Dear Mark,
>the best solution is to use prelude & fugue to unwarp the EPIs
>and sigloss to calculate a cost function weight based on the estimated signal loss
>de-weight (assign zero) to all areas on the EPI where
>there is clear distortion or signal loss
>(this will involve some subject generation of the weighting images
Do you have any scriptable suggestions (i.e. "default" settings) how to get cost function weights for "average" individuals from the fieldmapping? So far, I have always had very reasonable results by just registering the example_func EPI to its unwarped version (6 dof), the latter to the highres T1 and the highres T1 to the avg152T1_brain without using cost function weights. However, I'd like to check if the weighting would indeed further improve the registration.
Thanks a lot in advance and very best regards from Wuerzburg-
Andreas
-----Ursprüngliche Nachricht-----
Von: FSL - FMRIB's Software Library [mailto:[log in to unmask]] Im Auftrag von Mark Jenkinson
Gesendet: Dienstag, 24. Januar 2006 12:29
An: [log in to unmask]
Betreff: Re: [FSL] Registration question
Dear Richard,
We also see shifts in the z-direction with simple EPI registration. These are made better by weighting the ventricles, but the best solution is to use prelude & fugue to unwarp the EPIs and sigloss to calculate a cost function weight based on the estimated signal loss. However, all of the latter steps involve the acquisition of fieldmap data.
If you do not have fieldmap data then the best thing to do is to make a cost function weighting image that does the following:
- de-weight (assign zero) to all areas on the EPI where there is clear
distortion or signal loss (this will involve some subject generation
of the weighting images, which is the trade off for not having
fieldmaps)
- weights the ventricles highly (if these have good contrast)
- in the mid-sagittal slice of the EPI, find the main blood vessels
near the
ventricles (these usually appear dark) and also de-weight these
Note that the fact that blood, and especially large veins, appear very dark in the EPI but usually not in the structural (due to T2* weighting in the
former) and therefore do not match well. Although the cost functions correlation ratio and mutual information are somewhat insensitive to mismatched intensities - small structures like this, which do not heavily weight the statistics of the whole image, can still cause problems of mis-registration as the mismatch may be slightly less in the mis-registered position than in the well-registered position (since neither match well).
In general, the de-weighting of the large signal-loss and distorted areas in the inferior portion of the brain is the most important step as this often causes a lot of erroneous shifting in z.
As to the "fix" of add 3.0 to the z translation - I think that may work OK for images from the same scanner and similar individuals, but I wouldn't trust it in general. I would definitely advocate trying the above weighting options (and various cost functions - corratio, mutualinfo, normmi) to see if you can get the same or better results. And also think about using fieldmaps in the future.
I hope this is helpful.
All the best,
Mark
Richard Albistegui-DuBois wrote:
> This is a sort of followup to an earlier question of mine, when I was
> looking for ways to improve the registration of a BOLD EPI series to
> a T1 spin-echo structural. The problem ended up being susceptibility
> artifact in the EPI, and weighting the registration toward the
> ventricles improved things.
>
> I have since found that the weighting did not always improve things
> significantly, and seemed highly dependent on very small variations
> in the weighting mask. However, I have found that simply adding 3.0
> to the z-translation component of the affine matrix from the weighted
> transformation produces nearly perfect results.
>
> I am not sure why the weighted transformation would be so good, but
> consistently off by a small distance in Z, but I thought I would post
> this to share the solution I found, and to ask if this behavior is
> normal.
>
> Thanks, all.
>
> -Richard
>
>
> Richard Albistegui-DuBois
> UCLA NeuroRehab, lab of Bruce Dobkin, MD.
> Office: 1-132 Reed
> Phone: 310-825-4016
> Mobile: 310-774-1305
> Fax: 310-794-9486
> AIM: dubistegui
> email: [log in to unmask]
|