Hi,

I would like to use CMTK to correct susceptibility distortion in our T2*-weighted EPIs and am having problems where the corrected images look like the images were overlapped with a mirrored version (see attached) -- which perhaps is already apparent in the distortion/Jacobian determinant map. I was trying to roughly follow "Unwarping Echo Planar Images Using CMTK" as a guide although we are not dealing with diffusion-weighted images.

For some sessions, we only have an additional gradient echo image in the reverse phase-encoding direction as our T2*-weighted images. In this case, I treated a random T2*-weighted EPI image as b0_fwd and the extra GRE blip reverse image as b0_rev, which I fed as inputs to epiunwarp. I then applied the resulting Jacobian determinant from epiunwarp to all the T2*-weighted EPIs (as inputs to reformatx and imagemath).

For other sessions, we also have acquired two spin echo images with the same parameters as our T2-weighted EPIs in different phase encoding directions to each other. I treated these SE images as the b0_fwd and b0_rev inputs to epiunwarp.

Either way, we are having the same problem, and I want to say with my naive eye that this is already a problem during epiunwarp (hence I'm not including all the code here yet to reduce visual clutter). Could someone point out what I am doing wrong? I've tried the --no-flip switch, our images are acquired A-P (so I didn't change the default --phase-encode switch).

Could someone also explain to me why it is necessary to run and imagemath (step 3 "pixel-weise computations, outputting the distortion-corrected image" in the pdf guide to unwarp EPIs) after running reformatx which "applies the computed deformation to correct for susceptibility" (what is this set up applying and outputting if imagemath seems to be doing the applying and outputting)?

Many thanks in advance for any help and advice!
g

Hi Gina -

The EPI unwarping implemented in CMTK's epiunwarp tool is primarily designed for DWI, not fMRI, because there are systematic intensity fluctuations in the BOLD images that are independent of the distortion. By contrast, the assumption is that intensitites in b=0 DWI images are affected by distortion via signal pileup, but are otherwise stable.

That said, I suspect the reason you are getting the "mirrored image overlay" effect is that the algorithm requires a pair of images acquired with opposing phase encoding direction. For example, you'd need one image with phase encode A->P, and a second with phase encode P->A. If you have that, but you still get the mirror effect, then it should be sufficient to simply flip one of the two images along the PE direction, ie., in the previous example along the AP direction (if your images are oriented correctly, that would be the y axis in the image files).**

Regarding the use of the imagemath tool - there are two effects of the distortion on the images: one is a warping along the PE direction, the other is intensity increase or decrease, depending on whether the warping stretches or expands pixels. Basically, you have to multiply (or divide, I can't remember) the original pixel intensity with the Jacobian determinant of the warp at that position to get the corrected intensity. It is simpler to have the two steps be done independently, and because the intensity correction does not involve interpolation, there is no quality penalty from doing the step separately. Also, for DWI the intensity correction would be applied equally to the same pixel in all channels of the DWI, so for most tensor reconstruction methods it would probably not have any effect on the estimated tensor anyway.

Hope this answers your questions.

Best,
  Torsten

**Edit - or use the --no-flip command line switch, as you already are doing. By the way, with vs without that switch, you should see the "mirror overlay" in one case but not the other. Is that so? Anyway, you do need an image pair with opposing PE, there's no way around that.

Hi Torsten,

What problems would you foresee in using epiunwarp for fMRI? If I understand correctly, there is signal dropout in gradient echo images the way there isn't in spin echo -- I guess you're also saying that epiunwarp would not deal with these situations (but what does it do when it encounters them?)? We have quite a bit of distortion in our functional images (multiband factor 4).....

We do have images with opposing phase encoding directions (perhaps, though, the reverse phase encoding image got flipped during conversion). The --no-flip switch didn't help, but I ran convertx with --flip-y on the reverse phase-encoding image before running epiunwarp -- that helped with the mirror effect.

The correction with the additional gradient echo image in the reverse phase encoding direction as our functional scans, seems decent to my naive eye -- could I get your expert opinion (see attachment, p.1, left side)?

However, now I have a problem and a question.
When correcting the functional scans with the deformation matrix calculated from the pair of spin echo images (one A->P, other P->A), it seems like the correction is going in the wrong direction for the functional scans (though the correction on the SE images look pretty good -- see attachment, p.1, right side)

I also am wondering whether one should use reverse or forward Jacobian (attachment, p.2)?

I thought I'd continue this thread (as all related to SDC for BOLD images)....

Thanks a lot again!
Gina

Hi Gina -

This is the third time I am typing this, since thanks to f***ing Chrome and f***cking Comcast I have lost the previous two attempts at answering your questions. So I'll be exceedingly brief.

a) You cannot compute unwarping from one type of images and apply to another. All images have to be the same (minus reverse PE in one reference image)

b) The issue with signal in fMRI is that the BOLD effect modifies image intensity on top of signal pileup due to distortion. Therefore, it will affect the recovery of the warping transform itself. You may be able to mitigate this issue by acquiring several images for unwarping, say 5 with fwd and reverse PE, and averaging the images in each group before unwarping.

Best,
  Torsten