Hello, I have a very "dark" T1w image set where the 3dSeg is not able to differentiate the WM and GM towards the "top" of the brain. Is there any way to make this segmentation better? My ultimate goal is to run fMRIprep, but it was having problem with segmentation as well and thought I would try a manual preprocessing (just skull stripping and brain masking) so I can feed the rest of it to the pipeline again.
Below is the picture of the WM segmentation overlayed on my skull stripped anatomical (not warped to MNI yet).
have you run your structural through 3dUnifize first?
there are a ton of things one can do re: tissue segmentation but i'd start with that if you haven't already
I have actually run the 3dSeg after running SSWarper (skipping the warping part). But I was using the 3dcalc processed one (the skull-stripped copy of the input with no other processing).
I have re-run the 3dSeg on the second pass skull-stripped dataset which should include the 3dUnifize in its step..
The top is the non unifized one and the bottom is the unified one! It does make it much better.
I do have a quick question about the difference though, does the 3dUnifize do anything else than improve the intensity of the WM? I am trying to keep my own preprocessing as minimal as possible before running it on a pipeline.
that is definitely an improvement. 3dUnifize simply "evens out" the intensities across the image (aka "bias corrects" it) which then makes it easier for a tissue segmentation algorithm to perform successfully. 3dUnifize is conceptually similar to the "N4" algorithm in ANTs if you've heard of that one. What neither program will really do, however, is improve the actual contrast between tissue types (e.g. gm and wm); that's usually set at the scanner via multiple acquisition parameters (or specific sequences, e.g. MP2RAGE rather than an MPRAGE). So in essence, the bias correction sets the segmentation algorithm up for as much success as is possible given the tissue contrast you already have.
That was exactly what I was looking for. I have just fed the new T1 into the pipeline, so I am keeping my fingers crossed so that the segmentation works better on it!
3dLocalUnifize is a newer version of unifizing that might work better.
3dUnifize has a useful -GM option for gray matter scaling.
3danisosmooth with just a few iterations may be a good way to smooth some noisy data (-iters 3).
Modal smoothing of the segmentation with a low radius can remove some the small blips within the tissue segments. Use 3dLocalstat -stat mode -nbhd "SPHERE(-1.8)"
I have a question about these steps compared to SSwarper.
Are all of these steps included in the SSwarper?
I am getting different results of skull stripping compared to SSwarper and running them myself.
So far, I have run the 3dUnifize (3dLocalUnifize unfortunatlly did not work better for me!) & 3danisosmooth with the iterations that you suggested, and then tried skull stripping by 3dSkullStrip.
Although my manual skull stripping seems to be performing poorer compared to SSwarper. Would there be a reason why I am getting different results?
These are all different processing steps although SSwarper has its own recipe. Differences are expected. SSwarper combines 3dSkullstrip with template alignment. There is no explicit segmentation step except by transformation of atlases. sswarper2 uses another recipe to do the same thing - with no skullstripping step.
The
National Institute of Mental Health (NIMH) is part of the National Institutes of
Health (NIH), a component of the U.S. Department of Health and Human
Services.
