Open tsalo opened 2 weeks ago
Not sure if this was prompted by the fact that I'm currently at a Hackathon working on dwidenoise
improvements: #3029.
I also recently generated a meta-Issue for listing the various improvements that would be useful in this context: #3023.
I am interested in improving the efficacy of dwidenoise
for both DWI and fMRI data. @BahmanTahayori has a pending manuscript on the benefits of using MP-PCA for multi-echo fMRI data analysis. We will be advocating that this be done upstream of fMRIPrep, combined with particular options to provide to TEDANA after fMRIPrep. It would however be more elegant if PCA denoising, whether MP-PCA or something else, were to be executed by fMRIPrep itself, and I'd be quite happy if I could progress dwidenoise
to the point of being that choice.
The ability to have dwidenoise
take as input a pre-estimated noise map, and use that to threshold the eigenvalue spectrum rather than running the MP estimator, would both bring the usage more in line with NORDIC, and work nicely with fMRIPrep's new fit-apply model, as the final estimated nose level could be written as an intermediate derivative. This is listed as #2274 and I currently have a team member working on it.
I'm not intimately familiar with the exact distinctions between NORDIC and complex MP-PCA. I believe there's something about g-factor penalty, and I don't know if that's possible to estimate without explicit calibration data. We could at least have access to the slice timing vector to know which slices were members of the same group, but I suspect that's not adequate.
Regarding how to provide that input noise information, there's multiple sources of confusion:
-noise
option is an output, not an input; but as described above, there will soon be an alternative option that takes a noise map as inputNot sure if this was prompted by the fact that I'm currently at a Hackathon working on dwidenoise improvements
That was more of a happy coincidence, but I did look through your recent dwidenoise issues to see if there was any overlap. I didn't think there was, but I completely missed #2274.
The existing -noise option is an output, not an input; but as described above, there will soon be an alternative option that takes a noise map as input
Ah, sorry. I didn't realize that.
A noise map would be 3D rather than 4D
If possible, it would be great to be able to provide the no-excitation volumes (organized in BIDS as files with the _noRF
suffix) directly, rather than having to calculate a 3D noise map ahead of time, but I understand if you consider that out of scope. The protocol I've been seeing (and using) lately includes 3 no-excitation volumes at the end of the functional scan, so it's definitely a 4D file initially.
I'm not sure how a noise map would be complex. Is this claiming different noise levels between real and imaginary components? I would not expect to see such a thing from a quadrature receiver array. It seems from looking at the code, the eigenvalues are complex but the MP distribution is fit to their magnitudes.
I don't actually know either. NORDIC accepts complex-valued BOLD data with no-excitation volumes at the end, so they definitely have a phase component, but NORDIC could be discarding the phase component of the noise volumes within the workflow. Even if that is the case, it would be nice to be able to provide the complex-valued noise volumes instead of having to organize the BOLD data and noise volumes differently, but I understand if that's out of scope.
On the other hand, I don't know how to convert the phase part of the noise volumes to radians, which might make it hard to create the complex-valued version of the noise volumes. NORDIC just accepts the magnitude and phase data separately, without any scaling.
If possible, it would be great to be able to provide the no-excitation volumes (organized in BIDS as files with the _noRF suffix) directly, rather than having to calculate a 3D noise map ahead of time, but I understand if you consider that out of scope.
I don't think it's "out of scope", I would rather advocate for handling it differently.
If there's the capability to provide a noise level estimate as input, that noise level can be estimated through any means. My main goal in doing so is to estimate the noise level based on the MP distribution, filter that estimate (eg. smoothing, outlier detection & replacement), and then apply that noise level estimate for denoised series reconstruction. However if one has acquired data with no RF, you can do similar by explicitly estimating a noise level map from those data.
So while a higher-level workflow for data denoising may need to branch depending on what data are available / how one chooses to filter the noise map, and that may include splitting a 4D series based on the presence / absence of RF excitation, I would personally keep the dwidenoise
interface very simple, just taking a noise map as input.
Is there a publicly visible dataset that includes the no-RF volumes at the end of a 4D series? I'm curious to know how this is encoded in the DICOMs and whether our software needs to do something smarter with it. I've also never seen a pulse sequence with this capability; do you know if this is GE only?
I'm not sure how a noise map would be complex.
NORDIC accepts complex-valued BOLD data with no-excitation volumes at the end, so they definitely have a phase component, but NORDIC could be discarding the phase component of the noise volumes within the workflow.
I think we're describing different things. A noise level map I expect to be real-valued. What I think you're referring to is the volumes to be used for noise mapping. Those can be (and ideally should be) complex just as can / should be the input data to be denoised; but if an explicit calculation is performed using those data as input to estimate the noise level, I expect that to yield a real-valued result.
I don't know how to convert the phase part of the noise volumes to radians, which might make it hard to create the complex-valued version of the noise volumes.
This is a common existing step for denoising of complex DWI data. Some tools will look at the minimum and maximum values in a phase image and linearly transform the intensities so that these lie in for instance the [-pi, pi] range, though there can be some small error in doing it this way if one is not careful. For Siemens data phase images are known to be encoded in the [-4096, 4094] range. dwidenoise
expects data of type "complex float"; mrcalc
has a -polar
option that converts separate magnitude and phase images into the real & imaginary components of a native complex floating-point type.
If there's the capability to provide a noise level estimate as input, that noise level can be estimated through any means.
I see, so you'd prefer to have a separate step to create the noise map from the no-RF volumes. Would you be willing to include that as a separate command in mrtrix3 or would you prefer that be handled by a different package?
Is there a publicly visible dataset that includes the no-RF volumes at the end of a 4D series?
I have a multi-echo BOLD dataset on OpenNeuro with no-RF volumes included. I know of a few other folks who are acquiring similar data (e.g., Damien Fair's group at UMinn, a number of folks at WUSTL, Cesar Caballero-Gaudes, and Dan Handwerker), but I don't think anyone else has had a chance to share their data yet.
I'm not aware of any DWI datasets with no-RF volumes, but I am pretty sure that NORDIC is designed to leverage no-RF volumes in DWI as well.
I'm curious to know how this is encoded in the DICOMs and whether our software needs to do something smarter with it. I've also never seen a pulse sequence with this capability; do you know if this is GE only?
As far as I know, this is only (or at least most visibly) implemented in the Siemens CMRR-MB sequence. I've chatted with a number of other folks and have also looked at the DICOMs myself, and the consensus seems to be that there's no info in the DICOM that can be used to distinguish the regular volumes from the no-RF volumes. I typically include a post-heudiconv curation script to split out the no-RF volumes into files with the _noRF
suffix. Unfortunately, I can't share our DICOMs (sorry about that).
Perhaps the CMRR folks would be willing to modify the sequence to encode that info in the DICOMs though.
A noise level map I expect to be real-valued. What I think you're referring to is the volumes to be used for noise mapping.
Ah, that makes perfect sense.
This is a common existing step for denoising of complex DWI data.
I was concerned about rescaling the signal data and no-RF data separately in case that might result in different scaling, but I could probably concatenate them back together for the rescaling step.
Would you be willing to include that as a separate command in mrtrix3 or would you prefer that be handled by a different package?
It will be possible to estimate via the existing MP-PCA method. I am considering the prospect of having a dwi2noise
command that does almost everything dwidenoise
does, it just has as its compulsory output argument the estimated noise level rather than a denoised 4D series. That way one could run that command, do any kind of filtering of the resulting noise map they deem suitable, and then run dwidenoise
with an input noise map.
In the case of the no-RF data, it would require a separate implementation to estimate the noise map as written in the NORDIC manuscript. It's not the most complex thing in the world, but it'd need to be written regardless. The g-factor penalty may need to be computed differently given that sensitivity profiles would not be available. So as a first pass I'd probably be leaning on the MP-PCA estimator, and use of the no-RF data via the method described in the NORDIC manuscript would be a future augmentation.
I was concerned about rescaling the signal data and no-RF data separately in case that might result in different scaling, but I could probably concatenate them back together for the rescaling step.
If you know a priori how the phase images are numerically encoded in DICOM, then the multiplier to get to radians is fixed and calculable a priori also. There is a small risk with the data-driven approach if one series contains both a voxel with value -4096 and a voxel with value +4094 whereas the other does not; but I wouldn't conceptualise that as an error of consistency so much as that for at least one of the two, the conversion is wrong. I've a suspicion that the information about how phase is encoded may be buried somewhere within private fields, in which case conversion to radians could occur at the import stage (#3009).
I've opened a PR to fMRIPrep to add a denoising step with dwidenoise: https://github.com/nipreps/fmriprep/pull/3395. There are a few blockers, but I'd love your input on the overall structure.
Thanks for the ping; see response at https://github.com/nipreps/fmriprep/pull/3395.
Is your feature request related to a problem? Please describe. I would really like to include a NORDIC denoising step in fMRIPrep, QSIPrep, and ASLPrep, but the official implementation of NORDIC is in MATLAB and is pretty difficult to use. Would you be willing to incorporate the NORDIC algorithm in dwidenoise?
Describe the solution you'd like A new
-algorithm
or-method
parameter for dwidenoise that accepts a value of eithermppca
ornordic
would be amazing.Describe alternatives you've considered Could dwidenoise accept a 4D noise image (either magnitude or complex) instead of a float for
-noise
? This would be necessary for NORDIC, but I don't know if MP-PCA could leverage the same information.