Education, tips and tricks to help you conduct better fMRI experiments.
Sure, you can try to fix it during data processing, but you're usually better off fixing the acquisition!

Tuesday, February 9, 2016

Starting points for SMS-EPI at 3 T

Several people have approached me for advice on using simultaneous multi-slice (SMS) EPI for fMRI experiments. This is the sequence also known as multiband (MB) EPI. I'll come back to nomenclature in a moment. First, though, a brief introduction to what may become a lengthy series of posts. I'm going to focus on BOLD-based fMRI exclusively for the time being - sorry diffusion and ASL folks - and because I presently only have a Siemens Trio at my disposal, everything I write will have strong bias in that direction. That said, I do anticipate writing later posts dealing with SMS-EPI (for fMRI) on a Siemens Prisma at least, and I can already envisage a need for posts dealing with receive field normalization, in-plane parallel imaging, distortion correction options, reconstruction options and multi-echo SMS, to name just a few advanced topics. But first things first - to get going!

Options for SMS-EPI on a Siemens 3 T scanner

I am aware of three SMS-EPI pulse sequences for a Siemens Trio. One comes from the University of Minnesota's Center for Magnetic Resonance Research (hereafter CMRR), one comes from the Martinos Center for Biomedical Imaging at Massachusetts General Hospital (hereafter MGH), and one comes from Siemens as a work-in-progress (WIP) aftermarket sequence. For this post I'm going to be using the sequence provided by CMRR. Since CMRR refer to their sequence as multiband (MB) EPI I shall stick to this nomenclature here, and reserve the term SMS-EPI to apply to the broader family of pulse sequences. I may do posts on the MGH and WIP sequences in the future, but the CMRR sequence has been used the most broadly to date (e.g. the Human Connectome Project, which I'll discuss at length below) and so it offers the most immediate, road-tested place to start.

Other pertinent background. In addition to being on a Siemens Trio, I am going to assume you have a 32-channel receive-only head coil. The MB-EPI and other SMS-EPI sequences can be made to work with a 12-channel coil but only in a much reduced fashion because the 12-channel coil is a Total Imaging Matrix (TIM) coil that generates a lower number of effective channels than the headline number suggests. Specifically, the 12-channel TIM head coil acts like a 4-channel coil (in Triple mode) and that means it has limited RF heterogeneity with which to encode simultaneous slice information. Furthermore, the geometry of the 12-channel coil isn't ideal because it comprises nearly parallel, only slightly curved struts that have very little spatial variation along the coil axis, which happens to be the axis most people want to slice along (for axial images). There's a comparison of my 12-channel to 32-channel coils in an earlier post on receive field heterogeneity as an artifact rather than a feature.

Your first task is to obtain the CMRR pulse sequence, for which you'll need a master research agreement (MRA) with Siemens and then a C2P agreement with CMRR. Not 100% sure what C2P stands for but I believe it's Core Competence Partnership. In any event, once you have the sequence in-hand you should of course read the release notes and review their gallery of images before you do anything more. I would also suggest that you or your facility physicist spend a couple of hours perusing the history of the current and closed issues in their GitHub. You may well find some of your technical questions already answered. And given that this is a relatively high risk experiment compared to single shot EPI, you or someone around you is going to want to have a decent amount of expertise testing on phantoms and volunteers before you begin using SMS-EPI for an entire study.

Why SMS-EPI at all?

For most fMRI researchers, SMS-EPI is likely to be of interest when you want voxels with dimensions smaller than about 2.5 mm on a side, and/or you want a repetition time (TR) substantially below 2 seconds while maintaining at or near whole brain coverage. There are several published studies showing advantages of SMS-EPI for fMRI. A PubMed search for "multiband EPI" and then "SMS EPI" will locate them. Add "BOLD" as an additional search term if you want. Advantages of SMS-EPI versus regular multislice EPI may include better characterization of physiologic noise, leading to better discrimination of resting state networks, higher spatial resolution without losing whole brain coverage, and possibly higher sensitivity to BOLD changes. I'll leave the motivation to you, but you will definitely want to review all the published studies relevant to your neuroscience. If in doubt, keep it simple!

What the Human Connectome Project recommends

The HCP team wrote two papers containing essential information on their fMRI protocol. I'm going to review the salient parts of these papers before I get into their specific recommendations because what they did for HCP is actually different than what they recommend. Why? Because they had a customized 3 T scanner - the "Connectome Skyra," with more capable gradients than any stock 3 T.

In the paper describing the resting-state fMRI portion of HCP (doi:10.1016/j.neuroimage.2013.05.039) they explain how they settled upon 2 mm isotropic voxels, noting:
Pilot studies using a range of resolutions and EPI accelerations indicated that reducing voxel size to less than 2 mm was not beneficial at 3 T, in terms of the spatial detail discernible in the rfMRI resting-state correlation structure. This was not only due to the reduction in SNR with increased spatial resolution, but also the relatively large point spread function of the BOLD effect at 3 T (Parkes et al., 2005), which was measured to be ~ 3.5 mm at full-width-at-half-maximum (FWHM) caused by the dominant draining vein contribution at this field strength (Uludağ et al., 2009).
In the finalised protocol this acceleration is used to acquire 2 × 2 × 2 mm data with a temporal resolution of 0.72 s. While a larger voxel size would result in even faster imaging and better SNR, this choice provides a good overall balance in which both spatial and temporal resolution are a significant improvement compared with conventional rfMRI datasets.

Other key parameters followed from the voxel size and consideration of signal losses relative to BOLD sensitivity:
Echo-time (TE) was, after much discussion and evaluation, set to 33 ms. Again, this choice is a trade-off; long TE increases BOLD contrast, but decreases overall signal level and increases signal dropout in areas of B0 inhomogeneity. The TE for optimal functional CNR is equal to T2* when thermal noise dominates; however, T2* varies spatially, meaning that no single TE can be optimal throughout the entire brain. Acquiring multiple echoes in a single EPI readout train, or in separate acquisitions with multiple TEs, was not acceptable due to significantly prolonged readout duration and/or TR. Thus, the shortest TE that could be achieved without the use of partial Fourier or in-plane accelerations was selected, to minimise signal dropout. At 2 mm resolution, with the Connectome scanner gradients, this TE was 33 ms, given the excitation pulse width (~ 7 ms) required to achieve the (Ernst) flip angle (52°) for multiband x8. The use of partial Fourier to reduce TE resulted in larger signal dropouts than acquiring a full Fourier coverage of k-space with longer TEs (likely caused by local phase ramps in regions of B0 inhomogeneity shifting signal outside the acquired k-space region); hence partial Fourier was not utilised. The EPI echo train length is 52.2 ms for the final HCP fMRI protocol.

These are useful points to keep in mind as you proceed. If you want to review all the MB-EPI parameters as used in the HCP, check out pages 41-3 of the PDF accompanying their release of data from 900 subjects. The HCP team also wrote a paper on their minimal processing pipeline (doi:10.1016/j.neuroimage.2013.04.127) which you will want to read before setting up your own experiment, but we don't need to dig into that just yet because, as I mentioned above, what the HCP recommends we do is different to their setup.

So, just what do the HCP team suggest we do on a standard 3 T scanner such as my Trio? As noted above, we don't have a high-performance custom gradient set as on the "Connectome Skyra" so we have a few more limitations to deal with. (Perhaps those of you with a Prisma will be able to push things harder. I'll deal with Prisma-specific protocols in a later post.)

For the HCP recommendations we need to review pages 3-7 of that PDF accompanying the 900 subjects data release:
For functional imaging, key choice points relative to the HCP fMRI acquisitions involve the multiband (MB) factor, spatial resolution, TE, and phase encoding direction (the latter three of which all interact). While gradient strength is not as critical for fMRI (relative to dMRI), the Connetome Skyra gradients do allow it to operate at a lower echo spacing than a conventional 3T scanner (e.g., 0.58 ms vs. 0.69 ms at 2 mm, all other things being approximately equal). The limitations of maximal readout gradient [Siemens Trio (TQ) ~ 28mT/m, Verio (VQ) and Skyra (XQ) ~ 24 mT/m] and forbidden echo spacing (due to acoustic resonances) make 2 mm more of a “stretch” resolution on these 3T magnets.

More of a stretch? Yes, sadly that's not just a figure of speech. The echo train length will necessarily increase on a scanner with weaker gradients. We'll deal with this issue for the in-plane optimization, below. Until then, here's what the HCP summary recommendations look like:
Note that considerable benefit as regards the accuracy of the mapping of activation to the cortical surface is already achieved by going to a 2.5 mm isotropic resolution, albeit with further incremental gains in accuracy in going down to 2.0 mm (Glasser et al. 2013). Overall then, we recommend that users of Trio, Verio, and Skyra systems test resolutions of 2.0 to 2.5 mm for fMRI and make a selection based on their requirements for temporal SNR, statistical power, and acceptable degree of susceptibility distortion, and signal dropout.
The good temporal stability of the Connectome Skyra and the low electronic noise of the Siemens Tim 4G© platform allow the HCP to robustly generate good quality BOLD data at an MB factor of 8 without in-plane acceleration. Users of other systems will want to look carefully at whether they are happy with the levels of residual aliasing and temporal SNR at MB=8. In general, we recommend a MB factor of MB=6 for robust image quality while retaining high temporal resolution for these systems.

There's one final piece of advice concerning the phase encode direction. HCP opted to use a rectangular field-of-view, with phase encoding alternating L-R and R-L. But they don't suggest we do this:
....we recommend using either anterior-to-posterior (AP) or posterior-to-anterior (PA) phase encoding (rather than the RL and LR phase-encoded pairs used in the HCP acquisitions), so that there is not a right/left susceptibility asymmetry (bias) in the aggregate data. In pilot testing, we could not discern an overall preference for either AP or PA phase encoding, since each resulted in a different amount of signal dropout and local distortions in different brain areas with susceptibility artifact, and this dropout differs greatly depending on slice orientation (e.g., T>C vs C>T). Thus, we recommend that users make the choice between AP and PA phase encoding based on their own particular research aims and goals. Note that AP or PA phase encoding will require use of a full FOV in the phase direction (“FOV phase = 100%”), which will lengthen the total echo train, leading to some increase in T2* blurring, susceptibility distortion, and signal dropout (via increased minimum TE) compared to the HCP acquisitions. In practice, this effect will be at least partially mitigated given the shorter minimum echo spacing achievable in the AP/PA phase encoding direction (due to lower peripheral nerve stimulation limitations with AP/PA than RL/LR). “Compensating” for these effects via use of partial Fourier and/or in-plane GRAPPA involve their own tradeoffs (e.g., for in-plane GRAPPA, reduced image SNR and a lower acceptable maximum multiband acceleration factor and thus longer minimum TR). The HCP investigated these tradeoffs to some degree during pilot testing, and ultimately settled on RL/LR phase encoding with no partial Fourier or in-plane GRAPPA as yielding the best overall quality on the Connectome Skyra. For users of other Siemen’s 3T systems desiring 2.3 – 2.5 mm isotropic spatial resolution with only a single phase encoding direction, we suggest trying AP or PA phase encoding without in-plane GRAPPA or partial Fourier (allows a minimum TE of ~ 33 ms), and a multiband factor of 6. For users desiring 2.0 mm resolution, 7/8 partial Fourier may be desirable (allows a minimum TE of ~ 36 ms).

These suggestions give us sufficient information to proceed. To make this a tractable project I shall first describe in general terms the main factors to consider, then put all the decisions together into three explicit starting point protocols.

Slice dimension optimization

To a first approximation, the amount of slice acceleration - what is generally called the MB or SMS factor - depends upon the RF coil selection. As noted, we are assuming a 32-channel head coil. Not only does this coil have more independent channels than the standard 12-channel coil, but the elements also have very different geometry. In the 12-channel coil the elements are arranged as near parallel struts; there is little receive field variation along the coil's length. The 32-channel coil's elements are arranged more like the hexagonal patches on a soccer ball. There is a high degree of receive field anisotropy parallel as well as perpendicular to the coil axis. Thus, in principle we might choose to slice along any axis when using the 32-channel coil.

So, what direction should we place our slices? Efficiency dictates that we cover the shortest brain dimension with the slowest sampling. Since slices acquire more slowly than either of the in-plane dimensions, even with SMS enabled, it means that we should expect to use axial or axial-oblique slices. The 32-channel coil has plenty of receive heterogeneity along the axial direction (magnet Z axis) to perform good MB-EPI. You could use AC-PC or some other oblique axis equally well.

How fast should we go? The HCP suggests that we set MB as high as 6. This results in good images on my Trio and so this has become my practical limit, too. Would I consider going to, say, 8? I might, but it's worth bearing in mind that an MB of 6 is already six times faster than regular multislice EPI. That's a phenomenal gain! An MB of 8 is only 33% better than an MB of 6. And it's higher risk. Conversely, if an MB of 6 nets all the brain coverage you want with a minimum TR that is much shorter than you require, e.g. to use a particular task script that triggers off the TTL pulses happening once/TR, then you should consider dropping the MB to 4 or even 3 to reduce the possibility of reconstruction artifacts, and to decrease the image reconstruction time. (Vide infra.) In sum, then, accelerate as much as you need to, but not more!

In-plane optimization

You might think that in-plane optimization should be trivial because it's the slice dimension being accelerated in SMS-EPI. In practice, however, the separation of the simultaneously acquired slices needs some help from a slight modification to the in-plane gradients, via a method called blipped CAIPI. Reconstruction artifacts - say we try to push the acceleration factor higher than can be supported by the RF coil - may appear in the image plane, not just as modulations only visible in the slice dimension. You may see these artifacts described as "residual aliasing" or, more commonly for the SMS literature, as "leakage artifacts." I'll look at these in detail in a subsequent post.

There is also a prosaic reason why we shall want to change the parameters in-plane. When the slice thickness goes down there is a tendency to want to improve the in-plane resolution simultaneously. Many (most?) people seek voxels having nominally isotropic dimensions - cubes - or very close to it. What this means is that because we have enabled higher resolution in the slice dimension we're now motivated to drive the resolution in-plane. There's no obligation to do so, it is a choice but it's one that many of you will succumb to.

Following the lead of the HCP recommendations, there are a few guiding principles that we shall use for our starting point protocols. Unfortunately, high resolution in-plane takes gradient time and so the minimum TE will end up being longer for high in-plane resolution than for the 3 to 3.5 mm voxels you're used to. Our guide will be to try to set TE as close to 30 ms as possible, but not longer than 40 ms, in order to match the typical T2* of most of the brain. We should expect regions prone to signal dropout - the frontal and temporal lobes - to suffer more the longer TE gets.

We generally want to drive the echo spacing as short as it will go, while noting that some scanners have mechanical resonances that can make ghosting bad. Newer Siemens operating software - certainly the SyngoMR VD line as used on Skyra, for example - generally prohibits the use of an echo spacing known to cause mechanical resonances. But on my Trio I can set the gradients to any physically plausible value and thus I need to be aware of the consequences.

Finally, use of GRAPPA in-plane is not yet encouraged by me (or HCP) unless you take special measures to ensure that head/body motion is considerably lower than used in most routine fMRI studies. All the subject has to do is swallow during the autocalibration scans (ACS) and a whole time series can be wrecked. Also, in agreement with the conclusions of the HCP team, I've noticed that the raw image SNR is far lower when GRAPPA is enabled. So no GRAPPA for now. There are a few recent options for ACS that may make GRAPPA sufficiently robust we might want to consider it in future, so I'll do a dedicated post on using GRAPPA with SMS-EPI soon. Until then, if you insist on enabling GRAPPA with your SMS-EPI then you should probably take your cue from this paper. (Note, however, that GRAPPA was always enabled in their comparison, no non-GRAPPA option to compare against.)

Dealing with dropout

For voxels smaller than 2 mm on a side we shall necessarily have a TE in excess of 30 ms, causing more dropout in frontal and temporal regions than you would like. While the HCP was able to avoid using partial Fourier courtesy of their zippy custom gradient set, for us it is going to be a useful trick. It's not without consequences, but we have a degree of control over their severity by setting our partial Fourier fraction in conjunction with the direction of the phase encoding; either A-P or P-A. If you aren't familiar with partial Fourier, especially the interaction of partial Fourier with the phase encoding direction, then I strongly suggest you read the three posts linked in this paragraph. For MB-EPI we shall be using the Siemens default "early" echo omission, in order to reduce the minimum TE. (There is a "late" option but it's not useful for us when we want to shorten TE.)

Dealing with distortion

There are generally two reasons to want to use in-plane acceleration such as GRAPPA. One reason is to try to shorten the TE, the other reason is to reduce the amount of distortion in the phase encoding direction. We are already expecting to employ partial Fourier to reduce TE below 40 ms for MB-EPI, so that leaves the latter reason: distortion reduction. Now, we are going to have severely distorted MB-EPI images because we are expecting to employ high resolution, and high resolution in EPI means long echo trains. Long echo trains produce severely distorted EPIs. Sorry. This is one of those situations where we have to choose between two unpalatable options: motion sensitivity or distortion. For now I choose distortion because there is some hope to mitigate the worst effects with a post-processing step employing a field map, or functional equivalent. Motion, on the other hand, tends to trash a time series irrevocably.

So let's take a quick look at what the HCP did for distortion in their protocol:
To more rapidly measure the B0 field for correction EPI distortions, we acquire two spin echo EPI images with reversed phase encoding directions (60 s total for 3 pairs of images). (Note that we refer to these images as spin echo field maps, though they measure the field by reversing the phase encoding direction, which is a very different mechanism from standard field maps that use a phase difference calculated from two different TEs.) These spin echo EPI images have the same geometrical, echo spacing (0.58 ms in our scans), and phase encoding direction parameters as the gradient echo fMRI scans. These images enable accurate correction for spatial distortions in the fMRI images so that they can be precisely aligned with the structural images. Two of these spin echo EPI field mapping pairs are acquired in each functional session, for added robustness with respect to acquisition errors and subject movement, along with one set of B1− receive field-mapping images (with identical parameters to those described in the structural session).
For the functional pipelines, a field map is required, because any neuroimaging analysis that aims for precise cross-modal registration between functional and structural (or other data modalities) will require EPI distortion correction. In general, either standard gradient echo field maps or spin echo EPI field maps can be acquired, though spin echo EPI field maps can be acquired more quickly with less chance of motion corruption. The geometrical parameters (FOV, matrix, phase encoding direction, resolution, number of slices) and echo spacing must be matched between the gradient echo EPI fMRI timeseries and the spin echo EPI field maps.
EPI fMRI images contain significant distortion in the phase encoding direction. This distortion can be corrected with either a regular (gradient-echo) field map or a pair of spin echo EPI scans with opposite phase encoding directions. Reversing the phase encoding direction reverses the direction of the B0 field inhomogeneity effect on the images.

There are other options, such as the true (gradient-echo) field map mentioned by HCP. Siemens offers a sequence called gre_field_mapping and released as part of their BOLD software package. (Check the SIEMENS tree in the Exam Explorer.) But on my scanner at least, the minimum slice thickness is 2 mm and you may wish to match exactly the field map with the MB-EPI slices when the latter are thinner than 2 mm.

The issues of acquisition time, motion sensitivity, correction efficacy and so on make this a major enterprise and I don't have any quick suggestions for you, so I shall leave this extensive topic for a dedicated post once I have more details to offer. Until then I suggest you take a look at what Oregon's LCNI recommends, and perhaps consider using FMRIB's FUGUE software, in addition to considering HCP's suggestions.

Other issues

Shimming:  The CMRR-supplied protocols come with Advanced shim enabled and that should be considered the default. It takes about 90 seconds but the benefits in frontal and temporal lobes should be worth the extra time. If you're looking to nick a few minutes in a lengthy scan session then you might consider starting with the Advanced shim for the first fMRI run, then use Standard shimming for all subsequent runs.

Flip angle:  The HCP used an approximate Ernst angle for their TR; 52 degrees for TR = 720 ms. I tend to use a lower FA than the Ernst angle for two reasons. First, as shown in this paper from Gonzalez-Castillo et al. (doi:10.1016/j.neuroimage.2012.10.076), reducing the FA can reduce physiologic noise without reducing BOLD sensitivity. (Years ago on a 4 T we always used a very low FA, such as 20 degrees for a TR of 1-2 seconds, in order to reduce the inflow effects. It worked very well.) The second reason might be less of an option. The amount of subject heating (assessed by the specific absorption rate, SAR) scales as the square of the FA. So even a small reduction in FA can reduce SAR considerably. If you find yourself running into the SAR limit then you can reduce the FA until you're good to go. The down side to a lower than Ernst angle FA? Image contrast may be affected slightly, but since we prefer temporal stability over anatomical content then altered contrast shouldn't be a huge burden. (For Gonzalez-Castillo et al. it was an advantage because they were using water-excite rather than fat presaturation on a GE scanner and so they had low anatomical contrast at the Ernst angle.)

RF pulse width:  The HCP used an RF pulse width of 7120 us whereas the default on my protocols is set to the maximum of 10500 us. I haven't tried altering the RF duration at all yet, and I will note that I can get down to 1.5 mm slices using the default which is why I've not played with it. The "Connectome Skyra," with its higher peak gradient capabilities, could probably get thin slices with the shorter duration RF. On our stock scanners we should expect to need longer pulses. So until someone shows me why 10500 us is a bad choice I plan on continuing with it.

Receive coil heterogeneity:  According to the PDF accompanying data release, the HCP acquired receive field "bias" maps but actually used a different approach:
The BIAS_BC and BIAS_32CH scans are collected as analogs of Siemens' “Prescan Normalize” procedure, but these also are not being used. Rather, HCP is using the T1w and T2w scans for estimating the receive-coil bias field (see Glasser et al. 2013).
Having looked at both the PDF and Glasser et al. it's still not entirely clear to me whether they were using the T1W and T2W scans or the two named BIAS scans. Perhaps someone can enlighten me. In any case, I advocate enabling the "Prescan Normalize" option on MB-EPI provided you also select the option to save both the normalized and raw time series (enable "Unfiltered images"). Reconstruction time is dominated by the SMS unaliasing algorithm, not the normalization, and so the only significant cost to having two data sets is twice the amount of data to store. More on bias field correction in a later post, after I've had a chance to assess the pros and cons of the Siemens routine versus acquiring one's own reference data set.

Coil combine mode:  I would only use the default Sum of Squares for simplicity, as the HCP did. I don't know enough about the alternative methods yet, but my intuition is that fancier coil combination methods may not play nicely with the SMS unaliasing routine. 

Single band images (SBref):  The HCP used the single band references as their template for realignment (a.k.a. motion correction) and to permit more accurate registration of EPI to T1W anatomical scan. I can't say whether the SBref makes a better or worse template for motion correction than the first MB-EPI, the middle MB-EPI, or some average. This is a post-processing question that we can ignore at acquisition time, provided we ensure that we have the SBref saved along with the MB-EPI time series.

Reconstruction speed:  So this will likely be your loudest complaint when first using SMS-EPI. I am reliably informed that the GPU-based reconstruction engines on the newer scanners can process SMS-EPI in real time. Not so the Step IV MRIR board on my Trio, and this is the fastest reconstruction board you can get for the Trio. That's why it's useful not to set the MB factor to some gratuitously high number for the sake of having more data. More may or may not be better, but it will most definitely be slower. With MB=6 you can expect the reconstruction to run longer than the acquisition by between 1.5 and 3 times. Plan your session accordingly. Also note that a maximum of twelve reconstructions can be in the Siemens queue at one time. I've never hit that limit, or even come close. What tends to be a bigger deal is the need to keep the next user from terminating a reconstruction at the end of the current session.

Slice order:   Simultaneously excited slices already have a sizable gap between them such that (a) we don't need to specify additional gap, and (b) we can reconsider the desire to use contiguous slices for reducing motion sensitivity, as is common for regular single shot EPI. For an MB of 6 there will be a gap of five slice thicknesses between the simultaneous slices. For ordering, there is a difference between specifying interleaved slice order versus contiguous (ascending/descending) order, but it isn't a stark difference as for regular EPI. So for convenience I have stuck with interleaved slices. I doubt there is much practical difference if one selects a contiguous order, but that would necessitate me checking for artifacts all over again. I don't expect there to be an appreciable difference in motion sensitivity, for example.

Motion sensitivity:  I can only offer an anecdotal report on the motion sensitivity of MB-EPI. In my experience, moderate subject movement (a swallow, say) during acquisition of the single band reference images has a much smaller effect on image artifact level than the same motion would during ACS for in-plane GRAPPA. This is good! The HCP did some piloting and assessed the ability of ICA to clean the time series, but I suggest that you treat MB-EPI (and all SMS-EPI sequences) as intrinsically more motion-sensitive than regular EPI, and adapt your subject preparations and head restraint systems accordingly.

Putting it all together

I've created three introductory protocols for MB-EPI based on three different voxel sizes: 1.5 mm, 2.0 mm and 2.5 mm dimensions, all isotropic:

Main parameters for three starting MB-EPI protocols (click to enlarge)

The full parameters are available in a PDF, a printout of the full protocol from the Exam Explorer. The protocol was created using CMRR's MB-EPI sequence version R012c (R013 is out as of 14th December, 2015) on my Trio running SyngoMR VB17A. There are two sets of scan parameters for each voxel size in the protocol, one with phase encoding set A-P and the other set to P-A. Phase encoding reversal is achieved via the parameter "Invert RO/PE polarity" on the Sequence tab. Enabling this option is functionally equivalent to setting the "Phase enc. dir" dialog box to 180 (degrees), which is why you still see A-P listed under that parameter. All six scans have prescan normalization enabled, SBref saved and excite (RF) pulse duration set at 10500 us.

There are some caveats that require explanation. The 1.5 mm isotropic voxel protocol doesn't cover the whole brain in the TR of 1300 ms. It might just cover whole cortex with judicious slice tilting to snag temporal poles as well as the full frontal and parietal cortices. You'd need to extend the TR out to about 2000 ms to include cerebellum. You might consider setting MB up to 8 to improve the coverage, but I can't vouch for the leakage artifacts that might result.

For the 2.5 mm isotropic voxel protocol the whole brain is covered easily with 64 slices in a TR of 1200 ms. If you wanted an even shorter TR, down to 800 ms, you might consider setting the MB factor to 6 rather than 4.

The 2 mm isotropic voxel protocol uses an echo spacing of 0.69 ms, right in the middle of the range 0.6-0.8 ms where mechanical resonances for axial slices are expected to be worst on my Trio. Unfortunately, there is insufficient gradient strength to get below 0.6 ms, and setting the echo spacing at 0.8 ms would necessitate a TE above 40 ms, thereby requiring 6/8ths partial Fourier to get back below 40 ms. In my throwaway tests the ghosting from the echo spacing of 0.69 ms weren't terrible (see the example data below) whereas the dropout from using a longer TE and/or higher fraction for partial Fourier produced noticeable dropout. You would want to verify on your scanner that such an echo spacing doesn't produce show-stopping ghosts; test on a phantom. The mechanical resonances differ from installation to installation.

Example data

Below are 3D-MPR displays of a single volume of MB-EPI data acquired from the same brain using the three protocols just outlined. Here the "Prescan Normalize" filter has been applied. I also have 50-volume time series acquisitions with and without "Prescan Normalize" that you can download and tinker with yourself (295 MB zipped file). Included in the file are the "split" mosaic images I used to create the 3D-MPR views.

2.5 mm isotropic voxels (click to enlarge)

2 mm isotropic voxels (click to enlarge)

1.5 mm isotropic voxels (click to enlarge)

References and further information:

Human Connectome Project:
Appendix I – Protocol Guidance and HCP  Session Protocols
(900 Subjects Release, 8th Dec 2015)

The minimal preprocessing pipelines for the Human Connectome Project
MF Glasser et al., NeuroImage 80, 105-24 (2013)

Resting-state fMRI in the Human Connectome Project
SM Smith et al., NeuroImage 80, 144-68 (2013)

You may also want to look through the contents of the special issue of NeuroImage 80, 1-544, Mapping the Connectome, in which the above two papers were published. While I didn't use them for this blog post, the following two articles may be useful for context:
The WU-Minn Human Connectome Project: An overview
DC Van Essen et al., NeuroImage 80, 62-79 (2013)

Pushing spatial and temporal resolution for functional and diffusion MRI in the Human Connectome Project
K Ugurbil et al., NeuroImage 80, 80-104 (2013)

The following papers provide some useful background on the use of blipped CAIPI for reducing artifacts in SMS-EPI:
Rapid brain MRI acquisition techniques at ultra-high fields.
Setsompop K, Feinberg DA, Polimeni JR, NMR Biomed. 2016 Feb 2.
doi: 10.1002/nbm.3478

Evaluation of slice accelerations using multiband echo planar imaging at 3 T.
Xu J, et al., NeuroImage 83:991-1001 (2013).
doi: 10.1016/j.neuroimage.2013.07.055

A comparison of various MB factors with GRAPPA R=2 always enabled:
Evaluation of 2D multiband EPI imaging for high-resolution, whole-brain, task-based fMRI studies at 3T: Sensitivity and slice leakage artifacts.
Todd N, et al., NeuroImage. 2016 Jan 1;124(Pt A):32-42.
doi: 10.1016/j.neuroimage.2015.08.056


  1. Quick update for users of CMRR's MB-EPI: I tested the LeakBlock recon option and, as mentioned in the release notes, the noise is amplified. TSNR reduces as a consequence. For the 2 mm iso protocol the reduction in TSNR is around 10-20%, for the 2.5 mm iso protocol only about 5%, but it's 25-50% reduction for the 1.5 mm iso protocol. I would avoid using LeakBlock for the time being. That said, Todd et al. showed that it reduced false positives, but the problem there is that GRAPPA was always enabled in-plane (R=2) and so we can only guess that LeakBlock would similarly reduce false positives without GRAPPA. (As mentioned in the post, I don't advocate using GRAPPA because of the SNR reduction it causes, as well as its motion sensitivity.)

  2. Nice writeup.

    Your Trio actually does prohibit switching frequencies around known mechanical resonances, it just happens that for Trio the restriction (I think there is only one prohibited band for Trio) is not too bad for most things people want to do at 3T. On other systems this is not necessarily the case, and Siemens seems to be getting more conservative. For example, Prisma unfortunately has a resonance exactly where many of our users previously liked to operate on Trio, so converting protocols after the upgrade was not all joy.

    One thing that your Trio doesn't do very well, that is implemented more carefully in newer software, is model and predict the heating of the gradients. The Trio gradients are pretty robust, though, so it isn't usually a problem except for extremely long diffusion scan sessions.

    Regarding the RF pulse width, the best reason to keep it as short as possible is to keep the bandwidth as high as possible, which will reduce some dropouts due to intra-slice dephasing. This is an area where I believe the MGH approach is very different -- they tend to use VERSE fairly aggressively by default and also perhaps PINS pulses which have very low bandwidth. I don't know what Siemens is using in the various WiPs and the eventual product implementation.

    Another reason to keep the RF pulse short is to maintain the fidelity of the waveform. Siemens has a limit of 8192 support points per RF pulse object. In the CMRR C2P the pulses are all calculated on the fly with optimal resolution, so up to 8192 us you will have 1 us/point resolution. 10500 us will still give you 1.5 us/point resolution, though, which is probably fine. In early releases the resolution defaulted to 10 us/point, and at that resolution the waveform could be meaningfully distorted for large inter-band spacings (usually looks like intensity variations between bands).

    1. Many thanks for the feedback, Eddie. It will really help folks avoid pitfalls.

      Re.mechanical resonances, on Skyra and Verio systems I've seen "holes" in the echo spacing parameter whereas on my Trio it's solid green across the range. So I can park between 0.6-0.8 ms for axial (X frequency encoding) EPI slices, for example. There certainly doesn't seem to be any echo spacing I can't access.

      Re. gradient heating, putting aside the cooling errors the system has been generating coincidentally in the past 24 hours, our gradient cooling is probably about as good as one can get. We have a low humidity environment, 17 C in the exam room, and building chilled water at 8-10 C. We can remove an awful lot of heat. (I usually start to worry about the temp of the gradient cables between the magnet and the filter panel first, e.g. during overnight DW-SSFP of post mortem brains.) I've not tinkered much with DW-MB-EPI but readers would be well advised to assess their system's thermal stability.

      Re. the RF pulse width, this is good to know. I will try a comparison with a small handful of widths to ensure that I don't clip and to assess if there's a benefit in severe dropout regions. Since we have adopted relatively low flip angles (< 50 deg) for fMRI apps I would expect to have lots of flexibility.

      Thanks again for the comments, and also thanks for making the sequences available to the neuroimaging community. We very much appreciate it!