Dual optimization of Reconstruction and Motion estimation for self-navigated, free-breathing abdominal MRI using Gaussian Splatting.
DoReMe alternates between multi-coil image reconstruction and nonrigid respiratory motion estimation. Compact Gaussian primitives share spatial parameters across states, with state-specific displacement vectors describing motion. Motion-aligned total variation regularization shares information between states while accounting for anatomical displacement, and both images and motion fields are refined throughout reconstruction.
Requires an NVIDIA GPU, CUDA-enabled PyTorch, NumPy, and a compatible CUDA toolkit and C++ compiler. From the repository root, install the B-spline extension and run the dummy example:
cd bspline_gradient
python -m pip install --no-build-isolation .
cd ..
python run.pyrun.py checks the pipeline with random tensors; no dataset is required. It does not reproduce the paper's experimental results, and the study data cannot be made publicly available.
Settings are at the top of run.py. Motion fitting has many configurations to explore: Gaussian covariance, component and neighbor counts, voxel sampling, fitting schedules, reference-based or reference-free alignment, and smoothness or Jacobian penalties. The defaults are a starting point, not a universal configuration.
Memory usage: The algorithm is not optimized for memory. Large volumes, many motion states, or large Gaussian models can require substantial GPU memory.
- run.py: dummy example entry point.
- DoReMe/
- DoReMe.py: main reconstruction loop.
- modules.py: Gaussian motion model, covariance options, and displacement constraints.
- computational_functions.py: TV regularization, motion fitting, field inversion, and Jacobian penalties.
- mri_operators.py: multi-coil MRI encoding, k-space/readout conversion, and image warping.
- helper_functions.py: signal scaling, coordinate grids, masks, and point sampling.
- bspline_gradient/
- bspline_cuda_autograd.cpp: PyTorch bindings and automatic differentiation.
- bspline_cuda_kernel.cu: CUDA interpolation and gradient kernels.
- bspline_cuda_optimized.cu: parallel B-spline prefilter kernels.
- setup.py: builds and installs the CUDA extension.
- Docker/
- Dockerfile: CUDA/PyTorch container environment.
- requirements.txt: Python dependencies for the container.
If you use DoReMe in your research, please cite:
Majd Helo, Dominik Nickel, and Thomas Kuestner. DoReMe: Dual optimization of Reconstruction and Motion estimation for self-navigated free-breathing abdominal MRI using Gaussian Splatting. arXiv:2610.19120.