Title: Achieving Dynamic 3D MRI Reconstruction in Real-Time Through Latent-Space Motion Tracking of Single Measurements

Abstract:

In clinical settings such as MRI-guided radiotherapy, the ability to perform prospective reconstruction is essential. This process requires the simultaneous delivery of precise image reconstruction and rapid motion estimation based on data currently being acquired. Despite its importance, prospective reconstruction presents significant hurdles, primarily driven by the need to operate under conditions of ultra-sparse sampling and strict latency constraints.

To address these challenges, this study introduces PDMR, a novel framework for Prospective Dynamic 3D MRI Reconstruction that utilizes latent-space motion tracking. The foundational concept behind PDMR involves pre-learning an efficient and broadly applicable latent manifold of motion fields during an offline phase. This allows for swift online adaptation, facilitating effective prospective reconstruction.

Technically, the method represents deformation vector fields (DVFs) within a low-dimensional manifold. This approach significantly narrows the search space, thereby accelerating the online adaptation process. Furthermore, the framework incorporates a tri-plane representation to ensure that the encoding of 3D motion is both memory-efficient and aware of underlying geometry.

The efficacy of PDMR was validated through experiments involving XCAT digital phantoms as well as in-house abdominal MRI datasets. The results indicate that PDMR delivers high-fidelity, temporally consistent reconstructions across various prospective scenarios, including Immediate and After-2min conditions. In these tests, the proposed method surpassed existing state-of-the-art retrospective and online techniques. These findings highlight PDMR as a viable and promising strategy for implementing ultra-fast, motion-aware prospective MRI reconstruction in clinical environments.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC