Free-breathing abdominal MRI is limited by respiratory motion, which introduces artifacts and inconsistencies in k-space. Conventional approaches often rely on pre-estimated motion fields or external breathing monitors. The proposed method, DoReMe, instead couples reconstruction with motion estimation by representing respiratory motion as compact Gaussian primitives that are jointly refined during iterative reconstruction.
The algorithm alternates between data-consistency updates and motion modeling, integrating Gaussian splatting directly into the reconstruction. In tests on self-navigated free-breathing Cartesian gradient-echo acquisitions from four healthy volunteers, DoReMe reconstructed respiratory motion states in roughly 1.5 minutes. Compared with regularized compressed sensing, it improved visual image quality, particularly in coronal liver views, and better preserved anatomical structures when acquisitions were retrospectively shortened.
These results are presented as a proof of concept. The authors note that Gaussian primitives offer a compact and sufficiently flexible representation of nonrigid respiratory motion with practical runtime, supporting further work toward scanner integration and validation in larger cohorts. No limitations beyond the small volunteer group and need for future testing are discussed in the abstract.