Archive for the ‘SMS’ Category

Hyperband Transition

September 14th, 2020

Executive Summary

The GE product Hyperband BOLD and diffusion sequences are a result of close collaboration between GE scientists and engineers with the CNI staff.  The CNI SMS implementation, which we called the “MUX” research sequence, served as a resource and a benchmark for the GE Hyperband product.  Over the last several months Adam and Hua worked with our partners at GE to resolve a limitation of the initial Hyperband implementation. We are glad to report that the two sequences now have the same SNR (signal-to-noise ratio).  We describe the testing that convinced us and the benefits below.  With this improvement, we plan to move CNI users to the Hyperband sequence, which will offer many benefits. Specifically, we ask all users of the UHP system to migrate any protocols using the research MUX sequence to the product Hyperband sequence.  We do not plan to support the research MUX sequence in future releases of the GE platform.  Instead, we will coordinate with GE on improvements to their Hyperband product.

Advantages

The main benefit for researchers moving to the Hyperband product sequences is that GE’s product releases are fully integrated with other system improvements. The CNI cannot achieve the same level of integration when we port the MUX research sequences.  For example, GE’s product Hyperband sequences are fully integrated with the reconstruction engine.  For this reason unaliased images show up in the mini-viewer and are stored in the image database which makes it possible for operators to immediately detect image quality issues.   Second, CNI’s resources for testing and migrating research sequences each GE system update is limited.  GE’s product sequences go through rigorous in-house acceptance testing with each new release.  Third, by moving to the product users will also benefit from all other system developments for the baseline EPI product (higher-order eddy current correction, bug fixes, etc.).  Fourth, the product sequence includes seamless integration with the User Interface so users simply prescribe whole-volume slices as normally.  Fifth, the product delivers near real-time reconstruction, rather than the slower reconstruction used by the research MUX sequence.

Image quality

We did not recommend using Hyperband until now because testing showed that the GE product SNR was 30% lower than the MUX performance.  The reasons for this SNR drop have been determined by CNI staff, and these corrections are now incorporated into GE’s future product releases.  We also incorporated these fixes into the CNI versions of the base sequences on our system, namely cni_epi and cni_epi2 which correspond to the Hyperband BOLD and diffusion sequences.  These improvements are why we now recommend shifting to the Hyperband sequence.  The shift will enable users to take advantage of integration with the other features of the GE platform.

Our assessment is based on several sets of SMS BOLD and diffusion data across multiple GE systems, including phantom and human subject data.  We acquired on the previous CNI MR750 scanner and the new CNI UHP scanner, as well as on the product sequences compared to our research sequences.  These data are all available on the CNI Flywheel site in the scanner_comparison project which is accessible by all CNI flywheel users.  A brief summary is provided here.

The figure below shows the detrended SNR determined across 50 timepoints for a region at the center of a phantom image acquired on our UHP scanner using the Nova 32-channel coil with both research and product SMS sequences.   The Hyperband sequence shows slightly improved for all non-unity SMS acceleration factors except for the SMS factor 8.  We hypothesize that these variations may be due to the different approach used between CNI and GE SMS sequences in acquiring calibration data to determine how to unalias the data.  CNI sequences generally acquire calibration data at the beginning of the BOLD acquisitions while GE uses a fast separate calibration sequence to acquire this data.  We will continue to investigate the source of this SNR drop at SMS 8, but given the performance at the normal acceleration factors that are used we are not concerned.  The SNR is also still quite high from an absolute perspective, and it is likely physiological noise would be of more concern in an in vivo experiment.

 

The following figure shows results acquired using the MUX research sequence across multiple platforms at Stanford.  The data were acquired using different Nova 32 channel coils that were available at each site.  As a result only data acquired at CNI used the same receive array, and this shows a substantial increase in detrended SNR when moving from the CNI MR750 system to the CNI UHP.  The Lucas 3T Premier system has higher SNR, and we hypothesize this is due to the receive array.  We will be following up with a future experiment to using the CNI coil on the Lucas Premier system.These SNR values are all quite high however, and likely data will be dominated by physiological noise for in vivo data.  Even so, we are considering recalibrating or purchasing a new coil.

 

We also evaluated the Hyperband diffusion sequence in terms of baseline SNR and artifact.  The following figure shows compares the research and product SMS sequences when using a whole-volume acquisition of an agar phantom with an SMS factor of 3 and no inplane acceleration.  A nominal b-value of 2800 was prescribed but the diffusion tensor file hand-edited to acquire 50 b0 images that were used to calculate the SNR.  We measured using a peak gradient amplitude of 50mT/m for the diffusion encoding lobes and for the product Hyperband sequence we also measured with a peak gradient amplitude limit that takes advantage of the UHP capabilities, at 100 mT/m.  In this last instance the echo time (TE) for the acquisition is significantly reduced due to the shorter diffusion encoding lobe durations that are required. As shown below, the Axial, Sagittal and Coronal reformats of the data all show good unaliasing and no noticeable banding artifacts.  The SNR for instances when the gradient amplitudes are limited to 50 mT/m are comparable across the MR750 and UHP regardless of whether our research or product SMS sequence is used.   The last measurement with the peak gradient at the UHP limit of 100 mT/m shows an appreciable SNR gain due to the shorter TE.

CNI staff are engaged with GE scientists to investigate different reconstruction approaches that will improve the product Hyperband sequence and reconstruction.  The current product uses a reconstruction method analogous to 1D-GRAPPA;  we are investigating a method analogous to split-slice-GRAPPA.  This latter reconstruction method shows no benefits as of yet, and requires considerably more computation and is not available for real-time reconstruction on the scanner.  Given the performance of the existing product reconstruction method we are confident in recommending its adoption.  At the same time, we will continue to work with GE to investigate possible improvements.

The scanner_comparison project on Flywheel has more data analysis than we’ve reproduced here, and raw data is also available should users be interested. While there is undoubtedly more comparisons that could be performed, there is evidence that the performance matches our research sequence performance, and there are other benefits offered to researchers by moving to the product Hyperband sequences which are significant.  These data and the general considerations are why we now recommend researchers switch to Hyperband.

Transition Assistance

CNI staff will be available to help with transitioning protocols to using the Hyperband product. Given the full integration of these sequences with the user interface and reconstruction we’re sure that users will find it a very easy transition to make.  If warranted we’ll add a channel to our Slack workspace for Hyperband to help with common questions (and yes, if you’re not on the CNI Slack workspace, please sign up now — see https://cni.stanford.edu/slack-and-volunteers for details).

Upgrade Plan for DV26

August 28th, 2017

Executive summary

We will be upgrading the GE software and computational infrastructure in late September (from DV25 to DV26). The reasons and implications for your projects are explained below in detail.

To migrate the CNI sequences to this new environment, the CNI development team needs time on the scanner (we estimate about 12 hours).  As most of you know, the schedule is very full. Thus, starting September 1st and continuing for a few weeks, the CNI will have priority for any released time and all protocol development time.

If you have already booked protocol development time, we may contact you to negotiate an alternative slot. We will return to a more open policy after the transition is completed. Thank you for your cooperation.

Background

Like many GE sites, we are now planning for a significant upgrade.  This upgrade, DV26, includes new computational hardware and software (but no new MR gear, such as coils or gradients).  This upgrade includes features that will eventually be valuable for many of you.

At minimum before we make the upgrade, we will be making sure the existing CNI-modified sequences (cni_epi, cni_epi2, muxarcepi, muxarcepi2, sprt, cni_3dgrass, cni_efgre3d,  cni_ir_epi and Probe-MEGA) will all be working at DV26 as they do at DV25, together with offline reconstruction for the mux and spiral sequences using NIMS.  Other existing product sequences are not noticeably changed in this upgrade. As a result, the transition for users should in most cases be seamless.

A beneficial feature of the new system is that GE has incorporated the SMS methods that were implemented by Adam, Kangrong, Bob, Hua  and Matt Middione (GE) at the CNI.  GE refers to their implementation as Hyperband (love the marketing folks; multiband was not enough).  The DV26 product includes only a Hyperband DTI sequence, but GE has agreed to enable us to use a beta version of the Hyperband fMRI sequence.

The user-interface for Hyperband will operate as any normal sequence. Simply prescribe the whole volume you wish to acquire and the online reconstruction will perform the slice and inplane acceleration unaliasing so that undistorted images appear in the mini-viewer and scanner image database. The new computational hardware from GE will perform these reconstructions, eventually reducing the burden on our aging CNI computers.

However, there are some limitations of the new product Hyperband sequences.  GE has not yet implemented our preferred reconstruction algorithm for Hyperband acquisitions (split-slice GRAPPA). Also, there are some support features in the CNI versions of these sequences (e.g. triggering selection) that are not in the product Hyperband sequences.  The CNI team will make modifications to these Hyperband sequences to support the specific CNI features as well as augment the product reconstruction.  Some of these updates, in particular supporting online split-slice-GRAPPA reconstruction, will not be completed until after our migration to DV26.

As a result of these limitations to the Hyperband sequences, and until we have the opportunity to confirm these sequences have the same performance of our existing SMS sequences, we advise users to continue to use the CNI SMS sequences (muxarcepi, muxarcepi2). However we expect that we will be able to recommend users migrate to these Hyperband sequences sometime later this year.  We will keep you posted on our progress.

The CNI Team

Updates for multiband reconstruction

March 15th, 2017

The CNI has recently introduced a new option for reconstructing SMS (aka, multiband or mux) scans. The default reconstruction method in the SMS reconstruction pipeline is currently 1D-GRAPPA (Blaimer M. et al. MRM 2013). Based on recent research and testing, we believe that the split-slice-GRAPPA (Cauley SF, et al. MRM 2014) reconstruction algorithm does a better job at unaliasing the simultaneously acquired slices, especially in cases where the calibration data are corrupted by subject motion. This more robust unaliasing will help reduce the chance of false correlations in fMRI scans by reducing signal leakage across aliased slices.

Another advantage of the split-slice GRAPPA method is that it is less dependent on the image contrast being consistent between the calibration data and the SMS data, therefore it allows more flexibility in choosing the best calibration scans. So, we have also introduced a new SMS calibration scan option – using a separate single-band scan as an external calibration for the target SMS scan. Most of you doing SMS are using internal calibration, i.e. the first few volumes integrated in the SMS scans are used as the calibration data. And a few of you are doing an external calibration that has the same SMS (mux) factor as the target scan. Compared to these calibration methods, the single-band external calibration has higher SNR in the calibration data and is less sensitive to subject motion during the calibration. Therefore we believe it is a more robust calibration method, and in combination with the split-slice-GRAPPA reconstruction method, is likely to produce better image quality for the SMS scans, especially when you have wiggly subjects.

Here is a compelling example of how the single-band calibration scan can reduce a particularly insidious artifact due to excessive eye motion during the calibration scans (this subject was instructed to intentionally move their eyes during the calibration scans). In these standard deviation maps of the BOLD timeseries, the aliased eye artifact is clearly visible in the 1D GRAPPA reconstruction (top). The split-slice GRAPPA reconstruction (middle) shows a reduced eye artifact in the white matter, but there is still significant aliasing of the eyes. This aliasing is substantially reduced or eliminated in the same data reconstructed using a single-band calibration scan with split-slice GRAPPA (bottom), even though that also had subject eye movement.

s1_grappa
s1_ssg
s1_sbref

However, in cases where there is no excessive motion during the calibration, all the methods are quite comparable:

s2_grappa
s2_ssg
s2_sbref

The CNI SMS data processing pipeline will by default keep using the original 1D-GRAPPA reconstruction method for continuity of ongoing studies. And if you have compliant subjects who remain still, ideally with eyes-closed during calibration, this method should be fine. However, if you think your subjects may move during calibration, then we recommend switching to split-slice GRAPPA for image reconstruction. And you may also consider adding a single-band calibration scan to your protocol. Any SMS scan that doesn’t have a separate calibration scan setup in the same scan session will also be reconstructed using the internal calibrations. In order to use the new methods, you need to do the following in your protocol:

To use the split-slice-GRAPPA reconstruction method, include the keyword “_ssg” at the end of each series description that you want to be reconstructed with split-slice-GRAPPA. Note that this also enables SENSE1 coil combination (Sotiropoulos et. al., MRM 2014).

To use the single-band calibration, you need to set up a separate scan with SMS (mux) factor = 1 (CV 22). Include the keyword “_sbref” in the series description. This single-band scan needs to have the same coverage as the multiband scan, i.e. the same FOV and matrix size but X times the prescribed number of slices, X being the SMS (mux) factor. You will need to increase TR (by a few seconds) in order to accommodate all the slices in one TR. You only need 4 or 5 phases for this scan. Because the TR will be longer, you can set the flip angle to 90 to optimize SNR.

For scans using the single-band calibration, the calibration volume is not included in the reconstructed images, i.e. the NIFTI files will only contain the SMS volumes. This information is also shown in the JSON file that now accompanies all SMS NIFTI files in NIMS. The JSON file contains a list of important parameters related to the scan and the reconstruction.

Please contact Hua or Bob if you have any questions and/or want help setting up these new SMS features in your protocol. Also, we owe thanks to the Wagner and Poldrack labs for help in testing these new methods.