Accelerated MR Elastography Using Learned Neural Network Representation

To develop a deep-learning method for achieving fast high-resolution MR elastography from highly undersampled data without the need for high-quality training datasets.

We first reframed the deep neural network representation as a nonlinear extension of the linear subspace model, then used it to represent and reconstruct MRE image repetitions from undersampled k-space data. A multi-level k-space consistency loss was proposed to learn the network weights and mitigate overfitting. To further enhance reconstruction quality, phase-contrast specific magnitude and phase priors were incorporated, namely the similarity of anatomical structures and smoothness of wave-induced harmonic displacement. The network was trained in a self-supervised manner with empirically fixed iterations. Experiments were conducted using one multi-slice spin-echo and two 3D gradient-echo spiral MRE datasets.

Compared to conventional linear subspace-based approaches, the nonlinear network representation method produced superior image reconstruction with suppressed artifacts and noise from a single in-plane spiral arm per MRE repetition, yielding stiffness estimation (e.g., from a 3D GRE sequence at 2 mm resolution within 1 min, R = 10) comparable to that obtained from fully sampled data.

This work demonstrated the feasibility of using deep network representations to model and reconstruct MRE images from highly undersampled data, a nonlinear extension of the subspace-based approach. Future work will include validation on a larger and more diverse cohort of subjects to further assess its robustness and generalizability.