University of Michigan's PDMR framework enables real-time 3D MRI reconstruction with latent motion tracking

Prospective dynamic MRI reconstruction—critical for applications like MRI-guided radiotherapy—has been hindered by ultra-sparse sampling and strict latency requirements. A new framework from University of Michigan researchers tackles this head-on. PDMR (Prospective Dynamic 3D MRI Reconstruction) learns an efficient, generalizable latent manifold of motion fields offline, then enables rapid online adaptation to new measurements. It parameterizes deformation vector fields on a low-dimensional manifold, drastically reducing the search space during online inference. The method also employs a tri-plane representation for geometry-aware, memory-efficient encoding of 3D motion, allowing the system to handle complex organ movements without excessive computational overhead.

In experiments, PDMR was tested on both XCAT digital phantoms and real in-house abdominal MRI datasets. It consistently delivered high-fidelity, temporally consistent reconstructions across multiple prospective scenarios, including immediate and after-2-minute intervals. Compared to state-of-the-art retrospective and online reconstruction methods, PDMR achieved superior image quality and motion accuracy, all while meeting clinical latency constraints. This work suggests a promising pathway toward ultra-fast, motion-aware prospective MRI reconstruction that could transform MRI-guided radiotherapy and other real-time imaging applications by significantly reducing scan times and improving treatment precision.