How to correct for distortions in spinal cord diffusion MRI data?

MRI Acquisition
1

Context: distortions in EPI

Echo planar imaging (EPI) data usually suffer from susceptibility artifacts, manifesting as pixel displacement in the phase-encoding direction. For gradient echo EPI (used in fMRI), intravoxel dephasing also causes a decrease of the magnitude signal (Note: the information is lost and cannot be recovered with processing techniques). In diffusion MRI, the strong and rapidly-switching gradients also induce Eddy currents, which manifest as translation, scaling and shearing. Note that scaling and shearing displacements increase as a function of the distance from the center of the FOV [Bodammer et al. 2004], which is useful information for what comes later in this post.

What are the traditional methods for the brain?

Popular methods include:

  • Mapping the static B0 field using a fieldmapping sequence and then estimating and correcting pixel displacement [Jezzard et al. 1995].
  • Acquiring EPI data with two opposing polarities of the phase-encode gradients (technique also known as “blip up / blip down”). Both data are submitted to distortions with opposite displacement direction. These two data are then used to estimate a distortion field that brings a distorted EPI to the “mid-space”, effectively correcting susceptibility-related distortions [Andersson et al. 2003]. Such method is notably implemented in FSL’s topup.

Why it doesn’t work (as well) in the spinal cord

It is difficult to acquire a robust and accurate B0 fieldmap in the spinal cord region, partly due to the presence of strong CSF flow, low SNR and dynamic B0 variations in the vicinity of the lungs – up to 70Hz B0 offset were reported at 3T [Verma et al. 2014]. Moreover, the presence of multiple 2𝛑 wraps makes it difficult for unwrapping algorithms to recover an unwrap field map without biases. The danger of using an incorrect fieldmap is that you might end up adding more distortions than already present in the image.

Regarding the blip/up blip/down technique for correcting susceptibility distortions, while this technique usually works well for the brain because in general the entire volume is included in the field of view, in the case of the spinal cord where reduced field of view techniques are used [Samson et al. 2016] (precisely to decrease the amount of distortions at the first place), the two images with opposite gradient polarities end up having different information content, which could then lead to a wrong estimation of the deformation field (the one that finds the “mid-space” between the two images). Beyond the problem related to the reduced field-of-view, a mis-match of information content can also happen because of flow artifacts (producing a strong signal decrease in the hyperintense CSF on T2-like images) and tissue motion (producing signal loss on DW images). When it comes to Eddy-current distortions, the problem is not as bad in the cord vs. in the brain, because the spinal cord is a thin structure (~1 cm diameter) and is usually close to the isocenter (i.e. within the gradient linearity area) and is X-Y centered in the FOV. As mentioned earlier, the scaling/shearing displacements increase with X and Y, so we typically see minimum impact of those affine distortion parameters within the spinal cord region. Eddy-current-related translations could be present however, and those can be easily and robustly corrected using regularized slice-wise alignment [Cohen-Adad et al. 2015]. So, like for the fieldmap approach, a proper estimation of the deformation field is difficult and could lead to spurious deformation.

What are the alternatives?

An alternative method to deal with distortions is based on the non-linear registration of an EPI over a non-distorted reference image [Ardekani et al. 2005]. The reference image is typically a T2-weighted fast spine-echo sequence, that covers the same region as the DW-EPI, and with a similar (or better) spatial resolution. The registration method could be purely image-based (e.g. calculating the mutual information or cross-correlation between the two images), or could use a cord segmentation as pre-registration steps to make the registration more robust & accurate [De Leener et al. 2017]. This method is advantageous because no additional volume acquisition is required, except a reference image without distortion – which is incidentally generally acquired by default in standard protocols.

Of course, it is always a good idea to acquire data with minimal distortions in the first place (by minimizing the echo spacing, optimizing shimming, using reduced FOV techniques, etc.).

OK I get the theory. Now, what can I do with my data?

An example of processing pipeline that involves slice-based correction of DW time series and registration of the mean motion-corrected DW data to a template is available here. It has been tested in an open-access database of ~240 subjects [Cohen-Adad et al. 2019].

1 Like
2

Hi Julien!

This is an excellent forum post that I come back to from time to time.

I have three questions that I think your insight would be invaluable!

  1. With your recent works [Snoussi et al. 2021] and others [Dauleach et al. 2021] since this post, if you had any new insights or recommendations if we did happen to collect blip/up blip/down data? It seems as if there is some improvement in spatial alignment of the images and of tensor-derived indices utilizing these for distortion correction (using HySCO, TOPUP, TORTOISE, etc.). Would you use these, or do you think the improvement is minimal compared to all the challenges that can arise? Even though these algorithms aim to correct intra-scan movement (and do well in the brain), we’ve found that running sct_dmri_moco prior to TOPUP for example seems to improve results…although this is largely anecdotal and we haven’t fully studied this.

  1. Would your recommendations change with anisotropic acquisitions? We acquire data more similar to the spinal cord generic protocol [Cohen-Adad et al., 2021] and I see that the two studies above studied isotropic diffusion MRI data.

  2. Would your preprocessing recommendations change with very high b-value data (i.e. b>4000 where we expect a large majority of directions are noise). And if using the SCT motion correction tools, have you found the need to change any parameters for robust motion correction (larger grouping parameter to ensure signal in at least one volume?)?

Thank you!

3

Hi Kurt! Thank you for keeping the discussion alive. My 2 Canadian cents:

  1. With your recent works [Snoussi et al. 2021] and others [Dauleach et al. 2021] since this post, if you had any new insights or recommendations if we did happen to collect blip/up blip/down data? It seems as if there is some improvement in spatial alignment of the images and of tensor-derived indices utilizing these for distortion correction (using HySCO, TOPUP, TORTOISE, etc.). Would you use these, or do you think the improvement is minimal compared to all the challenges that can arise? Even though these algorithms aim to correct intra-scan movement (and do well in the brain), we’ve found that running sct_dmri_moco prior to TOPUP for example seems to improve results…although this is largely anecdotal and we haven’t fully studied this.

To be honest I haven’t “dirty my hands” for a while with the latest algorithms and software, so I wouldn’t be the best person to answer this. But in general I find that the quality of moco is soooo dependent on the data, that any study comparing algo A vs. algo B would do it on specific dataset(s), and the conclusion might not apply to the dataset you are working with.

  1. Would your recommendations change with anisotropic acquisitions? We acquire data more similar to the spinal cord generic protocol [Cohen-Adad et al., 2021] and I see that the two studies above studied isotropic diffusion MRI data.

I don’t think so, except of course when we talk about interpolating along a large dimension. Eg, for a voxel size of 1x1x5mm, I would avoid interpolating along the last dimension, ie, only work with 2D transformations. In the context of susceptibility distortions, this makes a lot of sense (because distortions are in-plane).

  1. Would your preprocessing recommendations change with very high b-value data (i.e. b>4000 where we expect a large majority of directions are noise). And if using the SCT motion correction tools, have you found the need to change any parameters for robust motion correction (larger grouping parameter to ensure signal in at least one volume?)?

Yes, they might (see first point). And yes, grouping could be a good strategy to mitigate ‘large patient motion’ during the course of an acquisition, although it would not be adequate to address a patient that moves every 5s.

Hope that helps :blush:

4

Thank you for the answers and feedback. Much appreciated!

Fully agree that susceptibility distortion and motion correction (or really any preprocessing in the cord) will be highly dependent on dataset. I suppose for every new study that changes scanner/resolution/qt sampling we will do small experiments using various tools and parameter settings - and use measures similar to the Snoussi et al. 2021 study above to ask ‘what is best for my dataset?’.

Thank you again!

1 Like
5

Hi Julien

Thank you very much for this forum article. It is really helpful, but I have some more questions.
I am currently working with EPI data used for IVIM purposes acquired in spinal cord injury patients. The standard processing pipeline includes fsl topup. However, if I read your article correctly, you suggest that it is not really suitable for SC and that one should use another approach?

My second question regards eddy current correction. So far, we do not correct for eddy currents as FSL eddy probably does not work on our data (we only acquire one diffusion-encoding direction at the time). However, I am wondering if it would be beneficial to apply eddy correction to our data, especially because we acquire close to metal implants. Would you expect that the metal implants make the eddy currents in our data stronger? Would you suggest a slice registration as proposed in your publication of 2015 ([Cohen-Adad et al. 2015]) to correct for the eddy currents? We actually use the sct_register_multimodal when we register our data to the template, but maybe it would also be beneficial to run it only by itself.

Thank you very much for a response!

6

Hi @fres!

It is not that clear cut. The purpose of my article is to essentially bring awareness to the possible issues that can arise when applying “traditional” distortion correction methods to spinal cord DWI data. Depending on many factors (FOV, SNR, resolution, dwell time, parallel imaging factor, fat sat efficiency, etc.), topup could or could not work “as good as it should”. My best advice is to look at your data and make a best judgement, but this step typically requires several years of expertise to be able to make a proper judgement. If you are novice in in the field, I suggest you team up with someone who can make the best call.

Eddy current might be negligible depending on many factors (type of sequence, FOV centering, parallel imaging, etc.) hence might not need correction. I suggest you scan an oil phantom with the exact same imaging parameters, to assess the presence of Eddy current in your imaging setup.

The presence of metallic implant is an interesting consideration. I don’t have enough insights to assess whether they might cause more Eddy currents in your case.

7

Thank you so much for your answer! It’s really helpful!