Previously, we presented Wave-Shuffling: a technique that combines random-sampling and the temporal low-rank priors of T2-shuffling with the sinusoidal gradients of Wave-CAIPI to achieve highly-accelerated, time-resolved acquisitions. In this work, we optimize and apply Wave-Shuffling to T2-SPACE and MPRAGE to achieve rapid, 1 mm-isotropic resolution, time-resolved imaging of the brain where we recover a time-series of clinical contrast. We also present preliminary explorations into Joint-Contrast Wave-Shuffling to boost signal-to-noise ratio and reconstruction-conditioning at high accelerations.
Previously, we presented Wave-Shuffling: a technique that combines random-sampling and the temporal low-rank priors of $$$T_2-$$$shuffling1 with the sinusoidal gradients of Wave-CAIPI2 to achieve highly-accelerated, time-resolved acquisitions. In this work, we optimize and apply Wave-Shuffling to $$$T_2-$$$SPACE and MPRAGE to achieve rapid, $$$1 mm-$$$isotropic resolution, time-resolved imaging of the brain where we recover a time-series of clinical contrast. We also present preliminary explorations into Joint-Contrast Wave-Shuffling to boost signal-to-noise ratio and reconstruction-conditioning at high accelerations.
Forward Model. A succinct description of Wave-Shuffling is as follows:$$b=\underbrace{MF_{yz}WF_{x}RE \Phi}_{\triangleq A}\;c,$$where $$$c$$$ represents the underlying temporal coefficients from the Shuffling model, $$$\Phi$$$ is the temporal basis that expands the coefficients to a model-based time-series of images, $$$E$$$ is the ESPIRiT3 operator, $$$R$$$ is the readout $$$(x)$$$ resize operator to a larger FOV to account for the readout alias-spreading as a function of spatial phase-encode positions $$$(y, z)$$$ induced by Wave-CAIPI, $$$F$$$ is the Fourier Transform, $$$W$$$ is the Wave-CAIPI Point Spread Function (PSF), $$$M$$$ is the phase-encode sampling mask, and $$$b$$$ is the acquired data.
Reconstruction. The reconstruction problem is posed as follows:$$c^*=\underset{c}{\mathrm{argmin}}\;\frac{1}{2}||Ac-b||_2^2+\lambda||T(c)||_1,$$where $$$T$$$ is the spatial Wavelet transform and $$$\lambda$$$ is the regularization. This problem is solved with BART4 using FISTA5.
A healthy subject was scanned on a 3T scanner using 32-channel head coil with$$$\;T_2-$$$SPACE and MPRAGE sequences that were modified to play sinusoidal gradients during readout and randomly reorder phase-encode sampling. The matrix size for both cases were $$$256\times256\times256$$$ at $$$1mm$$$ isotropic.
MPRAGE. Fully-sampled (randomly ordered) data with and without sinusoidal readout gradients was acquired using $$$TI=1.2s,TR=2.5s,$$$ flip angle of $$$8^\circ,T_{\text{acquisition}}=11\;\text{minutes}$$$ and an echo-train length of $$$256$$$. The data was retrospectively under sampled to contain $$$100\%,50\%,25\%$$$ of the possible $$$256\times256$$$ phase-encode points, which we denote $$$R1,R2,R4$$$ respectively.
$$$T_2-$$$SPACE. Fully-sampled (randomly ordered) data with $$$89\%\;$$$Partial Fourier (PF) in the$$$\;y\;$$$direction, was acquired with a variable flip angle train6 of echo-train length of$$$\;196$$$, echo spacing of$$$\;4.6ms$$$.$$$TR=3.2s,T_{\text{acquisition}}=16\;\text{minutes}$$$ and sinusoidal gradients. The data was retrospectively under sampled to contain $$$25\%$$$ of the possible $$$228\times256\;$$$phase-encode points ($$$22\%$$$ of $$$256\times256$$$), which we denote $$$R4$$$. PF was accounted for by applying homodyne projection-onto-convex-sets (POCS) to the reconstructed coefficients.
Wave Parameter Optimization. Through simulations, it was observed that larger readout alias-spreading provides better reconstruction conditioning. This spreading is proportional to the maximum amplitude of the sinusoidal gradient, which is maximized by minimizing the number of cycles given gradient-slew and readout duration limitations. In our previous work7, we showed that small cycles create large corkscrew k-space trajectories that result in signal mixing across partition-encoding due to $$$T_1/T_2\;$$$recovery over the acquisition-train, which causes ringing artifacts in standard wave-CAIPI. The Shuffling model resolves signal-recovery over time, allowing for small cycles without artifacts. Consequently, for$$$\;T_2-$$$SPACE, we used$$$\;2\;$$$ cycles with a gradient maximum-amplitude of $$$25mT/m$$$. For MPRAGE with a lower readout bandwidth,$$$\;8\;$$$cycles with a gradient maximum-amplitude of $$$16mT/m$$$ was used for good conditioning while avoiding artifacts of lower-cycle cases with undesirable velocity-encoding.
Given recent developments of joint reconstruction for multi-contrast acquisitions8-11, we present our preliminary explorations of Joint-Contrast Wave-Shuffling. The magnitude coefficients of the above reconstructed $$$R4$$$ MPRAGE and $$$R4\;T_2$$$-SPACE data are registered with respect to each other using FSL12 (since the same subject was scanned on different sessions), and a locally low-rank threshold is applied to promote structural similarities. The result is depicted in Figure 4, where the prior is seen to help remove visual noise with a marginal improvement to normalized root mean squared error with respect to the respective $$$R1$$$ magnitude coefficients.
We plan to explore this direction further by incorporating the different signal evolution of MPRAGE and $$$T_2-$$$SPACE into $$$\Phi$$$ so as to enforce joint-contrast with the data-consistency term. Simulation results of the same are presented in Figure 5. Brain web13 phantom courtesy of Dr. Bo Zhao.
This work was supported in part by NIH research grants: R01MH116173, R01EB020613,R01EB019437, U01EB025162, P41EB015896, and the shared instrumentation grants: S10RR023401, S10RR019307, S10RR019254, S10RR023043.