Hi Jeff,
On Tue, 2010-10-19 at 13:40 -1000, Jeff Sadino wrote:
Thank you for your feedback on my previous question regarding the surface area measurement. We found it very helpful to move forward.
I am beginning (again) to do longitudinal processing using fs4.5.0, but I just want to make sure that I understand everything right before starting this time. We have about 150 subjects with timepoints varying from 2-7, for a total of somewhere around 750 scans. All have already been processed cross-sectionally.
My recommendation to process with 5.1 once it is out. 4.5 was not designed with edits in mind. Some edits might work, other not (see the old emails). Even in 5.1 not every edit situation will be solved, but we hope to be able to give clear advice on what should be edited where.
- Reading through the wiki, the main thing to check on the
cross-sectional runs are the control points. I'm confused by this though because the base creation only needs the reconstruction ran through the norm.mgz step to start. But we won't know if we need to add control points to the cross-sectional scan until after the surfaces are made, later in the stream. So it seems like I need to run the cross-sectionals through the norm.mgz step to start the base creation, but then before I run the -long part, I need to run the cross-sectionals through the surface reconstruction so that I can check the control points so that they can be copied from the cross-sectional run to the -long run? Are there any other edits I need to look at before starting the -base run? The cross-sectional brainmasks? The brainmask QA is just to make sure that the skull and other things are correctly stripped away? What about intensity normalization and gm/wm contrast on the norm.mgz image?
For the -long runs BOTH -base and cross need to be fully processed. To be on the safe side when you have edits, run the -base after all cross sectionals are done. If no edits are involved, you can start the -base once the norm.mgz is available in the cross runs. If manual control points are available in the cross runs, we will try to use them in the long runs. The base is constructed from the norm.mgz thus intensity corrections should be indirectly incorporated into the base image automatically. The base brainmask is created from the cross sectional brainmask (but sadly in 4.5 this is before edits, will be fixed in 5.1).
- Also, in
http://www.mail-archive.com/freesurfer@nmr.mgh.harvard.edu/msg12895.html, > Martin says "the cross sectional asegs are 'fused' (probabilistic
voting), to initialize the labeling in the longitudinal runs." Can someone explain what this 'fused' means? This is the only place I have read this besides the short description in recon-all --help. So it seems like I need to finish the cross-sectionals pretty much completely before doing the -long run.
Yes, both base and cross need to be finished completely before -long. The fusion is a probabilistic voting on the labels (if the intensity at a given voxel is similar, the probability is high that this voxel has the same label). Basically this is a weighted average of the cross sectional aseg.mgz and thus automatically incorporates aseg edits. This fused aseg is used to initialize the long (it is not the final longitudinal aseg!).
- Why is the brainmask from the -base run used in the -long run? I
can see how this increases consistency between timepoints, but it seems like there would be a larger benefit of using the brainmask from the cross-sectional run?
The idea was: - head should not change, thus brainmask should be identical for all time points. - using the same info reduces variability - a single file needs to be edited instead of all time points brainmask If we start seeing cases where this does not work, we'll have to allow a flag to copy brainmask edits from the cross or to edit the longitudinal individually.
- I read that if the -base is stopped halfway through due to
earthquakes, then it can be restarted from some point in the stream. I will just restart it from the beginning for ease of use. My question is what if the -long stops halfway through? I forget the command I use to restart from the last step in the cross-sectional stream (our cluster is down now), but can I do the same thing: recon-all -long tp1 longbase -restart_from_last_step_command?
We have not tested any restarting functionality. Feel free to test, but until we know for sure, always restart if either base or long stop. We will have to look into some of this as we start incorporating edits.
And then I had some questions specific to our data, if you would permit me =)
- We have timepoints from the past 6 years, and so our scanning
parameters have changed. Our old data is 1x1x1.4mm and 144 slices and our new data is 1x1x1mm and 160 slices. Some subjects have an acceleration of 2, while most have no acceleration. Our scanner is a Siemens 3T Trio. Will I need to exclude any of these timepoints? From both -base and -long? The thread
http://www.mail-archive.com/freesurfer@nmr.mgh.harvard.edu/msg14014.html says that the new stream significantly reduces these effects, but I just wanted to doublecheck.
For any longitudinal study never - change the scanner - change scanning parameters - (even scanner software changes can introduce bias) You'll never know if what you see is due to anatomical change or differences in the processing. If you're lucky and have test-retest data just before and after the change, you might be able to show that there is no significant effect and then still use the longitudinal results (or correct for any changes). But this gets messy. And yes, the long is much better in dealing with these situations, as it reduces variability. But that does not change the underlying problem of introducing bias.
- Because of the way our group is set up, sometimes subjects will
have more than one mprage - from different days, say one week apart - for each timepoint. Even though I will only use one of these mprages for each -long timepoint, will the -base be better if I include all of them? As a tangential question, could I use these close together visits and average their mprages in the -motioncor step? My feeling would be no, that I could only do that if the subject does not move, but I wanted to check.
I would not include it in the base. Include only the time points that you also process longitudinally. I would also not use it in motion correction. The reason is not that they have moved (because that is undone by registration in the motion correction step anyway), but who knows if there are effects due to dehydration etc between the two visits. But this is just my opinion.
Sorry for the long questions! Thank you for all of your great help all the time! Jeff Sadino
Hope that helps ,Martin
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer