Hi all, not sure if this is a good place to post, but I asked the FS mailing list and no answer yet.
I use Freesurfer v6 to segment longitudinal scans in Huntington’s disease. I then QC the segmentations, but I only QC the cross-sectional segmentation where the failures or displacements are more clear than the longitudinal ones.
My main question is this: how do people account for failed Freesurfer segmentations at a single timepoint when using the longitudinal pipeline? Would you exclude just that timepoint and keep using the others in statistical analyses? Or would you reconsider whether the timepoint has also a corrupt scan (i.e., in scanner motion) and because the motion may corrupt the single subject template you would exclude the entire subject (or perhaps reprocess FS without the motion-corrupt timepoint).
I am trying to get a sense on how other people in the community do this if (1) the failure is simply to Freesurfer’s mis-segmentation or (2) the failure is due to the scan being of poor quality.
look at the euler number from Freesurfer and/or MRIQC to gauge scan quality.
include euler number in statistical models
I would not use a bad scan to make a longitudinal template
Beyond that, some suggestions may be specific to analysis goals and hypotheses. I can’t really say whether it makes sense to exclude a subject without knowing more about your cohort of what statistical techniques you are using.
Our processing is already done on 7000 timepoints using FS6, and we keep using that version to align with previous segmentations (FS6 and FS7 produce slightly different volumes, not sure about FS8).
About QC, I do that manually, so the problem is not on spotting failures but how to proceed once the QC fails. I used to tell my colleagues to throw out the subject if a timepoint has failed, because we had a lot of data. But now we have a study with just 70 subjects, and the question has come up if statisticians can just eliminate the offending timepoint instead of the entire subject.
As for your last tip, yes I agree that a bad scan should not be included in FS processing at all. But I have noticed when reviewing 7000 timepoints that some scans marked as good in the original study are still corrupt with motion or missing slices. I am thinking those cases should exclude the entire subject from the analyses to be on the safe side.
Again, this still will depend on your cohort, hypotheses, and statistical methods. For example, if you have 4 time points, and quality-related attrition will bring some subjects down to 3 timepoints, methods like mixed effect models should still work fine. That is not true if it is 2 time points being brought down to 1 time point.
Ah yes, statisticians do that fine. They just need to know if the other timepoints with good QC are still trustworthy or corrupt by the timepoint with bad QC/segmentation.