LSST: Learned Single-Shot Trajectory and Reconstruction Network for MR Imaging

TL;DR

The paper tackles speeding up 2D single-shot MR imaging by learning both the k-space trajectory and the reconstruction network under hardware limits and an unknown $T_2$-blur. It introduces Learned Single-Shot Trajectory (LSST), a differentiable framework that models the forward non-uniform FFT and a blur modulation $\mathcal{B}$, while enforcing $G_{ ext{max}}$ and $S_{ ext{max}}$ constraints in trajectory optimization. The reconstruction is a three-stage process: a density-compensation corrected SENSE step, followed by a Direct-Inversion Network (or a joint $\mathcal{D}_{\theta,\mathbf{k}}$), trained with a task loss plus a constraint loss. Experiments on the fastMRI knee dataset show that LSST outperforms CSTV and PILOT in PSNR and SSIM at 8x and 16x accelerations and is rated higher by a radiologist for sharper ACL fibers, indicating potential clinical impact for accelerated single-shot MR imaging.

Abstract

Single-shot magnetic resonance (MR) imaging acquires the entire k-space data in a single shot and it has various applications in whole-body imaging. However, the long acquisition time for the entire k-space in single-shot fast spin echo (SSFSE) MR imaging poses a challenge, as it introduces T2-blur in the acquired images. This study aims to enhance the reconstruction quality of SSFSE MR images by (a) optimizing the trajectory for measuring the k-space, (b) acquiring fewer samples to speed up the acquisition process, and (c) reducing the impact of T2-blur. The proposed method adheres to physics constraints due to maximum gradient strength and slew-rate available while optimizing the trajectory within an end-to-end learning framework. Experiments were conducted on publicly available fastMRI multichannel dataset with 8-fold and 16-fold acceleration factors. An experienced radiologist's evaluation on a five-point Likert scale indicates improvements in the reconstruction quality as the ACL fibers are sharper than comparative methods.

Figures (8)

• Figure 1: The training (a) and inference (b) pipeline of the proposed joint optimization framework. Here purple blocks contain trainable parameters.
• Figure 2: This figure (a) shows an example of a random set of points acquired to achieve an acceleration factor R=8. (b) shows corresponding TSP-based trajectory that does not satisfy physics constraints on maximum gradient strength and slew-rate. (c) shows optimized trajectory using the proposed LSST framework that has smooth curvature as seen from zoomed portions in range [$2\pi/3 ,\pi$] that optimized trajectory has smooth curvature compared to (b) since it satisfy physics constraints.
• Figure 3: This figure contrasts analytic and iterative reconstructions used as network inputs. A fully sampled test image shown in (a) when undersampled at 8x and reconstructed using adjoint NUFFT with DCF results in (b). (c) and (d) depict iterative reconstructions via the SENSE algorithm, without and with blur in measurements, respectively. The blurred measurements lower the PSNR from 33.83 dB to 19.36 dB, highlighting the challenge of single-shot reconstruction.
• Figure 4: A comparison of experimental results at 8x acceleration on a test slice. The CSTV output (b) exhibits noticeable artifacts due to high 8x acceleration and the use of a 1D Cartesian sampling mask. The zoomed area shows that ACL fibers are sharper than PILOT (c) whereas meniscus is sharp everywhere in the proposed method (d).
• Figure 5: This figure shows the benefit of optimizing the trajectory rather than using traditional 1D Cartesian sampling to acquire the data. The ground truth image (a) was undersampled using mask in (b) and trajectory in (c) that resulted in undersampled k-space data which was reconstructed using regridding reconstruction (regrid. recon.) $\widetilde{\mathbf x}$ to result in (d) and (e),respectively. Here, trajectory optimization also incorporated the T2-blur whereas Cartesian sampling experiment did not have additional blur. We note that (d) and (e) are not the final reconstruction but zero-filled reconstructions ($\widetilde{\mathbf x}$) as shown in Fig 1(a). Since, (e) has many details as compared to (d), it can act as an improved input to the SENSE reconstruction block and subsequent neural network block.