Hi,

The angulation and initial orientation of these images seems quite extreme, and so my suspicion is that some of the intermediate steps might not be working correctly.  Have you tried running fslreorient2std on all of your input images (including the fieldmap)?  That might make everything work better.

All the best,

Mark

On 15 May 2014, at 18:13, Julia Huntenburg <[log in to unmask]> wrote:

Hi Mark,

the smearing of signal mostly from OFC and/or brainstem occurs in the functional images registered to the individual's structural space with epi_reg. I did try to optimize the mask size but I cannot completely get rid of the smearing. I also tried the epi_reg correction with topup derived fieldmaps and have the same problem in general. Please find two snapshots of epi_reg outputs attached for an impression on what I mean.

I am implementing these corrections for comparison with another approach. To make it a fair comparison I am quite eager to get the best possible solution. I'm happy to share the commands and data if that helps.

Many thanks,

Julia

2014-05-09 21:00 GMT+02:00 Mark Jenkinson <[log in to unmask]>:

Hi,

thanks again for these clarifications. Do I understand correctly - for distortion correction of functional data with topup I could simply use the topup output field as it is (or the _fieldcoef?) for input to epireg?

You can use the topup output (from --fout) after scaling it to rad/s (by multiplying the Hz by 2*pi) as the input fieldmap to epi_reg.  The only tricky thing might be getting a non-brain-extracted magnitude image.  The unwarped images from topup are good for the brain-extracted magnitude image, but you may need to either edit the epi_reg script to avoid the non-brain-extracted registration step, or to register a non-brain-extracted image to the space of the topup output to act as a surrogate.

I can only repeat, this detailed support is highly appreciated.

Best, Julia

PS on the fieldmaps: when I used too small of a mask for preparing the fieldmap for my data, epireg started to strongly smear the signal in the OFC where the EPI was not covered by the fieldmap. I am just mentioning it, because before it sounded to me like it would always be the better option to use a smaller mask when in doubt. I don't know whether this is just something specific to my data though.

Was this smearing in the outputs registered to structural space, or after transforming into standard space?  It is fairly common to see a bit of smearing in standard space, but it isn't common to see in the structural space outputs, or the unwarped functionals from epi_reg.  However, the mask really only needs to be made smaller to avoid noisy voxels in the standard fieldmap phase images (which are just at the edge of the brain).  So normally this doesn't make things so small that there is a problem.  There is a step inside the processing that also tries to clean up these noisy voxels, but we recommend a tight mask to be a little conservative. If you are getting bad results then try to run things with a slightly larger mask and see if that works better.

All the best,

Mark

2014-05-06 8:39 GMT+02:00 Mark Jenkinson <[log in to unmask]>:

Hi,

What I wanted to make clear here was that topup itself was not performing the distortion correction, like it can for diffusion data (since topup cannot be used in this mode for functional data).  

Although there is the option in the HCP scripts to register the SBRef image directly to the spin echo output from topup, then convert the topup-derived fieldmap into a warpfield and apply this directly to the functional scan, I would still advise people to use flirt (or epi_reg) for the distortion correction in general.  This is because the registration of the functional image (SBRef) to the spin echo image will only work accurately if (a) the distortions are exactly matched in the acquisitions, which won't typically be the case, and (b) the signal loss in the functional images does not affect the registration accuracy.  For the HCP these two conditions will be met, but others would need to carefully check if it was true for their data before using this method.  However, the method of using flirt/epi_reg will work even when these conditions are not met and so I would recommend that as the more robust/reliable method in general.

All the best,

Mark

On 5 May 2014, at 13:58, Matt Glasser <[log in to unmask]> wrote:

Hi Mark,

In my implementation, the distortion correction is done before flirt BBR.  The reason is that the gradient echo SBRef and spin echo image of matching phase are very easy to register to each other rigidly (because they have the same resolution, ,FOV, and distortions).  The topup field is output as a FSL warpfield and used the do the correction.  

Peace,

Matt.

From: Mark Jenkinson <[log in to unmask]>
Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
Date: Monday, May 5, 2014 at 5:46 AM
To: <[log in to unmask]>
Subject: Re: [FSL] flirt -bbr -fieldmap compared to flirt -bbr / freesurfer bbregister

Hi,

Just to clarify here - for fMRI scans, topup was used to create a fieldmap which was then used in BBR to do the distortion correction, as opposed to using topup itself to do the distortion correction (as the latter is only possible for diffusion images, not for functional ones).  

So, in answer to your question about the order of steps, we do not register the fieldmap to the functional image directly.  We used epi_reg (in FSL) to register fieldmap to structural and then functional to structural, with the fieldmap correction.  This step did the distortion correction.  Later on a second BBR step was run with freesurfer, using the distortion corrected images as input, that refined the rigid-body registration between the undistorted functional image and the structural image.

All the best,

Mark

On 2 May 2014, at 14:41, Matt Glasser <[log in to unmask]> wrote:

HCP data were corrected with spin echo field maps (phase reversed spin echo images) using topup before running FLIRT BBR and FreeSurfer BBR.  

Peace,

Matt.

From: julia <[log in to unmask]>
Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
Date: Friday, May 2, 2014 at 4:24 AM
To: <[log in to unmask]>
Subject: Re: [FSL] flirt -bbr -fieldmap compared to flirt -bbr / freesurfer bbregister

Dear Mark, and also Matt before,

thank you so much for your detailed explanations which are extremely valuable to me.

As a last tiny follow-up: For the HCP data, did you register the fieldmap to the functional data, unwarped and then registered to anatomy? Or did you rather pre-register to anatomy (w/o bbr option?) both the fieldmap and the functional image, unwarped and then finetuned with bbr?

Thanks a lot for being so helpful,
Julia

On 05/02/2014 10:14 AM, Mark Jenkinson wrote:

Hi,

thank you so much for your comprehensive answer and sorry for not being clear enough on my question about the "magic". I am interested in the additional benefit of using the fieldmap within flirt bbr. Or: what is the difference between the workflow in epi_reg and an alternative workflow where everything is equal apart from the fact that the fieldmap (registered to anatomy) is applied to unwarp the EPI (pre-registered to anatomy with normal flirt like it is done in epi_reg) before running flirt bbr (without fieldmap) instead of simultaneous unwarping and bbr registration (flirt -bbr -fieldmap). Assuming the transformations would be combined in the end to avoid multiple interpolations.

That is a tricky question, although the short answer is probably "not much".  The use of bbr in the first step will already register to the structural image, so you are never avoiding that, but running things in the second stage, with an unwarped functional image as input can change things subtly.

The long answer is that the main difference will be that the second stage of the non-simultaneous version will need to use an interpolated image as the input image to work with.  When we investigated this for the HCP, we ended up using such a strategy with the non-fieldmap bbr step being run by freesurfer, and hence fed with an unwarped (and interpolated) image.  When I fed the same image into flirt (as a two-stage process, so replacing the freesurfer bbr with flirt) I found that there was very little difference between the freesurfer and two-stage flirt results (less than comparing to simultaneous flirt).  However, because we wanted to use the freesurfer segmentations for the best GM-WM edges, then we just stuck with freesurfer for the second bbr step, although I found that with the same input (surface and unwarped image) there was almost no difference between flirt and freesurfer bbr.  The difference between using flirt with the two stages rather than simultaneous is probably about slight smoothing benefits from the unwarping step, which on HCP data matter more than I suspect they do in most other data.  You could try this yourself and see.  It is likely to be hard to see the differences, and the exact benefit/detriment will probably be highly dependent on the quality of the structural, functional and fieldmap images.

Concerning the masking I am a bit confused now. Within epi_reg, if I see it right, the mask that is used in the first fugue step is a multiplication of the binarised magnitude brain image and the binarised fieldmap:

    $FSLDIR/bin/fslmaths ${fmapmagbrain} -abs -bin ${vout}_fieldmaprads_mask
    $FSLDIR/bin/fslmaths ${fmaprads} -abs -bin -mul ${vout}_fieldmaprads_mask ${vout}_fieldmaprads_mask
    $FSLDIR/bin/fugue --loadfmap=${fmaprads} --mask=${vout}_fieldmaprads_mask --unmaskfmap --savefmap=${vout}_fieldmaprads_unmasked --unwarpdir=${fdir}

This would always result in a mask which is as small as the fieldmap itself or smaller, right? Where does (or should) the extrapolation happen then? Maybe I don't understand the function of the mask in fugue or am misinterpreting the script?

The masking here is done to make sure that only valid fieldmap values are kept within the brain extracted mask.  It is usually quite tight and can remove a bit of brain tissue, which is fine as it is better to do that in this context rather than keep in any of the noisy voxels that sit at the brain edge in the fieldmap.  So we extrapolate the fieldmap (rad/s) image using fugue (as this always extrapolates internally), and this is explicitly specified via the --unmaskfmap flag (as otherwise the fieldmap would be masked on output).  It is important to do it this way to remove these noisy voxels at the brain edge, and because the fieldmap is pretty smooth, extrapolation by a few voxels near the edge is a fairly safe/accurate thing to do.

I hope this helps.

All the best,

Mark

Many thanks in advance,
Julia

On 04/23/2014 08:41 AM, Mark Jenkinson wrote:

Dear Julia,

I'm not quite sure what this "magic" is that you and Matt are referring to.

The process inside the epi_reg script is to register the fieldmap (which is undistorted) to the structural image and then use the fieldmap in structural space when doing the distortion correction and rigid-body registration of the EPI to the structural simultaneously.  If you are just using the flirt command line then you will need to get the fieldmap into structural space yourself.  However, if you've already applied a fieldmap distortion-correction step prior to flirt then you should not use the fieldmap again (as this would overcorrect) and can just use the bbr option with the 6 dof setting.  I wasn't very clear as to your pipeline, so I hope this helps explain what is appropriate for you.

As for the masking issue, it is necessary to have a tight mask into the fsl_prepare_fieldmap step in order to exclude noisy voxels that occur at the edge of the brain.  However, when applying the fieldmap it should not be masked in this way and does need to be extrapolated.  The "too small" hack in the script is not for this extrapolation (as this extrapolation is easily done with fugue) but is instead there to deal with cases where the FOV was cut-off, typically in the through-slice direction, by a much larger factor that goes beyond what the fugue extrapolation was good at dealing with.

Finally, to answer Matt's question, I have not added GIFTI input support to FLIRT at this stage.

All the best,

Mark

On 18 Apr 2014, at 17:25, julia <[log in to unmask]> wrote:

Thanks again!
I would be very interested in the "magic" in flirt bbr when used with a field map.

If I may I'd have one more question (possibly related to the magic):
 
It is advised to use a rather small mask for fsl_prepare_fieldmap. However, when I apply such a fieldmap directly in fugue (i.e. after unmasking), smaller maps tend to smear the signal in areas that are not covered by the fieldmap.
Is this dealt with by the "new hack for extrapolation if fieldmap is too small" part in the epi_reg script? What confuses me is that there, the dilated fieldmap version is only used inside flirt bbr, not for the later warping (if I read it right).

Best regards,
Julia


 

On 04/16/2014 05:49 PM, Matt Glasser wrote:

1. FLIRT BBR is a better and more robust initialization than the other FreeSurfer initialization methods (so I actually turn off bbregister’s initialization).
2. I don’t think it is possible to give flirt bbr a GIFTI surface, but perhaps Mark has made this change by now?
3. I’m not sure I follow the goal of the process here, but perhaps Mark does?  Mark will have to explain what happens inside flirt bbr when used with a field map, but I think there is some “magic” in it.  

topup and regular field map corrections appeared very similar to us, but topup’s can be acquired faster and has the advantage I mentioned below in terms of robustness.  

Peace,

Matt.

From: Julia Huntenburg <[log in to unmask]>
Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
Date: Wednesday, April 16, 2014 at 10:36 AM
To: <[log in to unmask]>
Subject: Re: [FSL] flirt -bbr -fieldmap compared to flirt -bbr / freesurfer bbregister

Thanks a lot, Matt, that is all extremely helpful.

Just three quick follow-ups:

1. If freesurfer bbr is more accurate, what would be the advantage of using flirt bbr first instead using freesurfer bbr straight away?

2. If the segmentation is the main difference, would it be reasonable to assume flirt bbr would perform similarly well when fed with a freesurfer segmentation?

3. In case I register both EPI and fieldmap to the anatomy before, then apply the fieldmap and then do another round of registration (unwarped EPI to anat) with bbregister, that would not be different (at least no disadvantage) to performing this last step with flirt -bbr -fieldmap (as in epi_reg)? I.e. there is nothing magical happening inside flirt -bbr when it is provided with a fieldmap?

Also the hints on topup correction are very valubale for me as I am currently implementing both methods for distortion correction for my data to compare them.

Many thanks again, great help!

Julia

2014-04-16 17:17 GMT+02:00 Matt Glasser <[log in to unmask]>:

FreeSurfer BBR is generally slightly more accurate than FLIRT BBR, probably because of the much higher quality surface being used (i.e. the FLIRT BBR makes a quick white matter surface based on a FAST segmentation, whereas FreeSurfer’s white matter surface is based on a lot more processing).  For HCP Minimal Preprocessing Pipelines where we cared about getting the alignment right down to submm levels, we used both (FLIRT BBR first and then FreeSurfer BBR).

Whether or not it is better to use the field map with FLIRT BBR or apply it before depends on how good of a field map to EPI registration you can get outside of FLIRT BBR vs a field map to structural registration (and EPI to structural registration) inside FLIRT BBR (i.e. epi_reg).  Generally one does a better job inside of FLIRT BBR, but this is not always the case.  Also, it can be harder to debug what is going wrong when the steps are not separated (is my registration not good because the distortion correction is not good, or because the EPI to structural registration is not good?).  

I actually prefer using spin echo field maps (phase reversed spin echo images matched to the gradient echo EPI acquisition) and topup, as these have the same distortion as the gradient echo EPIs and can be very precisely registered (EPI to matching spin echo image) without the complication of using the structural image as an intermediate.  Then one can use BBR on the undistorted EPI image.  This keeps each step separate for debugging and “easy” for the algorithms involved to achieve robustly.  

Peace,

Matt.

From: Julia Huntenburg <[log in to unmask]>
Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
Date: Wednesday, April 16, 2014 at 9:25 AM
To: <[log in to unmask]>
Subject: [FSL] flirt -bbr -fieldmap compared to flirt -bbr / freesurfer bbregister

Dear list,

Concerning the usage of flirt with bbr cost function and a fieldmap I was wondering if there is any experience on:

1. how different it is to use the flirt -fieldmap option as compared to unwarping with the fieldmap first and using flirt -bbr afterwards?
(-- and related: How exactly is the fieldmap applied in this implementation?)

2. how the FSL bbr implementation compares to freesurfer bbregister? E.g. if I use the freesurfer segmentation with flirt -bbr, would that give me similar results as bbregister?

Any thoughts or experience (even if anecdotal) would be highly appreciated!

Many thanks,

Julia

<smear1.png><smear2.png>