davidssmith/tron

Trajectory Optimized Nufft

TRON: TRajectory Optimized Nufft

Introduction

This is the repository for the paper, "Trajectory Optimized Nufft: Faster Non-Cartesian MRI Reconstruction through Prior Knowledge and Parallel Architectures".

Fast non-Cartesian MRI reconstruction ("gridding") requires interpolation of non-uniformly sampled Fourier data onto a Cartesian grid. This interpolation is slow due to complicated, non-local data access patterns that are hard to optimize for modern, parallel CPU architectures.

TRON is a gridding code customized for a standard radial (both linear and golden angle) trajectory and Nvidia's CUDA architecture to eliminate the bottlenecks that gridding algorithms encounter when written for arbitrary trajectories and hardware.

TRON is 30x faster than gpuNUFFT (a fast GPU code) and 75x faster than the image reconstruction toolbox (IRT) of Fessler (a ubiquitous CPU-based Matlab code). TRON eliminates the need to presort data or perform a separate sample density compensation, while comprising 50% less code than the minimal subset of IRT tested.

Reproducing the Paper Results

To get generate the paper figures and data, follow these steps:

Requirements: Newish version of MATLAB and CUDA.

0. Make sure you have a working CUDA installation with cuFFT and the compiler nvcc in your path.

1. Clone and compile TRON into your folder of choice.

2. Pull the TRON example data from the LFS server using git-lfs pull in the folder where you cloned TRON.

3. Clone and compile gpuNUFFT into your folder of choice.

4. Clone and compile BART into your folder of choice.

5. Correct the path to gpuNUFFT and BART in the RUNME2 script.

6. Run src/RUNME1_tron_degrid_phantom.sh on the command line.

7. Change to the src/ subdirectory and run src/RUNME2_others_degrid_phantom.m in MATLAB. This should not take long to finish.

8. Go back to the command line and run src/RUNME3_tron_grid_all.sh. This will also not take long.

9. Go back to MATLAB and run ./RUNME4_others_grid_all.m. This is slow. An hour is normal. Put your feet up and read our paper while you wait.

10. If all steps run without errors, you're done. Congratulations! The paper figures should be in src/figs and the reconstructions will be in src/output.

tron's Issues

Input file is hardcoded

Need to be able to pass input file on command line.