TensorFlow2 Implementation of Non-Uniform FFT with Kaiser-Bessel Gridding
This package is an adaptation of nuFFT and adjoint nuFFT functions in SigPy.
Currently supports two computation modes
- Exact kernel computation:
Kaiser-bessel kernel values are computed on the fly, no table interpolation is used. - Precomputed kernel:
Kaiser-bessel kernel values are first computed and stored. Table interpolation is not used for kernel computation. Precomputation must be repeated for each set of coordinates.
The computation speeds are given in milliseconds, for a 256x256 image with a spokelength of 512 and 405 spokes. The numbers are obtained on an NVIDIA GeForce RTX 2080 Ti graphics processor.
| Operation | CPU (eager) |
CPU (graph mode) |
GPU (eager) |
GPU (graph mode) |
|---|---|---|---|---|
| Forward | 395.1 | 259.3 | 111.7 | 13.3 |
| Adjoint | 568.8 | 438.1 | 119.3 | 14.1 |
| Operation | CPU (eager) |
CPU (graph mode) |
GPU (eager) |
GPU (graph mode) |
|---|---|---|---|---|
| Precompute | 15.8 | 15.1 | 12.4 | 12.9 |
| Forward | 275.2 | 249.5 | 21.3 | 12.8 |
| Adjoint | 429.1 | 395.7 | 23.9 | 13.4 |
The computation times can be obtained by running the following commands
python profile_tfnufft.py
python profile_tfnufft_precomputed.py
If you use this library for your work, please consider citing the following work:
@article{alkan2024autosamp,
title={AutoSamp: Autoencoding k-space Sampling via Variational Information Maximization for 3D MRI},
author={Alkan, Cagan and Mardani, Morteza and Liao, Congyu and Li, Zhitao and Vasanawala, Shreyas S and Pauly, John M},
journal={IEEE Transactions on Medical Imaging},
year={2024},
publisher={IEEE}
}