Auto-Differentiation Based MRI Pulse Design

Reference implementation of:
Joint Design of RF and Gradient Waveforms via Auto-Differentiation for 3D Tailored Exitation in MRI
(arXiv: https://arxiv.org/abs/2008.10594)

cite as:

@article{luo2021joint,
  author={Luo, Tianrui and Noll, Douglas C. and Fessler, Jeffrey A. and Nielsen, Jon-Fredrik},
  journal={IEEE Transactions on Medical Imaging},
  title={Joint Design of RF and gradient waveforms via auto-differentiation for 3D tailored excitation in MRI},
  year={2021},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TMI.2021.3083104}}

For the interpT feature, consider citing:

@inproceedings{luo2021MultiScale,
  title={Multi-scale Accelerated Auto-differentiable Bloch-simulation based joint design of excitation RF and gradient waveforms},
  booktitle={ISMRM},
  pages={3958},
  author={Tianrui Luo and Douglas C. Noll and Jeffrey A. Fessler and Jon-Fredrik Nielsen},
  year={2021}
}

Steady-state sequence optimization

Besides regular tailored-excitation design, the pulse can be optimized against the steady-state magnetization of an SPGR sequence. Use adpulses.optimizers.arctanLBFGS_spgr (Python) or adpulses.opt.arctanAD_spgr (MATLAB) — the steady-state counterparts of arctanLBFGS / adpulses.opt.arctanAD. demo/demo_ss.m gives a worked comparison of a regular design against a steady-state design.

This feature accompanies a technical note currently under review at Magnetic Resonance in Medicine; a citation will be added once available.

System Requirements:

• Python ≥3.8.

Tested on Ubuntu 18.04 / 20.04; other operating systems are untested but not explicitly excluded.

General comments

setup_AutoDiffPulses.m does the configurations for Matlab.
For the python part, in your command line, navigate to the repo's root directory, type:

pip install -e .

Demos are provided in ./demo. In particular, demo/demo_ss.m uses the included example dataset (demo/IniVars.mat) to compare a regular design against a steady-state design under steady-state Bloch simulation.

Dependencies

This work requires Python (≥v3.8) and PyTorch (≥v1.6).

Python

pip install -e . (see above) installs all Python dependencies, including the steady-state-enabled MRphy.py (pinned in setup.py):

• mrphy: tianrluo/MRphy.py (≥v0.2.1) — with steady-state (SPGR) Bloch simulation support (mrphy.steady_state.spgr).

Other Python dependencies (numpy, scipy, torch) are installed automatically.

MATLAB

setup_AutoDiffPulses.m bootstraps the MATLAB dependencies automatically: it clones the packages below into ./matlab_deps (git-ignored) at the pinned commits, compiles the C Bloch simulator (mex blochcim.c), and adds them to the path. This requires git on the system PATH. To force a fresh checkout, delete ./matlab_deps and re-run the script.

• +mrphy: tianrluo/MRphy.mat — the MATLAB counterpart, extended with steady-state Bloch simulation support.
• +attr: fmrilab/attr.mat.

MATLAB must also be configured (via pyenv) to use a Python environment that has the mrphy installation described above.