Dear Katy,
> I am running very standard flirt & fnirt commands (as described on the fnirt webpage) to register T1 MPRAGE to the MNI152_T1_2mm template. I used skull-stripped images for flirt & non-betted images for fnirt, as suggested.
>
> I have noticed in some subjects there is quite a bit of stretching around the edges of the brain. This occurs to varying degrees in my subjects, and is most noticeable at the top of the brain and bottom of the temporal lobes (so in the z direction). I have seen other people post what sound like similar problems, with not much resolution.
there are two reasons that together cause this problem.
One is that across scanners and sequences there is a large variability in the signal from, what we think is, marrow (i.e. fat) in the skull bone. In practice this means that in some T1 weighted scans you see a "sheet" of high signal (sometimes higher than in the brain even) close to the edge of the brain, and in particular at the top of the brain. In other scans, and in our template, this signal is largely absent which gives a nice "edge" as the signal drops from in the brain to outside the brain. Fnirt, and I imagine most registration methods, struggle when asked to register a scan with this signal to a scan without (such as our template). Surprisingly often it still does a decent job in these cases, but that often mean that i "pulls" this signal away from the brain and registers it to the scalp signal (which is always present).
The second cause is that the EPI->Structural registration often struggles a bit, and in particular around this region there is often some misregistration.
These two problems jointly causes what you see. Fnirt attempts to pull non-brain tissue away from the brain, but because of the EPI->Structural misregistration it instead gets hold of brain tissue in the EPI image and pulls this away.
Things you can do are
1. Use a structural sequence that doesn't give you that signal. In particular a bit of fat-sat can be very useful.
2. Make sure you acquire fieldmaps, which will improve EPI->Structural registration.
Things we are doing:
MJ has implemented Doug Greves new EPI->Structural registration method and it will be part of the next FSL release.
Things we might do:
There is talk of making a template with this signal present, and then change fnirt so that it registers to a linear combination of the two templates.
>
> I serendipitously realized today that this problem seems to go away (at least in the subjects I've test so far) if I use the undilated (MNI152_T1_2mm_brain_mask) instead of the default dilated one (MNI152_T1_2mm_brain_mask_dil).
>
> Since the distortion is gone when I use the undilated mask I am inclined to use it for all subjects. Is there some reason that this is ill-advised? And along the same lines, why is the dilated mask chosen as the default to begin with?
It's great you managed to solve your problem like this. I wouldn't want to use it as a general solution though since it pretty much means that the edge information when going from high signal in the brain to low signal outside the brain is lost. And that is potentially important information to guide the registration.
I would strongly recommend anyone planning to use fnirt, and especially for FSL-VBM, to use a sequence that minimises the signal from this area (i.e. the marrow in the skull bone, I think). When developing a clinical T1 structural sequence what matters is good CSF-gray-white separation and good SNR, and I imagine physicists and physicians alike don't care about a little signal just outside the brain. For automated tools like bet, fast, fnirt etc things are quite different and it can potentially matter a lot. So talk to your physicists and explain the problem, and hopefully they can come up with a quite minor tweak to your sequence that sorts the problem out.
Good luck Jesper
|