^{1}

^{2}

^{3}

^{*}

^{4}

^{5}

^{4}

^{1}

^{6}

^{5}

^{4}

^{4}

Conceived and designed the experiments: DAF SMS KU EY. Performed the experiments: DAF SM EA SR EY. Analyzed the data: DAF SM SMS EY KLM MFG. Contributed reagents/materials/analysis tools: DAF SM EA SR SMS MFG. Wrote the paper: DAF SM SMS KU EY KLM.

Authors of this work, David Feinberg and Sudhir Ramanna, are employees of Advanced MRI Technologies, which is engaged in the development of magnetic resonance imaging pulse sequences. Some of the innovation and design of the pulse sequence presented in this work was done by Advanced MRI Technologies. This work was jointly developed with researchers at the University of Minnesota and Oxford University. However, this does not alter the authors' adherence to all the PLoS ONE policies on sharing data and materials, as detailed online in the guide for authors.

Echo planar imaging (EPI) is an MRI technique of particular value to neuroscience, with its use for virtually all functional MRI (fMRI) and diffusion imaging of fiber connections in the human brain. EPI generates a single 2D image in a fraction of a second; however, it requires 2–3 seconds to acquire multi-slice whole brain coverage for fMRI and even longer for diffusion imaging. Here we report on a large reduction in EPI whole brain scan time at 3 and 7 Tesla, without significantly sacrificing spatial resolution, and while gaining functional sensitivity. The multiplexed-EPI (M-EPI) pulse sequence combines two forms of multiplexing: temporal multiplexing (

In the last two decades, magnetic resonance imaging (MRI) techniques such as functional magnetic resonance imaging (fMRI)

Since its initial application, EPI scan time has not substantially decreased. Nearly all the successful efforts to shorten EPI acquisition times have targeted reducing the number of refocused echoes needed for spatial encoding to form an image (by means of partial Fourier _{2}*; the consequence is distortions and blurring on two of the 3D image axes, as well as a loss in signal-to-noise ratio (SNR). Multi-shot (segmented with multiple excitation) 3D EPI approaches that have produced high quality images _{2} and T_{2}* become inherently short. Another approach to reducing the scan time per volume in fMRI uses UNFOLD

The pulse sequence we introduce is based on a combination of two techniques for multiplexing signal acquisition, generating several EPI images following the contrast preparation time of a single EPI image. To increase imaging speed, the EPI pulse sequence incorporates temporal (

Each row of images was obtained with a different pulse sequence and slice acceleration, producing 1, 4, 6 and 12 slices from the EPI echo train. The

Although the SNR decreased with higher slice accelerations and faster TRs (see _{0}

^{2} utilizing 256 b-values in the twice-refocused diffusion encoding scheme (48). Imaging parameters: TR/TE 2000 ms/124 ms,

Two projections of the 3D diffusion spectrum image (DSI) reveal large scale neuronal connections in major white matter pathways. 3T Diffusion acquisition parameters: DSI using b-maximum of 4500 s/mm^{2} and 256 b-value encodings; Imaging parameters: SIR2 × MB2, 3 mm resolution, total acquisition time 8.5 min., TR/2000 ms and signal bandwidth 2604 Hz/pixel.

Resting state networks (RSNs) are patterns of spontaneous fluctuations that are coherent

The color overlays are z-statistic images, thresholded at Z = 4 in all cases. In this central axial slice (Z = +18 mm in MNI152 standard space) the RSNs shown cover visual areas, the default mode network and a sensori-motor network.

These differences are quantified in

The top row in each sub-figure shows a separate boxplot for each of the three TRs, and the bottom row shows the ratio of the 0.8 s and 0.4 s values to the 2.5 s values, the ratio being calculated separately for each RSN (and each subject) before feeding into the boxplot.

In the multiple-regression, 100dim analysis, the median ratio of the 0.8 s and 0.4 s data to the 2.5 s is 1.44 and 1.56 respectively. For single-regressions, these ratios fall to 0.99 and 0.87. In the 10dim analyses, the ratios are 1.10 and 1.34 for multiple-regression, and 1.27 and 1.19 for single-regression. Hence the main result is that for single-regression analyses, the peak Z-stat is quite similar across TRs, but for multiple-regression, the peak Z is 56% higher at the shortest TR compared with the longest. This improvement lessens at lower dimensionality (i.e., 10 RSN regressors instead of 100). We discuss below where this improvement comes from.

These results reflect the size (spatial extent) of the estimated RSNs, as well as the statistical significance. In the multiple-regression, 100dim analysis, the median ratio of the 0.8 s and 0.4 s data to the 2.5 s is 1.94 and 2.04 respectively. For single-regressions, these ratios fall to 0.89 and 0.72. In the 10dim analyses, the ratios are 1.26 and 1.35 for multiple-regression, and 1.30 and 0.95 for single-regression. Hence the main result is that for single-regression analyses, the sum-of-Z-stats (size × significance) is quite similar across TRs, but for multiple-regression, the sum-of-Zstats is 100% higher at the shortest TR, compared with the longest (with this improvement lessening at lower dimensionality).

These values are expressed as the standard deviation of each RSN's timeseries (estimated at the peak-Z voxel), as a percentage of the mean signal level. As expected, there is not a great variation in RSN amplitude across the different TRs, regression methods and dimensionalities (range 0.75–1.09), with a slight increase in amplitude apparent in low-TR data, when looking at the ratios of the 0.8 s and 0.4 s data to the 2.5 s (range 1.06–1.25).

These values are expressed as the standard deviation of the regression residuals (estimated at the peak-Z voxel), as a percentage of the mean signal level. The residuals are considerably higher (a factor of up to 2.8) in the low-TR data, because the raw SNR per timepoint acquired is reduced with shorter TR. (It is not until the Z-stats are considered, which take into account the increased number of samples in the low TR data, that the low-TR CNR is seen in fact to be as good as or better than the unaccelerated data.) As expected the multiple-regressions have lower residuals than the single-regressions, and the higher-dimensionality has lower residuals than the lower-dimensionality.

In addition to improved sensitivity, the higher temporal resolution allowed for a richer analysis of the temporal dynamics of RSNs than has previously been possible; this will be presented separately in a more complete study and is beyond the scope of this paper.

The effect of simultaneously using two multiplexing schemes in the pulse sequence allowed unprecedented numbers of 2D images to be acquired simultaneously in one echo train without physical limitations from specific absorption rate (SAR) or echo time (TE). Slice acceleration by the SIR approach requires far fewer time consuming readout gradient rise times that can dominate EPI readout trains and thus it achieves greater efficiencies given

We anticipate overall improvements in imaging at higher _{2}* and the inverse of the maximum frequency range over the image FOV. Whether these demands are met depends on factors including spatial resolution, gradient ramp times, available read gradient strength and ramp sampling. The MB slice acceleration does not lengthen the echo train or echo spacing and therefore does not worsen EPI image distortion in regions of susceptibility gradients. However, effects of residual aliasing, as with conventional parallel imaging, could introduce spurious activations. We have previously evaluated this and shown it to be non-significant up to MB factors of 4

In principle the readout train length could have been varied for the different SIR acquisitions, resulting in lower bandwidths for the less accelerated data, while matching the distortions of the data with the higher SIR factors. This would yield a relatively higher SNR for the lower SIR factor acquisitions and potentially better fMRI performance. However, this would also come at the expense of longer TRs and potentially longer TEs, which could adversely affect fMRI performance. Further, given the fact that the highest readout bandwidths were

Higher magnetic fields provide higher image SNR and supra-linear increases in functional signals associated with the microvasculature (e.g. _{0}(Gs)^{−1} where δ is the chemical shift and B_{0} the magnetic field strength and Gs is the slice-selective gradient. Therefore, fat suppression in M-EPI at high field is inherently achieved due to the weaker Gs of the refocusing pulses in SIR, without using fat saturation pulses that increase SAR. In SE EPI sequences, a lengthening of the RF refocusing pulse relative to the excitation pulse

A comparison of M-EPI to other fast imaging techniques, particularly 3D acquisitions, can be considered. Since 3D imaging utilizes 2 phase encode directions, reductions in the echo train can be attained by undersampling along the two dimensions, reducing significantly the spatial encoding times. As such, it was shown that 3D acquisition times could be reduced to levels near or faster than what is obtained with 2D multi-slice imaging, greatly improving the performance of 3D functional images over 2D _{1}R_{2}^{1/2} where _{i}_{2}* and consequently the shorter TEs used, there is typically very little dead time to obtain the shifted echoes from the previous excitation. Furthermore, lower SNR due to the use of lower flip angles

There are substantially greater efficiency gains using M-EPI in HARDI as the large time spent on diffusion encoding (typically 60 ms–100 ms) is shared for multiple slices and not repeated for each slice. Consequently, higher spatial and diffusion resolutions that may be prohibited by EPI scan times as long as an hour, could be achieved with M-EPI in much shorter times, tolerable by most subjects. Alternatively, the shorter scanning times for a complete data set can be utilized to acquire several such complete data sets in the time that would have taken to acquire a single EPI based data set. This approach would have the advantage of increasing the probability of achieving a complete data set even if some are rendered useless due to motion. An example of tractography data extracted from a DSI HARDI measurement obtained using the M-EPI sequence is presented as a demonstration of feasibility. A quantitative comparison of gains relative to performance criteria such as resolution of crossing fibers is beyond the scope of this work at this stage and will be pursued and presented separately taking into account trade-offs in SNR, acquisition time and acceptable levels of distortions.

It is known that resting state functional connectivity studies benefit from higher sampling rates to adequately sample undesirable respiration and cardiac effects

For single-regressions, the increase in noise (i.e. decrease in raw SNR per single image) at low-TR is nearly balanced by the statistical advantage of the increased number of samples (timepoints), resulting in peak Z-stats being similar across TRs. This is as predicted by the Bloch equations assuming thermal noise; the reduction in gradient-echo signal, as a function of TR and T_{1}, is well-balanced by the sqrt(N_{timepoints}) increase, over this range of TR values, with a predicted improvement of ∼15% in raw peak Z-stats at 0.4 s, compared with 2.5 s.

However, for multiple-regressions, the short TR data performs significantly better than the longer TR. This is partly because of the reduction in residuals when using all regressors together, a factor which benefits the short TR more than the long TR data, with some group-ICA components (and hence the final regressors) modelling some physiological noise processes with better sampling of these effects at lower TR. It is also partly because the reduction in temporal degrees-of-freedom caused by the use of a large number of regressors impacts more on the data with fewer timepoints. Finally, given that statistical significance in a multiple-regression is driven by a regressor's

Both single-session ICA and the application of dual-regression to map group-ICA results into individual datasets are effectively based around a multiple-regression, and hence benefit from the increase in temporal information shown here to be valuable in low-TR data. However, methods related to single-regression, such as seed-based correlation, will not see this advantage, nor will model-based analysis in a task-FMRI experiment (although artefact removal would probably be improved in both scenarios when using lower-TR data). Nevertheless, even in the ‘worst case’ scenario, the Z-stats (effective CNR) of low-TR data are at least as good as higher-TR data (and, in other scenarios, are considerably better).

The M-EPI approach described here may have a significant advantage over conventional segmented 3D EPI approaches in that the k-space data used to form images results from a single RF excitation. The ultimate determinant of functional CNR in most fMRI applications are temporal fluctuations due to physiologic sources and not the thermal SNR of a single image (e.g.

In conclusion, the novel Multiplexed EPI pulse sequence significantly increased the temporal resolution of whole brain fMRI, and substantially reduced diffusion scan times. As such, this methodology can be used for expanding and enriching the functional and anatomic information obtained from MRI. Further, the reduced scan times may help the clinical acceptance and translation of functional MRI protocols and HARDI neuronal fiber track imaging.

The imaging protocol used for human studies was approved by the institutional review board (IRB) at the University of Minnesota. Ten subjects were scanned using this IRB approved protocol. Each of the subjects provided informed written consent prior to participating in the research. Imaging at 7 Tesla was conducted on a Siemens (Erlangen, Germany) system with a 90 cm bore magnet from Magnex Scientific (currently Agilent Technologies), equipped with a head gradient set (AC84, Siemens) operating at up to 70 mT/m with a slew rate of 333 mT/m/ms. The 7T RF coil consisted of a single transmit channel with 16 receiver coils. Imaging at 3 Tesla used a standard commercial scanner (Siemens Trio) equipped with 40 mT/m gradients with a slew rate of 200 mT/m/ms and utilized the 32 channel head receiver coil system.

The separation of multibanded spatially multiplexed signals was performed as previously described

The RF coil employed for the 7T composed of 16 azimuthally-distributed loops dedicated for signal reception. Despite a lack of distinct coil distribution along the z-axis, at 7T, some z-encoding is nonetheless achieved due to the complex 3-dimensional heterogeneities in RF distribution in the human head. A 4-port driven TEM coil surrounding the receiver array was employed for RF transmission. The 3T array coil employed a distributed coil design along the lines presented by Wiggins et al

The optimization of M-EPI is different at 7T and 3T, given the differences in SAR and parallel imaging performance. At 3T the MB SIR sequence acquiring 2 mm isotropic pixels had the following imaging parameters: 1680 Hz/pixel, matrix size 96×96. All sequences used 6/8 partial Fourier, _{PE}_{0} of respective SIR images into different readout periods. The M-EPI sequence was used to acquire resting state fMRI (R-fMRI) data, and compared to standard EPI at 3T. 3 mm isotropic resolution images were acquired with: 2604 Hz/pixel, matrix size 64×64, _{PE}

The twice-refocused diffusion encoding sequence, first described by Feinberg and Jacob ^{2} and 256 samples in q-space, SIR2 × MB2, 3 mm resolution, total acquisition time 8.5 min, TR/2000 ms and signal bandwidth 2604 Hz/pixel. The TE was 124 ms, increased by 12 ms in comparison to an otherwise identical EPI based sequence. Images were reconstructed in the TrackVis program

At 7T, the SE EPI images were acquired with isotropic 1.5 mm voxels, with a 128×128 matrix, BW/2400 Hz/pixel. Both the SE-EPI and SIR EPI sequences utilized 90° pulse durations of 2.56 ms and 180° pulse duration of 5.12 ms, with the refocusing pulse lengthened to reduce SAR.

Analysis was carried out using FSL (FMRIB's Software Library)

_{1}-weighted, 1×1×1 mm^{3}), and this was aligned to MNI152 standard space, again using FLIRT. The various affine transforms were combined and all 4D resting FMRI datasets resampled into 2×2×2 mm^{3} standard space. Spatial smoothing of 5 mm full-width-half-maximum was applied.

RSNs and structured artifacts in the data were identified using MELODIC (Multivariate Exploratory Linear Optimized Decomposition into Independent Components

From the second multiple regression we can investigate the residuals (which will include thermal noise and any remaining artefacts not modelled in the 100 components), the RSN BOLD fluctuation amplitude (or ‘PE’, i.e., parameter estimate in the multiple-regression), and the Z-stats (basically the amplitude normalised by the residuals). We divide both the residual standard deviation and the amplitude by the original mean signal, so that the first 4 columns in the top row of the boxplots are in units of % signal change.

Although these regressions are able to provide Z-stats for comparing effective CNR across TRs, simple ordinary-least-squares regression does not correct for the true (temporal) degrees-of-freedom (DoF) in the data. This is an important issue if the residuals are temporally smooth (auto-correlated), which will be the case if the residuals contain physiological noise, but not for thermal noise. For the former, the temporal smoothness of the residuals will rise as the TR is reduced. If this is not corrected for, it will yield artificially high Z-stats for the low-TR data. To correct for residual autocorrelation we applied mixture-modeling

In addition to carrying out the regressions as multiple-regressions (i.e., regressing all 100 timeseries into the data simultaneously), we performed separate analyses using single-regressions (i.e., regressing each timeseries into the data one at a time, independently). This results in less specific spatial maps. For example, if one major resting state network is split across 5 of the 100 ICA components, then multiple-regression will return these 5 sub-regions of the network as 5 distinct maps (as each part of the multiple-regression is driven by the unique component of the variance for the relevant timeseries), whereas single-regression will return 5 very similar spatial maps, each of which looks similar to the complete original network. Thus the latter is quite close to the seed-based-correlation

In addition to carrying out the high-dimensional analyses, we also used a set of 10 spatial maps from another study (

For each regression type (multiple or single) and for each RSN map, we have a ‘mixture-model-corrected’ Z-stat image. The two main measures we take from that Z-stat image are the peak value and the sum of all Z-stat values over a threshold (this gave similar results to the supra-threshold

(PDF)

(PDF)

(PDF)

(PDF)

(PDF)

(PDF)

Supplemental text.

(PDF)

The authors would like to acknowledge Rajesh Venkataraman for help with the g-factor calculations, Lawrence Wald and Kawin Setsompop for useful discussions, and to thank the anonymous reviewers for their helpful comments on the manuscript.