-
Notifications
You must be signed in to change notification settings - Fork 59
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[User] qmt-spgr number of pre-pulses #292
Comments
New question:
There might be many reasons why you obtain such small values for F. I would say that field maps errors and poor spoiling are the most probable. Their effects on the parameter F could be simulated. @mathieuboudreau is certainly more used to qmt spgr dataset than I am and could help you find the issue in your dataset. @mathieuboudreau What are your thoughts??
The preprocessing is not included in qMRLab. However, you can run the following command in your Matlab Command Window to create the file B1 = 'B1.nii.gz';
save_nii_v2(smooth3(load_nii_data(B1)),'B1_smooth.nii.gz',B1);
Sorry the MP2rage module is not ready yet... see #255 (still open issue). Best, |
Thanks @tanguyduval ! A few quick thoughts: 1 - To my knowledge, fixing R1f*T2f = constant has only been done in studies using the Yarnykh model, I'm not sure how it will behave with other ones. Last notes, MP2RAGE is going to get working on in the coming months, and a B1 filtering module is currently being worked on by @ileppe in #135 |
I can't see what their R1obs value is from their second image, but assuming that it's very close to R1f, it might be a bit low (1.28 s vs ~1.5 for Sinclair et al 2010), but I don't know how fixing R1f*T2f and/or using the Ramani model impacts the sensitivity of F to R1obs (for me, not fixing it and using Sled, underestimating T1 resulted in overestimating F, which is the opposite that the authors might be experiencing). |
I agree with @mathieuboudreau's points. The ratio/product of R1f/T2f that Yarnykh has used was derived for brain, I see no reason for it to be the same in muscle. B0 maps should have only a small impact on F, if it was having an effect you would see variations in F that resembled the B0 map. However I have an additional theory that MP2RAGE is likely not appropriate to use to measure R1obs for qMT. I have not seen a comparison of MP2RAGE and VFA, but every T1 method disagrees so I strongly suspect that MP2RAGE disagrees with VFA. I have also not seen the MT effect of MP2RAGE investigated or quantified. Hence I'd recommend sticking with VFA, with similar parameters to the qMT experiment itself, to get the R1obs measurement. |
Thanks @usinha2 ! For everyone's convenience, here is what the document says: Thanks for all your comments. I just registered on GitHub and saw these comments. To give a little background on our muscle qMT work:
|
Just to clarify, my paper only investigated the Sled and Pike model, and didn't restrict any of the parameters (besides R1r). So, changing some of these options could potentially lead to different results, so take my conclusions with a grain of salt. I saw paper more as a framework on how to investigate the B1 sensitivity for a particular fitting model, and the results I concluded can only be applied to my settings unless as of now. |
Another note, for my paper, I observed that this was the case if you are only interested in F - the noise or error is mostly propagated to other parameters (e.g. kf, T1f). If these parameters might be of interest to you (I think kf may be for MSK studies), then I would certainly recommend that you acquire a B1+T1 map, just be aware that different setup (e.g. B1-dependent or B1-indepent T1 technique) may propagate errors differently to different parameters. At this stage, I don't think you should be too concerned with this, as I think you have a much larger parameter bias to correct (so odd that your VFA T1 maps were underestimating by 50%! I've heard of severe overestimations on GE scanners, but not under and not on Siemens) |
Again, at this stage, I would recommend that you do B1 correction with VFA. I don't think this is the major source of your bias at the moment, you can explore this type of hyper error-propagation optimisation later on, once you've bridged the larger gap in your parameter estimates. Sorry that my conclusions for that paper might have gotten you slightly off-track! When I first mentioned it to Tanguy to be forwarded to you, I intended to mean that generating my Figure 2 from that paper with simulations using your setting could give you some additional insight on if T1 errors could be a major contributor in the underestimation you were observing in "F". |
@usinha2 What range of values did you get for F when you switched to MP2RAGE? Did you F increase closer to the 8% reported, or decrease further? |
Hi Mathieu
Thanks for all your suggestions. I will get back on the T1 and f with
VFA/MP2RAGE T1 mapping soon after I look at all our data and analysis.
I see your point that the discrepancies we are getting for the f values are
large and that has to be solved before going onto exploring what T1 mapping
method is better.
I was wondering if the 3D MT SPGR sequence needs to have spoiler gradients
between the MT pulse and the read Rf pulse?
Tanguy had indicated that inadequate spoiling maybe a problem: we are
using the default Siemens RF phase cycling (of read RF pulse) and spoiler
gradients after readout; should we include spoiler gradients after the MT
pulse as well?
One other point: when we normalize the MT signal intensity, we are using
the signal intensity of the largest offset image (100 Khz) at the given
flip angle of the MT pulse : should we normalize all to an image without
a MT pulse
Usha
…On Thu, Jan 10, 2019 at 1:11 PM Mathieu Boudreau ***@***.***> wrote:
We tried several combinations of TR and several flip angles for VFA, but
the T1 values continued to be underestimated. Our values of ‘f’ were small
at the underestimated T1 values from VFA (2-3% compared to the 8% reported).
Then we switched to MP2RAGE as it gave us T1 values closer to that
reported for muscle.
@usinha2 <https://github.com/usinha2> What range of values did you get
for F when you switched to MP2RAGE? Did you F increase closer to the 8%
reported, or decrease further?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#292 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AsTb6SmothHcPQPWtfZjjnR9IoGyAdhwks5vB6yPgaJpZM4ZPKRc>
.
--
*******************************************************************************
Usha Sinha, Ph.D.
Director, Medical Physics
Professor and Chair, Physics
San Diego State University
San Diego, CA 92182
|
Yes. For small MT offset frequencies, your MT pulse can start to excite the water pool (from the tails/sidebands of the MT pulse). This can leave some water magnetization in the transverse plane before your imaging pulse, leading to image artefacts. Spoiling this transverse magnetization prevents the artefacts. The longitudinal water magnetization will still be reduced, but the Sled & Pike model accounts for this. The Ramani/Yarnykh models do not, which is why they have a limit to the lowest offset frequency that can be used (1 kHz and 3 kHz respectively IIRC). If your lowest offset frequency is above 1 kHz, then artefacts should be minimal depending on the sidebands of your pulse. However, I would still recommend having the spoiler in the sequence.
What you absolutely must avoid is introducing different scaling between the different flip-angles. Hence the way I would do it is to ensure that all receive gains etc. are the same for the two flip-angles, and then normalize to one image only. Your method ought to work, because the 100 kHz offset images should be essentially identical between the different MT flip-angles. This is worth checking. Good luck! |
@usinha2 I agree with everything that @spinicist said. Acquiring an image with no MT RF pulse is typically recommended; I believe this is done in most qMT studies that I've read. If you do this, please note that I would recommend you still have the spoiler gradient that as you suspected should be used and that @spinicist recommended (between the MT pulse and the readout excitation pulse). Now this is an anectodal tip and don't have a reference for it, but the idea is that if the gradient impacts the image in any way (e.g. through eddy currents), then this the "imaging environment" will be the same for the MT on/off iamges if you keep the gradient on (which should be desirable, because the MT off image is supposed to be used for normalisation). If you're really tight for time, and want to omit acquiring an MT-off, then I second that it may be best if you only use one of the high offset images to normalize all your data; I think the lowest flip angle image should be the best choice, to avoid any on-resonance effects as Toby mentioned. But feel free to compare both high-offsets to verify if they have any bias between them for all the tissue parameters in your region of interest. |
Thanks, Mathieu and Spinicist: will update after implementing your suggestions. So is it 1 kHz the minimum lower offset frequency for the Ramani model and 3 kHz for the Yarnykh model? What is IIRC? |
IIRC = if I remember correctly. For the limits, those sounds about right. See the discussion in Yarnykh's 2002 paper for that number; I'm not sure if it might be field dependent though. See below for a screenshot of the relevant section. |
Thanks--will look at Yarnykh's paper as well
…On Fri, Jan 11, 2019 at 7:41 AM Mathieu Boudreau ***@***.***> wrote:
IIRC = if I remember correctly.
For the limits, those sounds about right. See the discussion in Yarnykh's
2002 paper <https://onlinelibrary.wiley.com/doi/full/10.1002/mrm.10120>
for that number; I'm not sure if it might be field dependent though. See
below for a screenshot of the relevant section.
[image: capture d ecran 2019-01-11 a 12 10 41]
<https://user-images.githubusercontent.com/1421029/51043727-f9677f80-1599-11e9-9494-438a1b726729.png>
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#292 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AsTb6RUSYjEObHxgJWCR17psF3b9WKtNks5vCLCpgaJpZM4ZPKRc>
.
--
*******************************************************************************
Usha Sinha, Ph.D.
Director, Medical Physics
Professor and Chair, Physics
San Diego State University
San Diego, CA 92182
|
Hi Mathieu
One point that has been bothering me:
When I acquire the images (Siemens PRISMA 3T) for qMT, I have no way of
skipping the tuning between scans (though Siemens support tells me that the
receive gain does not typically change between these acquisitions;
strangely the actual value of receive setting is not given but it is only
set to Low or High gain but we also checked the image header scaling factor
that did not change between scans). initially I had planned to tune the
protocol for the largest offset frequency (having the highest signal
intensity) and using that setting of receive gain for all the other qMT
scans: I did play around with the manual tune and I believe I can open
that window for each scan (after the first tuning for the highest offset MT
pulse) and prevent an auto-tuning: is it how you recommend doing it.
Thanks
Usha
…On Fri, Jan 11, 2019 at 7:41 AM Mathieu Boudreau ***@***.***> wrote:
IIRC = if I remember correctly.
For the limits, those sounds about right. See the discussion in Yarnykh's
2002 paper <https://onlinelibrary.wiley.com/doi/full/10.1002/mrm.10120>
for that number; I'm not sure if it might be field dependent though. See
below for a screenshot of the relevant section.
[image: capture d ecran 2019-01-11 a 12 10 41]
<https://user-images.githubusercontent.com/1421029/51043727-f9677f80-1599-11e9-9494-438a1b726729.png>
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#292 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AsTb6RUSYjEObHxgJWCR17psF3b9WKtNks5vCLCpgaJpZM4ZPKRc>
.
--
*******************************************************************************
Usha Sinha, Ph.D.
Director, Medical Physics
Professor and Chair, Physics
San Diego State University
San Diego, CA 92182
|
Hi Usha, That is a troubleshooting problem that I don't know how to fix. When I was acquiring qMT data during my PhD, we used a custom sequence and each qMT protocol parameter were different "measurements" within the same sequence, so no tuning was being done between images (to the best of my knowledge). I was working on a Trio, so I don't know what challenges you might be facing on a Prisma or how to tackle that. Maybe @ileppe encountered something like this and could chime in? |
re: adjustments. I have asked Siemens this question as well. It seems you can't easily 'copy' the receiver gain but that it would be very obvious in the image if it had changed from one acquisition to the next (i.e. very quantized). To avoid re-shimming, we were told to 'Copy Adjustment Volume' from the previous sequence, this should force at least the shim parameters to be the same (we do this when we manually shim). I hope that helps. |
Hi ileppe |
@usinha2 I have an example for B1 (acquired using the Siemens values are around 1.4-1.5s, in decent agreement with those reported in the literature. For VFA, I used two FLASH volumes with 5 and 25deg flip angles. 2.86ms TE and 35ms TR. As @ileppe suggested, I divided the B1 map given by How off your values are from what you expect? |
@agahkarakuzu did you save the protocol file for that VFA acquisition? Maybe you could contact @usinha2 by email and exchange some Siemens-specific details, and see where both your protocols might differ to help resolve their very low T1 values. |
They said in a comment above.
|
Sorry @usinha2 I did not notice that you've already mentioned the T1 values you get, thanks @mathieuboudreau for bringing that to my attention. I will share the protocol and sequence params as soon as I have access to my workstation. |
@usinha2 I am sharing the protocol and my sample data with you: The following contains 3 DICOM images of fa5, fa25 and b1, so that you can see all the sequence params: |
Hi Agah
Thanks so much. Your calf muscle T1 values looks pretty much what is
reported in the literature.
We also use the 3D FLASH sequence on the Siemens (FLASH is the equivalent
of SPGR --acronym on GE scanners): with TE=3 ms, TR= 25 ms, FA= 5, 15, 25
degrees.
The B1 mapping sequence we are using is I believe the same sequence as you
refer to since we also get a map scaled by 10 (as the nominal 80 degree
saturation pulse is scaled by 10), then we divide by 800 to get the
relative map and then smoothed : we acquired with the same in plane
resolution as the T1 map (what was your in-plane resolution for the B1
map?). Now I am thinking that maybe, it is best to keep it low resolution
as this is a tuboFLASH acquisition even though it is centric reordered.
I have your data and am going to look at it and process it: will also set
the protocol on our scanner to match yours. Our T1 values after correction
were as low as 800 ms.
Thanks, will get back to you soon
Usha
…On Mon, Jan 14, 2019 at 6:17 AM Agah ***@***.***> wrote:
@usinha2 <https://github.com/usinha2> I have an example for B1 (acquired
using the Siemens B1_map_for_T1_mapping) corrected VFA T1 map of the
lower leg at 3T here:
[image: screenshot from 2019-01-14 09-00-19]
<https://user-images.githubusercontent.com/9632322/51117238-2e164980-17db-11e9-8949-6b6a18d5328a.png>
values are around 1.4-1.5, in decent agreement with those reported in the
literature. For VFA, I used two SPGR volumes with 5 and 25deg flip angles.
2.86ms TE and 35ms TR.
As @ileppe <https://github.com/ileppe> suggested, I divided the B1 map
given by B1_map_for_T1_mapping sequence by 800, upsampled to the T1
domain, then smoothed.
How off your values are from what you expect?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#292 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AsTb6U5d7ghSTUKrYAmLJGlRKfPdnMrRks5vDJF_gaJpZM4ZPKRc>
.
--
*******************************************************************************
Usha Sinha, Ph.D.
Director, Medical Physics
Professor and Chair, Physics
San Diego State University
San Diego, CA 92182
|
Hi All 3DFLASH | Fat Suppression | T1 (Seconds) It is not clear why fatsat should decrease T1s-- it may saturate water signal but this should be the same for all the flip angles used in the VFA maps-- it will reduce SNR but not clear why T1 is underestimated. |
Hello, The problem here is that a FatSat pulse is also an MT saturation pulse - both are off-resonance saturation pulses. The only difference is the specific offset frequency used - but the MT resonances will overlap with fat. Hence by adding FatSat you are adding extra saturation of the MT pool, which will change your apparent T1. If you were on a GE scanner I would suggest using Spectral-Spatial pulses, which avoid exciting the fat resonance at the cost of a slightly increased imaging TR. I've used this trick to avoid exciting scalp fat when doing CEST in the brain. However, I do not think these are available on Siemens? EDIT: Hang on, I guess "Water Exc." might be the Siemens name for Spectral-Spatial? In which case I would go with that option. |
@usinha2 First off, I'm happy you were able to reproduce Agah's VFA T1 mapping sequence/results! 🎉 I second what @spinicist said; adding fat sat pulses to your VFA and MT measurements will add an additional MT-weighting to your signal, which will ultimately impact your parameter quantification efforts. I've never tried avoiding fat in my own measurements, so I would try @spinicist's suggestions if you are concerned with fat in your image. I don't recall this being discussed a lot in brain qMT research, but I'd have to check again to be sure. Maybe you could contact the authors from the Sinclair 2010 paper and ask if this was a consideration for them; I know that the senior author Prof. Xavier Golay is very active on Twitter.
This would need to be checked, but my first instinct if that no it won't be the same for all VFA maps. The MT signal equations are non-linear, and the signal saturation from MT is not simply "added on" to the steady-state SPGR signal. The MT pulse (i.e. in your case, fat sat) and excitation pulse do affect the steady-state signal independently. It is likely that if anything, as @spinicist hinted, it would change the apparent T1, meaning that the entire SPGR might be shifted for a constant MT pulse but varying excitation flip angles (i.e. shifted relative to the no MT saturation case). This could be checked with simulations; if I find some time this week or the next, I might explore that in a Jupyter Notebook and share it. |
Also, this recent open-access paper on how MT impacts VFA measurements might contain some relevant information, if you want to inform yourself a bit further: https://onlinelibrary.wiley.com/doi/full/10.1002/mrm.27442 . Note that this isn't completely what is happening to you though. |
@mathieuboudreau I was going to recommend that exact paper. Rui works at the other KCL campus 👍 |
PS - As to why FatSat isn't discussed in the brain qMT literature, that's because the brain itself does not contain much "fat". Yes, it contains lipids in myelin, but fat as a separate tissue is not found in the parenchyma (to my knowledge). There is the layer of scalp fat, which most of the time is not an issue but was causing recon artefacts for a collaborator of mine. |
Hi all |
Hi All |
Hi, (For brain qMT, here we use a 20ms Gaussian pulse that has a bandwidth < 100 Hz, and on a different project with a Hanning pulse I once had very odd off-resonance problems above the sinus due to bad sidelobes on the pulse) |
Hi @usinha2 , I second what @spinicist said, however it looks like you were already aware that this might have been an issue with your sentence quoted below:
Yes, I think pulse truncation can be an issue (I've heard so from experience of others). The sharp drop from the truncation on both sides of the Gaussian will cause some high frequency components to the MT saturation, and I don't know if these are modelled well in the methods. That's why an apodized Gaussian pulse is often reported in papers; it attenuates these components. If you have access to modify your MT pulse shape on your scanner, you could re-write the pulse shape to contain some appodization (see here for a Gaussian-Hanning example; all our pulse shapes are here). It may be possible for you to simulate these and see the difference between fitting with data simulated with Gaussian and Gaussian-Hann for your parameters (e.g. maybe simulate using the Sled and Pike model, then fit with the Ramani which I think you are using?), however nothing will ever beat the real experiment (particularly in MT). If you have MT phantoms at your disposal, these might help for this debugging too (since those parameters are theoretically calculable). Hope some of this info help! |
This issue was moved to https://forum.qmrlab.org/t/user-qmt-spgr-number-of-pre-pulses/34 |
Email received
Yes, that is correct, # of MT pulses is the number of pre-pulses used to achieve steady state. Usually a dummy scan is performed prior to the scanning.
Without dummy scans, you have an unstable image while filling your k-space --> blurring, signal dispersion.
See, for instance, this paper for some illustrations:
I see your point about the center of the kspace being after 1024 pulses. Some clarifications about
the # of MT pulses
:Some simulation with kspace filling are required to investigate the impact of the transient phase. If you do not see any artefacts and signal in the background, and if the fitting looks good (especially the first volume), you might be right that dummy scans are not necessary. But I cannot advise you to do so...
I would recommend to skip the first volume (assuming that all volumes have been acquired in a row). Or, as you said... redo the experiments.
Best,
Tanguy
The text was updated successfully, but these errors were encountered: