1 INTRODUCTION

In contrast to CT, the acquisition of a single MRI contrast can take several minutes, and a whole MRI exam often takes tens of minutes. Long scan times can cause patient discomfort, which occasionally leads to premature termination of an exam. Moreover, the probability of motion artifacts tends to increase with both the acquisition time of each contrast as well as the duration of the entire scan. Consequently, for uncooperative patients, or for time-critical situations such as stroke or head trauma, CT is usually the preferred option—despite the ionizing radiation being particularly unwanted for pediatric patients due to the long-term cancer risks.1, 2 The alternative of sedating an uncooperative patient for an MRI examination is typically accompanied by a significant organizational effort, high costs, and long wait times.3

Consequently, shortening scan times and increasing motion robustness have been long-lasting endeavors since the early days of MRI. Improved gradient hardware has reduced scan times substantially and is now largely limited by peripheral nerve stimulation thresholds.4 On the acquisition side, parallel imaging5-7 and compressed sensing8 marked significant breakthroughs to cut down the scan time. Moreover, motion-robust sampling strategies have been developed, most importantly self-navigated techniques, such as PROPELLER,9 which has been widely adopted in the field, but also other non-Cartesian sampling strategies like radial sampling10 or stack of stars.11 Another approach is single-shot Cartesian imaging, such as single-shot EPI12, 13 or single-shot fast spin echo (SSFSE),14 in which the temporal footprint of a single slice is short enough to freeze the motion in most cases.

A comprehensive clinical brain MRI scan includes several contrasts such as T1-weighted (T1w), T2-weighted (T2w) and T2*-weighted (T2*w), T2w fluid-attenuated inversion recovery (T2-FLAIR), and DWI,15-17 each playing a part in determining the diagnosis. For example, T1w provides excellent gray matter–to–white matter contrast; white-matter hyperintensities can be easily depicted using T2w and T2-FLAIR weighting; T2*w is very sensitive to bleeding; and DWI highlights restricted diffusion of water molecules, which plays a key role in stroke assessment.18 For many pathologies, such as trauma, stroke, edema, inflammation, or hemorrhages, this list of contrasts is sufficient.

With the actual sampling time getting shorter due to advancements in hardware, acquisition and reconstruction, non–acquisition time becomes increasingly important to monitor. This non–acquisition time includes slice planning, sequence download time, and prescan/tuning time. This also includes dummy cycles to prepare a certain magnetization state and wait times for the longitudinal magnetization to recover. Much of this dead time could be avoided if several contrasts were acquired in a single scan. One option is synthetic MRI, in which the conventional MRI contrasts are synthesized in the reconstruction through parametric mapping.19-21 A disadvantage of synthetic MRI is the oversimplification of the underlying model, which fails to capture the complexity of the tissue resulting in differences between the synthetic and the conventional contrasts.22 In addition, the T2*w and DWI contrasts are usually not included in these models. Alternatively, all of the desired contrasts can be acquired in a fast and efficient way by combining the conventional acquisitions into one sequence. Recently, we presented EPIMix,23, 24 in which we combined a single-shot EPI readout with different magnetization preparations, yielding a highly motion-robust sequence acquiring the top-five contrasts T1w, T2w, T2-FLAIR, T2*w, and DWI in a little bit more than a minute. Several clinical studies have acknowledged that the quality of EPIMix is sufficient with only marginally lower diagnostic performance compared with conventional imaging.25-27 The main clinical drawback for EPIMix was the inherent geometric distortions caused by the EPI readout,25-27 which makes diagnosis near the skull base particularly problematic. The EPIMix method was therefore not sufficient to use as the only scan of a brain MRI screening protocol.

In this work, we present a new, generalized, multicontrast framework called NeuroMix (to emphasize brain applications, not solely using EPI readouts), which encompasses the following: An additional software top-level layer to KS Foundation28 that further modularizes the programming code to create and run an MRI sequence such that the acquisition of several contrasts can be set up with minimal programming effort. T1w, T2w, T2-FLAIR, T2*w, and DWI contrasts acquired with single-shot sequences to maintain the motion robustness of EPIMix, but the two most important contrasts for clinical neuroimaging (T2w and T2-FLAIR15) are acquired without EPI distortions using a SSFSE readout14. Optional additional acquisitions are added: isotropic T1w 3D-EPI,29 SWI 3D-EPI,30 and a high-resolution multishot T2w fast spin echo (FSE).

With the addition of FSE contrasts and optional high-resolution acquisitions, we envision NeuroMix to be used as a standalone single-scan brain MRI exam for screening. We describe the acquisition order of each contrast and the corresponding slices to minimize the time when no data are acquired. The performance of NeuroMix is compared against its predecessor, EPIMix, as well as conventional imaging. Finally, we show early clinical results on a pediatric and an adult patient.

2 METHODS 2.1 Programming framework

NeuroMix, as well as its predecessor EPIMix, was programmed for GE Healthcare MRI systems using the vendor-provided programming SDK EPIC together with an abstraction layer called KS Foundation,28 which is publicly available (www.ksfoundationepic.org). The multilayered programming structure of NeuroMix is overviewed in Figure 1, starting with the lowest-level EPIC at the bottom, where specific instructions in the waveform memory are created. This propagates up to the highest level, “ksneuromix.e,” which for the most part is a collection of desired sequence parameters, or sequence recipes. In EPIMix, all layers above the KS Foundation abstraction layer were collapsed into one. In that setting, it was challenging to optimize each specific contrast individually, such as changing the readout from EPI to FSE, or from 2D to 3D. Moreover, changing specific properties of the slice stack, such as orientation and slice thickness, on a per-contrast basis was not possible. Therefore, all of the layers above the KS Foundation layer were implemented for this work to enable NeuroMix. In the “sequence-independent design” layer, we share functionality such as slice orientation, slice acquisition order, TR computations including specific absorption rate (SAR), and gradient heating or k-space properties such as resolution, and phase-encoding plans. The “sequence generator” layer uses the capabilities of all underlying layers to create and serve an MRI sequence to the top layer, including timing calculations and waveform generations, but also real-time execution such as slice looping and data tagging. As of now, we have implemented FSE and EPI sequence generators, in which only EPI has 3D capability. Importantly, the sequence generator layer takes care of inversion preparation as well as including the generation of the inversion sequence and timing computation, linking the main sequence (in our case EPI or FSE) to the inversion sequence, and playing both jointly in real time. Finally, in the top layer, “ksneuromix.e,” all contrasts are designed to be completely self-contained, with no direct dependence on the user interface, using a recipe-like list of input arguments such as TE, TI, TR, FOV, and matrix sizes, but also slice prescription or saturation bands. Crucially, only the top layer is connected to the user interface. Moreover, the user interface has been restricted to only allow the selection of the FOV, number of slices, optional contrasts, and number of averages. This allowed us to optimize NeuroMix for the use case of brain scanning without exposing the complexity of those optimizations, which will be described later, to the user. Some examples of how NeuroMix is structured at a programming level are given in Supporting Information Figure S1.

The layered programming structure of NeuroMix starts with the vendor-provided EPIC SDK at the bottom and “ksneuromix.e” as the top sequence layer. All layers above KS Foundation have been implemented for this work

2.2 Pulse sequences

The waveforms of all subsequences played in NeuroMix are depicted in Figure 2, and all corresponding default sequence parameters are overviewed in Table 1. NeuroMix uses four single-shot 2D-EPI sequences, including an inversion-prepared T1w sequence (T1-FLAIR EPI; Figure 2D), a T2*w gradient echo (T2*w EPI; Figure 2F), a fully sampled, sequential low flip gradient echo (FLEET EPI; Figure 2E) for parallel imaging calibration and a diffusion-weighted spin echo (DW EPI; Figure 2C). Moreover, NeuroMix accommodates a T2w SSFSE (Figure 2H) and an inversion-prepared T2w single-shot FSE (T2-FLAIR SSFSE; Figure 2I). These single-shot sequences belong to the motion-robust core of NeuroMix, where we expect usable images even under severe head motion, similar to EPIMix.23 Optionally, the end user can also include a high-resolution multishot FSE (T2w FSE; Figure 2G) in the NeuroMix scan with almost half the voxel size of T2w SSFSE. All FSE sequences used a linear phase-encoding order, where the refocusing flip angle of all FSEs is modulated using the TRAPS technique.31 Finally, NeuroMix consists of two optional 3D-EPI sequences: a sagittal T1w 3D-EPI (Figure 2B) with isotropic voxels to enable slice reformatting in any plane; and SWI 3D-EPI (Figure 2A) with higher in-plane resolution.

Overview of all NeuroMix sequence diagrams. All diagrams share the same x-axis scaling. The units for the gradient axis Gx, Gy, and Gz are mT/m, and for the B1 axis, µT

TABLE 1. Overview of all default sequence parameters of NeuroMix and EPIMix NeuroMix EPIMix T1-FLAIR EPI FLEET EPI T2*w EPI DWI EPI T2w SSFSE T2-FLAIR SSFSE T2w FSE T1w 3D-EPI SWI 3D-EPI EPIMix T1-FLAIR EPIMix FLEET EPIMix T2* EPIMix T2w/DWI EPIMix T2-FLAIR Scan plane Axial Axial Axial Axial Axial Axial Axial Sagittal Axial Axial Axial Axial Axial Axial TE/effective TE (EPG, FSE only) 17.9 ms 12.2 ms 25 ms 62.3 ms 156.4/93.3 ms 156.8/108.8 ms 125.2/90.0 ms 7.4 ms 23.8 ms 17.0 ms 11.3 ms 20.5/36.2 ms 45.8/102.3 ms 100.3/119.8 ms TR 1300 ms 25 ms 1000 ms 3450 ms 7700 ms 11600 ms 8940 ms 17 ms 52 ms 1300 ms 42.5 ms 930/1220 ms 2410 ms 5500 ms TI 580 ms – – – – 2900 ms – – – 600 ms – – – 2750 ms Partial Fourier 108/180 48/48 156/180 150/180 138/180 132/136 288/288 168/192 312/312 108/180 108/180 120/180 108/180; 180/180 120/180; 120/180 FOV 240 × 240 mm² 240 × 240 mm² 240 × 240 mm² 240 × 240 mm² 240 × 180 mm² 240 × 180 mm² 240 × 180 mm² 240 × 240 mm² 240 × 240 mm² 240 × 240 mm² 240 × 240 mm² 240 × 240 mm² 240 × 240 mm² 240 × 240 mm² Matrix size 180 × 180 180 × 48 180 × 180 180 × 180 240 × 180 220 × 136 384 × 288 192 × 192 312 × 312 180 × 180 180 × 180 180 × 180 180 × 180 180 × 180 Voxel size 1.3 × 1.3 mm² 1.3 × 5.0 mm² 1.3 × 1.3 mm² 1.3 × 1.3 mm² 1.0 × 1.0 mm² 1.1 × 1.3 mm² 0.6 × 0.6 mm² 1.2 × 1.2 mm² 0.8 × 0.8 mm² 1.3 × 1.3 mm² 1.3 × 1.3 mm² 1.3 × 1.3 mm² 1.3 × 1.3 mm² 1.3 × 1.3 mm² Number of slices 36 36 36 36 36 36 36 152 84 36 36 36 36 36 Slice thickness 4.0 mm 4.0 mm 4.0 mm 4.0 mm 4.0 mm 4.0 mm 4.0 mm 1.2 mm 2.0 mm 4.0 mm 4.0 mm 4.0 mm 4.0 mm 4.0 mm exc flip 90° 15° 30° 90° 90° 90° 90° 22° 17° 90° 5° 18° 90° 90° Number of shots 1 3 1 1 1 1 5 24 12 1 3 1 1 1 Acceleration factor 3 3 3 3 2 2 2 3 2 3 3 3 3 3 ACS lines – – – – 4 4 6 8 8 – – – – – ESP/effective ESP (EPI only) 800/267 μs 800/267 μs 800/267 μs 800/267 μs 5792 μs 4544 μs 8192 μs 1424/59 μs 1576/131 μs 800/267 μs 800/267 μs 800/267 μs 800/267 μs 800/267 μs Echo train length 36 16 52 50 73 70 30 7 26 36 36 40/60 40/60 36/36 Readout gradient amplitude 34.58 mT/m 34.58 mT/m 34.58 mT/m 34.58 mT/m 8.16 mT/m 12.23 mT/m 8.16 mT/m 16.31 mT/m 24.32 mT/m 34.58 mT/m 34.58 mT/m 34.58 mT/m 34.58 mT/m 34.58 mT/m Dummy TRs 1 0 0 1 0 0 0 3 5 1 0 0 1 0 Duration 16.8 s 5.4 s 2 s 33.5 s 15.4 s 23.3 s 44.7 s 22.5 s 28.3 s 18.5 s 9.8 s 4.1 s 28.9 s 11.0 s Abbreviations: ACS, autocalibration; EPG, extended phase graph; ESP, echo spacing; FLEET, fully sampled, sequential low flip gradient echo; s, seconds.

The FOV coverage and acquisition order for each contrast is depicted in Figure 3A. NeuroMix is prescribed in the axial scan plane with frequency-encoding direction left–right (Figure 3A, green box), with slices typically aligned with the anterior commissure–posterior commissure line. Importantly, the end user only prescribes a square FOV for the slice stack of the 2D-EPI sequences. For the SSFSE and FSE sequences involved, NeuroMix swaps the phase/frequency encoding directions and reduces the left–right (phase-encoding) FOV to 75%. Furthermore, the resolution in the slice direction is doubled for SWI 3D-EPI, and an inferior saturation band is automatically added to suppress in-flow artifacts. Finally, the scan plane of T1w 3D-EPI is automatically rotated from the axial to the sagittal plane, and the left–right FOV is reduced to 75%, similar to the FSE sequences. In this way, the frequency-encoding direction of T1w 3D-EPI is turned into the superior–inferior direction, which helps to suppress flow artifacts. Moreover, we can use a non-slice-selective water excitation pulse for improved gray matter/white matter contrast stemming from reduced magnetization-transfer effects.29

(A) Illustration of the FOVs, order of the contrasts, the order of the slices. The scan times shown refer to the default protocol given in Table 1. The green box indicates the FOV that the radiographer has to prescribe. Everything else is set up automatically. (B) Moving averages for the squared B1, which is proportional to the specific absorption rate (SAR) for one NeuroMix sequence run using the default protocol

Designing the acquisition order of NeuroMix is a delicate process, particularly the challenge of avoiding unnecessary wait times for magnetization recovery. The two 3D sequences are acquired in steady state, and few dummy TRs are needed because the acquisition starts with the high frequencies of k-space. Thereafter, the DWI-EPI sequence is played, which necessarily requires one dummy TR for accurate apparent diffusion coefficient (ADC) values. The user also has the option to acquire an additional b = 0 volume with opposite blip polarity for distortion correction purposes.32 In addition, the dummy TRs of T1w 3D-EPI, SWI 3D-EPI, and DW EPI are used to acquire Nyquist ghost-correction data, in which the data of DWI EPI are used for all other 2D-EPI sequences.

Inversion-prepared contrasts benefit from thicker slices of the inversion sequence compared with the main imaging sequence to compensate for slice profile imperfections and CSF inflow. Inflow can cause false hyperintense signals, especially in T2-FLAIR, which could mimic or obscure pathology. Therefore, we have chosen to double the inversion slice thickness and to acquire T1-FLAIR EPI and T2-FLAIR SSFSE in two passes, where one pass covers the odd (slice groups A + C in Figure 3A) and one the even slices (slice groups A + C in Figure 3A).

After the DWI-EPI block, the first pass of T1-FLAIR EPI is played, including one dummy TR. Thereafter, FLEET EPI and T2*w EPI data are acquired without dummy TRs. Note that due to the low flip angles of FLEET EPI and T2*w EPI, the impact on the longitudinal magnetization is small. Moreover, during FLEET EPI, the magnetization can recover just enough for a reasonably attenuated signal of CSF in the T2*w EPI. If the user has chosen to acquire an additional b = 0 volume with opposite blip polarity, an additional FLEET volume with the same blip polarity is acquired after T2*w EPI to avoid ghosting due to distortion mismatch of the calibration data and the undersampled data.

A T2-FLAIR sequence consists of an inversion block and the main readout block, separated by the TI. For the following, we refer to them combined as a T2-FLAIR block. For the default parameter choices (Table 1), two T2-FLAIR blocks are needed to accommodate all slices of one pass. The crucial challenge, however, was to achieve a homogenous CSF suppression across the slice stack for T2-FLAIR. We assign all slices of a T2-FLAIR block to a group, resulting altogether in four slice groups: A to D (Figure 3A).

For T2w SSFSE, we change from odd to even slices and play slice group D, which contains fairly well-recovered magnetization, only saturated by the widened inversion pulse of the T1-FLAIR EPI sequence. Next, T2-FLAIR SSFSE is played for slice group B, which was also last saturated by the inversion of T1-FLAIR EPI. Next, the second T2-FLAIR block of the first pass is played on slice group D, which was last saturated by the T2w SSFSE sequence. Finally, T2w SSFSE is acquired for slice group B, in which the magnetization has had time to recover during the T2-FLAIR block. The reasoning behind this slice order is that the time between T1-FLAIR EPI and the first T2-FLAIR block on slice group B is approximately the same as the time between T2w SSFSE on group D and the second T2-FLAIR block on slice group D. Consequently, the magnetization state of the two T2-FLAIR blocks is similar. We found the corresponding TI that achieves homogeneous suppression of the CSF to be 2900 ms. The corresponding in vivo experiments with multiple TIs and a comparison against a conventional slice order are shown in Section 3. The second pass is acquired similarly to the first one, with the slice groups adjusted as shown in Figure 3A. If the user chooses more than 36 slices, the number of T2-FLAIR blocks per pass, termed N, needs to be larger than 2 to accommodate all slices. In this case, we generalize the acquisition as follows: T2w SSFSE is split into N slice groups per pass, where N−1 slice groups are played before T2-FLAIR SSFSE and the remaining one afterward. In this case, the TI was adjusted to 3050 ms. For the interested reader, an audio file of NeuroMix running the default protocol (Table 1) is provided in Supporting Information Audio S1.

The optional multishot T2w FSE is played on all slice groups at the end of the second pass. Magnitude inconsistencies between the first and all subsequent shots are problematic, especially for CSF due to the long T1 relaxation. Two mitigation options were explored in this work: a dummy TR and second an incoherent sampling scheme to disperse the formation of visible artifacts due to shot-to-shot magnitude differences.

In this work, EPIMix and conventional imaging protocols are used as references. The imaging parameters of EPIMix are also provided in Table 1. For EPIMix, all contrasts were acquired in two passes. Moreover, EPIMix acquired dual EPI readouts for the T2w/DWI and T2-FLAIR contrasts to improve SNR. More details of the EPIMix implementation can be found in Skare et al23 and Sprenger et al.24

At 3 T, SAR needs to be taken into account for rapid FSE sequences, resulting in additional dead time after each sequence playout to limit the average SAR.33 The SAR limits are defined in most countries by the International Electrotechnical Commission, defining maximum values for 6-min-average and 10-s-average SAR, the latter which shall not be higher than twice of the former.34 In this work, the TRAPS technique was used to reduce the SAR of the FSE sequences.31 Moreover, we took advantage of the fact that the durations of all subsequences are known from the beginning and are significantly shorter than 6 min. Consequently, each sequence (contrast) is evaluated with twice the SAR budget than normal, by obeying only the 10-s-average SAR limit. Thereafter, all sequences of NeuroMix were jointly evaluated against the 6-min-average SAR limit, resulting in no SAR-related dead time at 3 T. Figure 3B depicts moving averages of the squared B1 of a complete NeuroMix sequence run, indicating the relative SAR levels of the different subsequences.

All imaging was performed on a 3T MR system (Signa Premier; GE Healthcare, Milwaukee, WI) using a 48-channel head coil from the same vendor. In vivo data were acquired on 2 volunteers and 2 patients in accordance with the institutional review board policy, and informed consent was obtained. The movement of the head of the pediatric patient was recorded using a markerless motion tracker (Tracoline TCL3.1m; research version provided by TracInnovations, Ballerup, Denmark),35 but no prospective updates of the slices were performed.

2.3 Reconstruction

The image reconstruction was implemented in MATLAB (MathWorks, Natick, MA) using a vendor-provided SDK called Orchestra. Linear Nyquist ghost correction was applied to all EPI data using the calibration lines acquired in the dummy TRs of SWI 3D-EPI, T1w EPI, and DWI EPI, in which the latter was also used for all 2D-EPI sequences. This was followed by ramp sampling correction of the EPI data to Cartesian coordinates in the frequency-encoding direction. Parallel imaging reconstruction was performed for the 2D-EPI contrasts using GRAPPA with FLEET EPI as a motion robust external calibration.7 All other contrasts included autocalibration lines, and the parallel imaging reconstruction was done using autocalibrating reconstruction for Cartesian imaging.36 When applicable, partial Fourier reconstruction was done using projection onto convex sets. For the coil combination, sum of squares was used, except for SWI EPI, in which adaptive coil combine was used.37 Intracontrast 2D motion correction was done using a rigid model for T1-FLAIR EPI and an affine model for DW EPI, which also corrected for the zero-order and linear eddy currents induced by the diffusion gradients. The diffusion processing included averaging of all b = 0 volumes and all DWI volumes, and their log ratio was used to compute the isotropic ADC. The vessel contrast of SWI 3D-EPI was improved using SWI processing, combining magnitude and phase.38 The extra b = 0 volume with reverse blip polarity was reconstructed to later allow for distortion correction of all 2D-EPI contrasts.32 However, in favor of increased motion robustness and shorter image-reconstruction time, distortion correction was not enabled in the clinical NeuroMix reconstruction pipeline, nor for the volunteer experiments. However, distortion correction of existing NeuroMix data at a later point will enable a better voxel-wise multicontrast analysis.

3 RESULTS

Figure 4 compares all EPIMix contrasts (Figure 4A) with the corresponding single-shot contrasts of NeuroMix. For a fair comparison, and contrary to Table 1, NeuroMix was acquired without optional contrasts, with four diffusion directions (tetrahedral) and no reverse polarity volumes, resulting in an overall scan duration of 1:20 min for NeuroMix versus 1:12 min for EPIMix. Unsurprisingly, the biggest differences can be appreciated for the T2w and T2-FLAIR contrasts, in which NeuroMix uses an SSFSE instead of an EPI readout. Although the EPIMix T2-FLAIR sequence provided slightly stronger contrasts, SSFSE T2-FLAIR significantly outperformed EPIMix in terms of image sharpness, SNR, and absence of geometric distortions. The T2w-SSFSE sequence was acquired with 1-mm in-plane resolution compared with 1.3 mm for EPIMix (Table 1), resulting in a clear sharpness gain and an improved gray matter/white matter delineation. The other three contrasts (T1-FLAIR, DWI, and T2*w) are very similar. The only obvious difference is the increased SNR in NeuroMix T2*w compared with EPIMix T2*w due to the higher flip angle used. For the interested reader, we provide a set of dicoms including all contrasts of NeuroMix and EPIMix in Supporting Information Data S1.