i-fMRI: Prospective motion correction for fMRI?

An ideal fMRI scanner might have the ability to update some scan parameters on-the-fly, in order to reduce or eliminate the effects of subject motion. Today, this approach is commonly referred to as "prospective motion correction" because the idea is to adapt the acquisition so that (some of) the effects of motion aren't recorded in the data, in contrast to the routinely employed retrospective motion correction schemes, such as an affine registration algorithm applied during post-processing; that is, in between the acquisition and the stats/modeling, which can lead some people to refer to such steps as "pre-processing" if you have a stats/modeling-centric view of the fMRI pipeline.

On the face of it, ameliorating motion effects by not permitting them to be recorded in the time series data is a wonderful idea. Indeed, as the subtitle to this blog attests, I am a huge fan of fixes applied during the acquisition rather than waiting until afterwards to try to post-process away unwanted effects. But this preference assumes that any method actually works, and works robustly, in everyday use. For sure there will be limitations and compromises, yet the central question is whether the benefits outweigh the costs. In the specific case of prospective motion correction, then, does a scheme (a) eliminate the need to use retrospective motion correction, and (b) does it reduce the effects of motion without bizarre failure modes that can't be predicted or circumvented easily?

A good place to begin evaluating prospective motion correction schemes - indeed, all motion correction schemes - is to first asses their vulnerabilities. It's no good if the act of fixing one part of the acquisition introduces an instability elsewhere. Failure modes should be benign. Below, I list the major hurdles for motion correction schemes to overcome, then I consider how elaborate any solutions might need to be. The goal is to decide whether - or when - prospective motion correction can be considered better than the alternative (default) approach of trying to limit all subject motion, and deal with the consequences in post-processing.


What do we mean by motion correction anyway?

As conducted today, motion correction applied during post-processing generally refers to an affine or sometimes a non-linear registration algorithm that seeks to maintain a constant anatomical content in a stack of slices throughout a time series acquisition. Prospective motion correction generally refers to the same goal: conserving the anatomical content over time. But, as is well known, there are concomitant changes in the imaging signal, and perhaps the noise, when a head moves inside the magnet. Other signal changes that are driven by motion may remain in the time series data after "correction." Indeed, depending on the cost function being used, the performance of the motion correction algorithm to maintain constant anatomy over time may be compromised by these concomitant modulations.

Now, we obviously want to try to maintain the anatomical content of a particular voxel constant through time or we have a big problem for analysis! But as a goal we should use a more restrictive definition for an ideal motion correction method: after correction we seek the elimination of all motion-driven signal (and noise) modulations. The only signal changes remaining should be neurally-driven BOLD changes (if we're using BOLD contrast, which I assume in this post) and "physiologic noise" that isn't strongly coupled to head (skull) motion. (Accounting for physiologic noise is usually treated separately. That's the assumption I'll use in this post, although at a very fine spatial scale it's clear that physiologic noise is another form of motion sensitivity.)


Motion sensitivities in fMRI experiments

A useful first task is to consider all the substantial signal changes in a time series acquisition that can be driven by subject motion. What signal changes are concomitant with changes of anatomical content as the brain moves relative to the imaging volume? How complicated is this motion sensitivity? What aspects of the signal changes will require hardware upgrades to the scanner, and/or pulse sequence modifications in order to negate them? And are these capabilities already designed into a modern scanner or will they require substantial re-design? These are the questions to keep in mind as we review the major motion sensitivities.


T1 steady state perturbation

EPI for fMRI is usually acquired with significant T1 weighting because the TR is insufficient to permit complete T1 relaxation for the excitation flip angle used. (We acquire EPI in a time series so a steady state is achieved after three or four TRs, a condition usually established by some dummy EPI scans at the start of the time series.) Any displacement in the through-slice direction will cause perturbations to this T1 steady state. Perturbation of the signals from their T1 steady state causes bright/dark artifacts in the current volume of data, and possibly some 'hangover' artifacts in one or two later volumes as the steady state is reestablished (assuming there is no further motion).

There are a few simple tactics that can lessen, but won't eliminate completely, the effects of through-slice motion on the T1 steady state, including the use of contiguous rather than interleaved slices, and a small RF flip angle. With respect to slice ordering, brief, acute movements along a contiguous slice axis tends to restrict perturbation of the T1 steady state to just one or two slices whereas the same movement along an interleaved slice axis will tend to produce banding across the entire slice dimension.

Using a small RF flip angle makes the need to maintain a T1 steady state less stringent, but again the issue isn't entirely eliminated. A flip angle of, say, 15 degrees and a TR of 2000 ms will provide near complete relaxation from TR to TR. However, we aren't out of the woods because a movement in the slice axis that is also in the direction the slices are being acquired might cause quick repetitive excitation of some or all of the slice(s) just acquired. Rotations create further complication, introducing the possibility of regional steady state perturbations in one or a few slices. So low flip angles can help but are not a complete fix.

A method presented in a recent paper sought to eliminate TR-by-TR steady state effects through the use of a 2D navigator echo ("A dual echo approach to motion correction for functional connectivity studies." A. Ing and C. Schwarzbauer, NeuroImage 2012.) Essentially, two complete sets of slices are acquired each TR, with the first set of slices used as non-BOLD-weighted navigators for the second set. There are important methodological limitations - parallel imaging was needed to attain a practicable per slice acquisition time, introducing a new form of motion sensitivity - and there is a considerable temporal overhead inherent in the scheme - essentially, we have to acquire the same k-space twice per slice - but it is an intriguing idea that might stimulate further pulse sequence-based approaches to motion mitigation; perhaps as a component of some ideal prospective motion correction scheme.

Thus, in order to properly compensate for T1 effects we should be considering only slice-by-slice correction schemes rather than TR-by-TR correction schemes for multislice 2D imaging. (True 3D imaging has its own set of problems that I won't address in this post.) If the motion detection method is perfect such that the slice prescription can be maintained precisely on the anatomical targets then, in principle, the T1 steady state won't get perturbed no matter how much movement occurs. The scanner would simply 'lock on' to the brain anatomy and follow it faithfully. So far so good.


Shim effects

It is possible to update the shim gradient coils in real time on a suitably modified scanner. However, the problem is in knowing what the correct shim currents need to be for each potential head position inside the magnet! You will know that it takes some 20-30 seconds to acquire a single 3D field map shim on a modern scanner; this gives (hopefully) the best currents to use in the shim coils for the head position during the field map shim acquisition. The moment the subject moves his head then the non-linear magnetic fields inside the brain will change and the shim currents will be inappropriate to some extent. Where and by how much the shim field is now incorrect depends on the nature of the movement, the size and shape of the head, the number of shim coils, etc. But the bottom line is that this is an incredibly complex spatial dependency.

In principle, it might be possible to acquire a series of shim fields for different head positions in some sort of pre-scan, then use an interpolated set of shim values in concert with slice position updates as components of a simultaneous prospective motion correction method. But the way I see this problem today, the non-linearity and complexity of the magnetic susceptibility gradients in many interesting brain areas, particularly the frontal and temporal lobes and the inferior brain surface, limits even the potential of such methods. Correcting for small regions, such as for an MR spectroscopy volume, seems eminently feasible, but I don't foresee a global brain solution arising any time soon. Feel free to prove me wrong, I'd love to have such a method on my scanner!

There's another issue that needs to be considered, too: non-head motion. We already know that chest movement from breathing modulates B0 across the head. What about movement of the lower body or extremities? A subject could move his arms and legs to get comfortable, because he's fidgety or because the task requires it. How will these movements affect the magnetic field homogeneity across the head? (Extremity movement was shown to be a problem for GRAPPA.) Which body movements can be ignored safely, if any? In evaluating any prospective motion correction method we will need to consider whether the effects of non-head movements will remain in the data and perhaps require a further correction step during post-processing.


Receive field heterogeneity

As was shown in a post on EPI artifacts, modern phase array RF receive coils impose a spatial heterogeneity across the brain. (In a forthcoming paper (see Note 1) we investigate how this lab frame contrast mechanism interacts with movement to produce major challenges for motion correction.) The receive field is fixed to the magnet, so any change in the subject's head position will cause changes in signal intensity that will either confuse the prospective motion scheme, if the correction is based on the content of the images, or be missed by the scheme entirely. Perhaps it is possible to make a set of reference maps in a pre-scan, as suggested for the shim fields, but the performance of such an approach is likely to be severely limited; the more so the higher the coil array, as the receive field heterogeneity is increased.

There could be a simple solution to the receive field issue: eliminate it. One could nix the phase array coil in favor of a plain vanilla birdcage receiver; one that is sufficiently large that its receive field can be considered constant relative to the brain displacements during each fMRI run. But this hardware isn't common on modern MRI scanners, and the decreased SNR that comes from a birdcage coil would likely degrade the performance of other scan types where thermal noise (rather than physiologic noise) may be limiting, e.g. diffusion-weighted imaging or high-resolution anatomical scanning. Then again, we're talking about subjects who move a lot, so the SNR and thermal noise limits could be utterly irrelevant! Remember, we're in a subject-motion-limiting regime. If we weren't we wouldn't need to be considering such convoluted acquisition strategies!


Transmit field heterogeneity

This isn't much of a problem right now because most of us use large 'body' B1 transmit coils that can be considered homogeneous relative to the dimensions of a human brain. But the advent of multi-channel transmission, especially for high magnetic fields (7 T), will provide another challenge analogous to the receive field heterogeneity issues just considered. Prospective motion correction schemes that work perfectly well on today's scanners might end up compromised on future scanners if the B1 transmission contains significant heterogeneity relative to the image contrast. But I'm afraid I don't know enough about parallel transmission to offer anything more concrete than a vague concern, based entirely on what I do know about receive field heterogeneity. It's an issue that I shall watch with interest as parallel transmission moves out of the engineering realm and towards neuroscience.


Other considerations for prospective motion correction schemes

Now that we have assessed the primary motion-dependent instabilities in the time series data we can start to look at a few general principles of prospective correction schemes. I'm not going to try to specify particular hardware or software designs, rather I want to consider whether certain approaches might contain intrinsic problems that mandate one approach over another.


Aims of a prospective motion correction scheme

As mentioned at the outset, what we generally mean by 'motion correction' is the maintenance of anatomical content over time. We thus require as a basis the measurement of brain and/or head location inside the magnet. The usual approach is to measure displacements of the subject's head - usually it's just the head, not the entire body - either externally with some sort of non-MR measure (such as a camera or other optical recording mechanism), or internally via some metric derived from MR data itself (such as the anatomical content of EPI slices). The slice positions are then updated based on the displacement information just recorded, and hopefully the MR slices remain at a constant position relative to the subject's brain anatomy. Simultaneously there is an implicit goal to maintain all other sources of signal (and noise) modulation constant across time.


Prospective correction of head movement effects: when should we do it?

Most prospective schemes proposed to date, e.g. PACE, have operated on a TR-by-TR basis, which is equivalent to a volume-by-volume update. A more sophisticated but technically more challenging approach is to attempt to update each slice position in "real time," i.e. to apply a correction on a slice-by-slice basis. However, as we saw above when considering the main motion dependencies, only slice-by-slice correction is expected to be capable of avoiding some major limitations.

Essentially, what we're considering is the relative time scale of the motion and the motion correction. If the motion is slow with respect to the TR - a drift over tens of seconds to minutes, say - then a TR-by-TR correction is probably sufficient. But if the motion is relatively fast - a swallow, a sneeze or adjusting for comfort, say - then the motion will likely be faster than the TR. For typical repetition times of 2-3 seconds the motion correction isn't going to be able to keep up; some of the slices will be acquired assuming old, incorrect position information.

Of course, motion that is rapid compared to the acquisition time of a single 2D slice, which might be around 50-70 ms, is going to cause problems even for a slice-by-slice correction scheme. But, given that such rapid motion will also likely cause problems during a slice acquisition as well as between slices we can see that there will be a fundamental limit to how well any prospective motion correction scheme can perform. For the remainder of this post I shall assume that (correctable) motions of interest may be rapid relative to TR but are slow relative to a single slice acquisition. We can debate whether this assumption is useful in practice, given that motion during a slice is essentially uncorrectable. (See Note 2.)

There is one further benefit to a slice-by-slice correction which was pointed out by my colleague, DS in a comment on Neuroskeptic's blog:

"Another potential benefit of independent prospective correction of motion would be the decoupling of the temporal interpolation (often used in fMRI) and motion correction. With retrospective correction this decoupling is not possible. Nevertheless folks have been proceeding as though it were. It would be really great to decouple these problems."

Temporal interpolation is also known as slice timing correction. Decoupling the slice timing from the motion correction steps should, in principle, enhance performance of both steps, but it remains to be seen how much benefit might be attained in practice. Still, for our i-fMRI scanner I think we should be aiming for slice-by-slice correction.


Measurement of head movement: how should we do it?

I am ambivalent on the particular measurement scheme used to provide spatial information for a prospective motion correction scheme. However, there is an important observation that can be made. As already noted, there are essentially two forms of spatial information we could use. Either we could record position information using the MRI data itself - an internal measure - or we could use something else entirely; any external scheme that isn't using the time series MRI data itself as the template. In the latter camp I include video and optical measures, for example.

If the motion correction technique uses the MRI data as the basis for the correction then, for optimal results, the measurement will need to be updated slice-by-slice, as already specified. We then need to add one more restriction: the spatial information obtainable from the most recently acquired slice must be sufficient to accurately correct the next slice. If there is poor or low signal, e.g. an axial slice that captures just the very top of the head, or perhaps it's pure noise in a slice above the head, then there is very little (or no) information with which to establish whether the subject has moved recently or not. Significant motion could go undetected for several slices. And if that motion causes subsequent slices to be contaminated by motion, e.g. perturbation of the T1 steady state, because of undetected movement in the slice select axis, then it is possible to obtain corrupted target information that leads to an inaccurate specification for the update. It could take several subsequent slices to get back on track. (See Note 3.)

Perhaps the previous paragraph implies that external motion detection is the only way to go. I don't know. But I do know that the external motion detection schemes publicized to date only measure the head position, they make no attempt to estimate the shim field or the receive field, say. Perhaps this implies a combination approach; MRI plus external measures. There are several fields that need to be measured/estimated if we are to approach a "fix" for subject movement....

Wither prospective motion correction?

Certain types of motion are easier to fix than others. Slow drifts, whether from scanner heating, muscle relaxation or (head support) foam compression, should be amenable to prospective motion correction. But under the right circumstances they are amenable to retrospective motion correction, too. Would we have produced sufficient gain from our prospective correction scheme if it works only under a restrictive set of conditions, especially if we can't know ahead of time whether those conditions can be maintained?

I can see strong motivation for using prospective motion correction methods for clinical MRI, and perhaps for fMRI subjects who, for whatever reason, can't or won't remain still for a scan. Candidate subjects might include neonates, children and poorly compliant adolescents, adults with movement disorders, and perhaps "normal" fMRI subjects performing an experiment where head restraint is incompatible with the task. In all these situations the alternatives might be worse than dealing with the limitations of a prospective correction method, when faced with the option of not acquiring fMRI at all. However, we do need to recognize that what is on offer probably won't be a panacea for motion; there are too many complicated ways that motion can modulate the fMRI time series to assume that "motion is fixed" by using any sort of motion correction, prospective or otherwise.

People are belatedly recognizing systematic motion confounds in group studies, especially in resting state fMRI data, and realizing that motion correction algorithms don't remove all the effects of motion from the data. It would be nice if we could avoid repeating this mistake with prospective motion correction methods. We need validation, and we need to know what happens when the methods fail. Thus, when someone offers you the option of some fancy new prospective motion correction scheme you should ask "What residual motion dependency is there in my data, and what does it mean for my experiment?"

FMRI is endless compromises, we should expect prospective motion correction to be but another one on the list. At the present time I remain unconvinced that prospective motion correction - however it is implemented - is a panacea for motion confounds in fMRI. I think we have some good starting points for more engineering, but that's about it. What do you think?

__________________

Notes:

1.  The paper, written by my colleague, DS from the MathematiCal Neuroimaging blog investigates the magnitude of spurious signal changes arising from small movements, and shows that these changes compete with, and possibly exceed, the magnitude of BOLD changes. We suspect that what we term the "receive field contrast-motion correction" (RFC-MoCo) effect is responsible for a significant fraction of the erroneous correlations in resting-state fMRI data. As soon as the paper is available I'll link to it from this blog.

2.  Movement during k-space acquisition is always going to be a major problem that will lead to blurring, dropout, ghosting and other unwanted image artifacts. We tend to assume that single-shot EPI is sufficiently rapid to avoid motion contamination during the multi-echo readout train, but it's clear that rapid movement will compromise this assumption. At the present time there's not much that can be done about corrupted k-space, unless one has excellent navigator echoes and/or independent information on the interaction of the movement and the modulated k-space trajectory. And even then the results may not be pretty! Hey, we can't expect to fix every form of movement! There has to be a limit. I mean, we pose for photographs for a reason, no? Every technology has a limit.

3.  The PACE method, which does a TR-by-TR update on the Siemens Trio scanner, nicely demonstrates this limitation. Motion that happens during the previous TR is embedded into the template for the subsequent volume of data to attain. Thus, movement during the previous TR contaminates that volume but also has a "hangover" effect for the next TR, too. The method tries to make the next TR match the current, artifact-contaminated one! A single acute movement may take several TRs to "work its way out" of the compensation scheme, thus prolonging the deleterious effects in the data. A slice-by-slice correction using image information as the basis for correction would have much reduced sensitivity to this "hangover" effect, but it wouldn't be completely immune. It all depends on how well the subject's motion - that is, the new position - can be deciphered from the most recently acquired image plane.