2495

Superresolution MRSI: a desirable acquisition trajectory.

Claudiu Schirda¹

¹Radiology, University of Pittsburgh, Pittsburgh, PA, United States

Synopsis

The smoothly varying waveform and sampling that starts at k=0 and the innate property of rewinding periodically to k=0, makes the rosette trajectory achieve the same spatial resolution and spectral bandwidth as other trajectories (EPSI, SSI, CONCEPT) using less than twice gradient strength and slew rate. This makes it an ideal candidate for superresolution MRSI and ultra-high field SI acquisitions.

INTRODUCTION

Interest in achieving high resolution spectroscopic imaging (SI) is not necessarily new and research in superresolution¹ and ultra-high field (UHF, >=7T) MRSI seem to have produced developments at an increased rate in recent years. Approaches to high resolution MRSI include using readout gradients that simultaneously sample the spectral information and one (echo planar EPSI²) or two spatial dimensions (spiral-SSI³, rosette-RSI⁴), concentric circles-CONCEPT⁵), or conventional CSI with very short TR⁶ and some form of sensitivity encoding. Among mentioned trajectories, each has its appeal: EPSI samples data on a cartesian grid with constant sampling uniformity (η=1) and is the most established, SSI and RSI start acquisition at k=0, making them more robust to motion⁷, with SSI being able to achieve up to η=0.97 sampling uniformity, while CONCEPT allows for easy design of the sampling density function, by choosing number of rings accordingly⁸.

Acquiring data at high (3T) or UHF, is appealing for increasing the SNR. However, for gradient readout trajectories, this setting brings the additional challenge of having to reach an increased spectral bandwidth⁹ (SW), demanding the use of increased gradient strength (Gmax) and slew-rate (SRmax) which, if not addressed, can result in decreased data quality¹⁰. Methods such as SPICE¹¹ (based on EPSI) or dual-density SSI acquisition¹², sample the extended k-space to reach for an in-plane resolution (IPR) of 2.5-3 mm and, using low-rank denoising, achieve super-resolution¹. However, this is done by pushing the scanner gradient system close to its limits. Temporally interleaving the trajectories can typically alleviate the hardware demands but results in a proportional increase in total scan time and it may not always be an option. Thus, the question is, what acquisition trajectory would be more desirable for superresolution MRSI?

METHODS

To demonstrate the rosette trajectory (k=kmax*sin(2π*f1*t)*exp(1i*2π*f2*t)) capabilities at high Nx, we use Hadamard-RSI¹³ to scan a BRAINO phantom on a Siemens MR/PET scanner (2.9T) with a 20-channel head-neck coil (12-channels used), with TR=1s, WET water suppression, 4-slices/averages, 10mm slice thickness, 2mm gap, TE=56.2/45/61.8/50.6ms (inferior to superior), FOV=200mm, and Nx=100/80/40/20 for IPR=2/2.5/5/10 mm and SW=4.95ppm (610Hz), using Nsh=180/120/60/20 trajectories, rotated by 2π/Nsh to fully sample k-t space, with a maximum gradient strength Gmax=11.2/9/4.5/5 mT/m and SRmax=30/25/15/25 mT/m/ms (Figure 1). A 1.6min 20x20x4 Hadamard-RSI acquisition was first validated against 4 individual 20x20 CSI slices with elliptical encoding at corresponding TEs, 4.2 min each: Bland-Altman 95% confidence for the (RSI-CSI) difference was within 8.6% of mean. The Nx=100 acquisition was reconstructed on a 128x128x4x512 matrix and processed with LCModel¹⁴ in the 4.2-0.2ppm range. For all data sets, the center k-space corresponding to Nx=20 (kmax=.5/cm) was reconstructed on a 32x32x4x512 grid, and average Crammer-Rao lower bounds (CRLB) for tNAA/tCr/tCho were compared, after being normalized for acquisition time.

A single, fully sampled k-t space slice, FOV=240mm, acquisition matrix 96x96 reconstructed to 128x128, SW=575Hz (4.67ppm), TR=1.5s (matching the Nx, SW and TR of the dual-density SSI in Ref. 12) was also acquired in 3.75mins.

No eddy current or trajectory correction or low-rank denoising was used in processing any of the data.

RESULTS and DISCUSSION

CRLB maps for tNAA for the 12mins 100x100x4 acquisition are shown in Figure 2, with a sample spectrum in Figure 3. Average CRLBs for tNAA/tCr/tCho, were 8.3/9.4/10.2. While these results highlight the high SNR sensitivity of the RSI acquisition, obviously denoising would be needed for in-vivo acquisitions (because of much greater B0 inhomogeneities, etc). The time spent for Nx=100/80/40/20 in the Nx=20 center of k-space can be estimated by multiplying the total acquisition time by 2/π*asin(20/Nx) and is 13/16/33/100%, thus 1.5/1.3/1.3/1.3 mins. The average of the tNAA/tCr/tCho CRLBs normalized to time, relative to the Nx=20 acquisition are 1.06/1.0/0.98 for the Nx=100/80/40 acquisitions. Figure4 shows the common region used for comparison, with sample spectra in Figure5. While the normalized CRLBs show no difference down to IPR=2.5mm (up to Nx=80), the average 6% increase in CRLB for the IPR=2mm (Nx=100) acquisition, suggests the detrimental effect of increased gradient strength is starting to become observable.

The Nx=96, SW=575Hz was acquired with Gmax=8.5mT/m and SRmax=25mT/m/ms in half the time of a dual-density SSI acquisition¹². These numbers are almost one third of the gradient strength of the 22.4mT/m and one fifth of the slew-rate SRmax=125mT/m/ms used for SSI. While, as implemented, only 17% of the data is collected in the Nx=16 center k-space, sampling uniformity is η=0.9 across the entire k-space and 100% of data is used as opposed to having to discarding 41% of time for ramp-down and rewinding (23% SNR penalty).

CONCLUSION

The much lower gradient demands of the rosette gradient readout waveform makes it a very desirable acquisition trajectory for superresolution MRSI and for ultra-high field.