A Dynamic Reconstruction and Motion Estimation Framework for Gaussian Representation-Based Time-Resolved Volumetric MR Imaging (DREME-GSMR)
Abstract
Purpose
Time-resolved volumetric MRI reconstructed from minimal k-space samples is critical for motion-adaptive radiotherapy to capture real-time deformable motion. We propose a Gaussian representation-based one-shot learning framework that models patient anatomy and motion as 3D Gaussians (DREME-GSMR). DREME-GSMR combines pre-treatment dynamic MRI reconstruction with intra-treatment real-time motion/image estimation, without relying on prior anatomical/motion models.
Methods
DREME-GSMR consists of three components: (i) a dense 3D Gaussian set to reconstruct a reference-frame volumetric MRI, (ii) a second Gaussian set that models motion via coarse-to-fine motion-basis components (MBCs) to capture voxelwise motion patterns, and (iii) a dual-path MLP-CNN motion encoder that estimates temporal MBC coefficients directly from time-resolved multi-coil k-space data. The MLP path captures global bulk motion based on the k-space center, while the CNN path models finer dynamics from k-space data and enables motion augmentation for robust real-time motion/image estimation. The coefficient-weighted MBCs are combined into deformation-vector-fields (DVFs) to warp the reference-frame MRI into time-resolved volumes. DREME-GSMR was evaluated using XCAT phantom simulations, physical phantom measurements, and clinical data from a 1.5T MR-LINAC (6 volunteers and 20 patients), all acquired with a stack-of-stars trajectory. Metrics including the image structural-similarity-index-measure(SSIM), tumor/target center-of-mass-error(COME), and tumor/target DICE coefficient(DSC) were evaluated.
Results
DREME-GSMR reconstructs volumetric MRIs (94x368x368) of a ~400ms temporal resolution, with an inference time of ~10ms/volume. For XCAT, DREME-GSMR achieved mean(s.d.) SSIM, tumor COME, and DSC of 0.92(0.01)/0.91(0.02), 0.50(0.15)/0.65(0.19) mm, and 0.92(0.02)/0.92(0.03) for dynamic reconstruction/real-time imaging, compared with 0.84(0.02)/0.82(0.02), 0.66(0.31)/0.74(0.27) mm, and 0.92(0.02)/0.91(0.02) for an implicit neural representation-based variant (DREME-MR). Compared with DREME-MR, DREME-GSMR reduced training time by 30%. For the physical phantom, the mean target COME of dynamic/real-time imaging was 1.19(0.94)/1.40(1.15) mm, and for volunteers/patients the mean liver COME of real-time imaging was 1.04(0.70) mm by DREME-GSMR.
Conclusion
DREME-GSMR enables accurate, efficient time-resolved volumetric MRI from ultra-sparse k-space, supporting real-time motion-adaptive radiotherapy.