GAN-based iterative motion estimation in HASTE MRI

TL;DR

This work tackles motion artifacts in MRI, especially within fast HASTE acquisitions, by coupling motion estimation with data-consistent reconstruction. It introduces an unrolled iterative network that uses GAN predictions to estimate deformation fields while enforcing data fidelity through a tractable forward model $\mathcal{A}_{t,c}(U,s) = M_t \mathcal{F}[S_c U_t[s]]$, and a deformation basis together with time interpolation to keep computations feasible. A cascade training strategy and a specialized loss ensure stable learning despite non-unique references, enabling reliable motion correction without trajectory redundancies. Experimental results on brain and abdominal HASTE MRIs show improved PSNR/SSIM and detail preservation, suggesting clinical viability for motion-robust MRI without requiring extra acquisition requirements.

Abstract

Magnetic Resonance Imaging allows high resolution data acquisition with the downside of motion sensitivity due to relatively long acquisition times. Even during the acquisition of a single 2D slice, motion can severely corrupt the image. Retrospective motion correction strategies do not interfere during acquisition time but operate on the motion affected data. Known methods suited to this scenario are compressed sensing (CS), generative adversarial networks (GANs), and explicit motion estimation. In this paper we propose an iterative approach which uses GAN predictions for motion estimation. The motion estimates allow to provide data consistent reconstructions and can improve reconstruction quality and reliability. With this approach, a clinical application of motion estimation is feasible without any further requirements on the acquisition trajectory i.e. no temporal redundancy is needed. We evaluate our proposed supervised network on motion corrupted HASTE acquisitions of brain and abdomen.

Figures (6)

• Figure 1: Results of $U^{-1}[U[s]]$ by using different versions for discretized transforms. The bilinear interpolation smooths out details. The sinc interpolation with 7 times 7 patches is visually almost artefact free.
• Figure 2: $\text{V-netD}^{^{D,C}}_\Theta$ with parameters $\Theta$, depth $D=4$ and $C=12$ channels for images of resolution $N\times N=192 \times 192$.
• Figure 3: The time dependence function $f(\cdot, K_j)$ for a few instances $K_j$ of the set of parameter distributions.
• Figure 4: Test Error Loss performance of proposed network over iteration depth. At iteration number zero the loss of an identity deformation is shown.
• Figure 5: Results on Brain HASTE MRIs $N^\mathrm{coils}=4$. All reconstructions have a resolution of $192\times192$.