Click here to

Session: MRI for Adaptive Treatment Planning and Delivery [Return to Session]

Real-Time MRI Motion Estimation Through An Unsupervised K-Space-Driven Deformable Registration Network (KS-RegNet)

H Shao*, Y Zhang, UT Southwestern Medical Ctr at Dallas, Dallas, TX


WE-C-TRACK 6-3 (Wednesday, 7/28/2021) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]

Purpose: Real-time MRI is highly desired for intra-treatment target localization to guide radiotherapy. However, imaging acquisition constraints limit the amount of K-space data that can be acquired for real-time MRI. We developed an unsupervised, K-space-driven, end-to-end deformable registration network (KS-RegNet) for real-time MRI motion estimation from fully-sampled priors and under-sampled K-space data.

Methods: KS-RegNet solves deformation-vector-fields (DVFs) to deform prior MRIs (for instance, a fully-sampled setup MRI) to real-time new MRIs, through direct data fidelity matching in K-space. The network takes in complex-valued prior MRIs and real-time under-sampled K-space data, and outputs DVFs for real-time motion tracking. Without access to high-quality ‘gold-standard’ real-time MRIs, the loss function directly measures similarity between K-space transformation of the DVF-morphed prior MRI and the real-time new K-space signal. Inverse-consistency is incorporated into the network design to reduce the DVF solution space and improve its accuracy. A dynamic cardiac MR set with raw K-space data from 16 patients were used to train and evaluate KS-RegNet (12/1/3 for training/validation/testing). Radial sampling masks were generated to down-sample the K-space to 10% - 30% of full sampling. The real-time MRIs estimated by KS-RegNet, respect to each under-sampling mask, were compared against the fully-sampled gold-standard images via metrics including mean-squared-error (MSE) and relative-error (RE).

Results: Under 10% sampling, KS-RegNet generated dynamic MRIs with average MSE and RE of 0.95e-4 and 16.2%, as compared to 1.79e-4 and 21.3% for prior MRIs (without deformable registration), and 7.63e-4 and 47.6% for non-uniform fast-Fourier-transform (NUFFT) images. The corresponding results were 0.83e-4/15.0% and 2.86e-4/28.9% for KS-RegNet and NUFFT by 20% sampling, and 0.76e-4/14.3% and 1.41e-4/20.3% by 30% sampling, respectively. KS-RegNet solved each DVF within 200 ms, meeting clinical demands for real-time imaging.

Conclusion: KS-RegNet allows motion estimation from substantially under-sampled K-space acquisition, paving the way for real-time MRI motion tracking and treatment adaptation.

Funding Support, Disclosures, and Conflict of Interest: The study was supported by funding from the National Institutes of Health (R01CA240808) and from the University of Texas Southwestern Medical Center.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email