseg.presurf volume to add in correction of unlabeled cerebellum to our standard edit checklist rather than our previous workflow of having a Freesurfer fail our quality control standards and be excluded from analysis if too much cerebellum is unlabeled on aparc+aseg segmentation. We're consistently running into an issue when we have a Freesurfer that requires corrections to both brainmask.mgz and aseg.presurf.mgz.
An example of when this happens is if the Freesurfer output labels dura as cerebellum cortex on aparc+aseg (which requires erasing the included dura from brainmask.mgz) and the output also has an exclusion of cerebellum from the aparc+aseg segmentation (which requires correctly labeling the excluded voxels on aseg.presurf.mgz). We have found that if we edit both the aseg.presurf.mgz and brainmask.mgz volumes and then rerun recon-all with both, the cerebellum exclusion segmentation fix does not "take" completely and individual voxels in the cerebellum that were corrected in the edited volume are still left unlabeled. We have 100% confirmed that our labeling edits to aseg.presurf are correct and complete before this happens. To reduce confusion and attempt some consistency across our editing team, we relaunch recon-all with the same flags every time: -autorecon2 -autorecon3 -qcache
To work around this issue, we implemented a workflow where editors will make brainmask.mgz corrections, then rerun recon-all (rerun 1), and then make aseg.presurf corrections, and rerun recon-all again (rerun 2). This seems to work although it is frustrating since we are trying to minimize the number of reruns of recon-all and we typically tell people to stop and make a final quality control decision after 3 reruns.
What we are finding is if another issue pops up on brainmask.mgz after rerun 2 that then needs corrections (e.g. other areas of dura are incorrectly included in the pial surface or aparc+aseg cerebellum labeling but weren't included before so weren't erased) then after making those edits and rerunning recon-all, the previously corrected aparc+aseg cerebellum segmentation shows "missing" unlabeled voxels, same as if we had run recon-all once with both aseg.presurf and brainmask volumes edited.
The attached image shows an example of our process and the issues we're seeing. The end result (F) also happens if we edit both brainmask.mgz and aseg.presurf.mgz edits and rerun recon-all once with both included.
screenshot A is the aparc+aseg.mgz of the original recon-all output (run 0) where cerebellum is unlabeled (red arrows). There is also visible dura labeled as cerebellum that we need to fix (yellow arrow).
screenshot B is the aparc+aseg.mgz of the recon-all rerun output after just the brainmask.mgz was edited and saved (run 1). Additional segmentation issues have appeared in addition to the original ones (red arrows). This screenshot also shows an example of the type of edit we are doing to brainmask - we have erased dura from brainmask.mgz to correct the cerebellum cortex labeling and the edit worked (yellow arrow).
screenshot C is the aseg.presurf.mgz volume after run 1 before editing for the cerebellum exclusion issue.
screenshot D is our edited and saved aseg.presurf volume BEFORE we run the second recon-all relaunch. We have corrected the labeling (green arrows).
screenshot E is the aparc+aseg.mgz after a recon-all relaunch with only the new edit to aseg.presurf. (run 2). The cerebellum exclusion error has been fixed in aparc+aseg (green arrows). This run 2 of recon-all caused more issues to show up that required more brainmask.mgz edits.
screenshot F is what we see if we edit brainmask.mgz again and rerun recon-all (run 3). The output of run 3 shows missing voxels in the cerebellum where we had fixed them before. Unlabeled voxels appear throughout the areas that were edited in C and were corrected in D.
We try to minimize edits and maintain consistency across our editing team by asking people to only erase dura from brainmask.mgz that has been erroneously included in the cerebellum labelling or in the pial surface (rather than other dura that also happens to be included in the brainmask.mgz). We also ask people to relaunch recon-all with the same -autorecon2 -autorecon3 -qcache flags each time for consistency across our data. We really really want to be able to cut down the number of times we are relaunching recon-all but we are unable to fix both these issues at once (it ends up like screenshot F after 1 rerun).
Is there a way to prevent the aseg.presurf.mgz edits from "undoing" voxels when brainmask.mgz edits are included or done after aseg.presurf? Any assistance or alternative techniques for solving this would be much appreciated.
We are primarily using the Freesurfer 7.1.1 Docker container but I have also been able to replicate this issue on CentOS 8 and Ubuntu 20.04 with build stamp freesurfer-linux-centos7_x86_64-7.1.1-20200723-8b40551.
Thanks so much,
Sarah
________________________________
The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail.