Click here to

Session: Machine Intelligence Efficacy and Quality II [Return to Session]

Comprehensive Evaluation of a Real-Time 3D MR Imaging Technique Using a Deformation-Driven Deep Convolutional Neural Network (KS-RegNet)

H Shao1*, T Li2, M Dohopolski1, J Wang1, J Cai2, J Tan1, K Wang1, Y Zhang1, (1) The University of Texas Southwestern Medical Center, Dallas, TX, (2) The Hong Kong Polytechnic University, Hong Kong, HK

Presentations

MO-B-BRC-2 (Monday, 7/11/2022) 8:30 AM - 9:30 AM [Eastern Time (GMT-4)]

Ballroom C

Purpose: Real-time 3D MRI is highly desired for MR-guided radiotherapy but challenged by the limited signal acquisition speed that leads to severely under-sampled k-space data. We developed a deep learning-based, k-space-driven deformable registration network (KS-RegNet) for real-time 3D MRI (within 0.5s temporal resolution).

Methods: KS-RegNet is an end-to-end, unsupervised network that performs deformable registrations between an input fully-sampled prior MRI and on-board MRIs acquired as severely-under-sampled k-space data, to generate high-quality on-board MRIs for target localization. To train KS-RegNet to learn the artifacts from limited k-space sampling, the k-space data of the prior MRI was purposely down-sampled using the same readout trajectory as the real-time MRI. The fully-sampled prior MRI, the real-time k-space data, and the under-sampled k-space data of the prior MRI, were combined and fed into KS-RegNet to predict deformation vector fields for motion estimation and real-time MRI generation. KS-RegNet combines k-space-based loss and an inverse-consistency strategy to further address the under-sampling issue. It also serves a generic framework applicable to different k-space trajectories. KS-RegNet was evaluated using both cardiac and abdominal MR datasets. We evaluated the network for various under-sampling ratios and k-space trajectories, including stacked 2D golden-angle radial trajectories and 3D golden-mean Koosh-ball trajectories.

Results: KS-RegNet outperformed both conventional deformable registration algorithms (Elastix) and other neural network structure variants (via ablative studies). The average (± s.d.) DICE coefficients of KS-RegNet for real-time ventricle localization on the cardiac dataset were 0.884±0.025, 0.889±0.024, and 0.894±0.022, based on stacked 2D golden-angle trajectories of 5-, 9- and 13- spokes/slice, respectively. Using 3D golden-mean Koosh-ball trajectories, the average DICE coefficients for liver tumor localization on the abdominal dataset were 0.720±0.152 and 0.736±0.125 for 100 and 200 spokes/volume, respectively

Conclusion: KS-RegNet allows real-time MR imaging and motion estimation with <0.5s latency, enabling potential real-time MR-guided target localization, beam gating, and MLC tracking.

Funding Support, Disclosures, and Conflict of Interest: The study was supported by funding from the National Institutes of Health (R01CA240808, R01CA258987) and a seed grant from the Department of Radiation Oncology at the University of Texas Southwestern Medical Center.

Keywords

MRI, Registration, Localization

Taxonomy

IM/TH- MRI in Radiation Therapy: MRI/Linear accelerator combined- IGRT and tracking

Contact Email

Share: