Skip to content

Latest commit

 

History

History
52 lines (40 loc) · 5.47 KB

CHANGES.md

File metadata and controls

52 lines (40 loc) · 5.47 KB
Raw

Changelog

v1.01

  • The 🏎️ edition. RIESLING is now much faster due to multiple optimizations (judicious use of fast-math, gridding algorithm improvements, better and more widespread threading, and more besides). Many thanks to Martin Reinecke for suggestions.
  • Completely new SENSE calibration algorithm based on NLINV. Sensitivities are now estimated and stored as kernels in k-space, and only inflated to full maps when required. The NLINV regularizer is used during calibration to ensure smooth maps.
  • New low memory mode --lowmem. This performs the SENSE and NUFFT operators sequentially for each channel and accumulates the result. Reduces the memory required for the oversampled grid to a single channel. The speed penalty is acceptable.
  • Experimental Direct Virtual Channel Combination --decant, i.e. apply SENSE via convolution in k-space not multiplication in image space. This is much slower, but saves even more memory than --lowmem.
  • Virtual Conjugate Channels --vcc for real-valued image reconstruction. Experimental.
  • The way FOV cropping and oversampling works has been refined. Now, the --fov=x,y,z option refers to the gridding FOV used during the reconstruction, and the oversampling is applied in addition to this. The --crop-fov=x,y,z option is applied at the end (default is the nominal FOV). This makes some behind-the-scenes calculations much simpler.
  • Implemented the multi-channel version of Frank Ong's preconditioner. This speeds up convergence, but takes some time to calculate itself. Hence the single-channel is still the default. Specify --precon=multi if you want to try it.
  • Multi-channel TV regularizer.

v1.00

  • I now consider RIESLING useable enough to declare version 1 🥂.
  • The trajectory scaling now matches BART (-M/2 to M/2 where M is matrix size).
  • The interface has been much simplified - input .h5 files are expected to have a dataset named "data" and commands take both input and output filenames. This removes the need to remember the specific dataset names and suffixes added by commands.
  • Useful commands have been renamed and reorganized to aid discoverability. The most important are recon-lsq and recon-rlsq, which are now named by the problem they solve (Least-Squares and Regularized Least-Squares respectively) instead of the particular algorithms used (previously they were lsmr and admm).
  • The new montage tool can create pretty montages from RIESLING datasets. These can be saved as PNGs or displayed in the terminal if you use KiTTY. Sadly, due to dependency issues, this command is not included in the Github downloads and you will need to compile on your local machine if you want to use it.
  • Some commands have been removed, in particular cg. The remaining commands have many advantages and maintaing all the different algorithms was costing considerable time.
  • Behind the scenes, this version can be considered the "preconditioned least-squares everywhere" release. The operator commands now use a few iterations of preconditioned LSMR to calculute the inverse NUFFT etc. in preference to Density Compensation. This approach was already used for SENSE calibration and I consider it the superior approach over Density Compensation. A discussion on this is welcome at ISMRM, especially if you provide a glass of riesling.
  • RIESLING now uses the DUCC FFT library instead of FFTW. This has comparable performance, no planning, and a much better multi-threading implementation.
  • New and improved regularizers including wavelets and TGV on the L2-norm of multi-channel images.

v0.12

  • A much improved ADMM implementation, including the residual balancing scheme from Wohlberg 2017. This shows much more robust convergence behaviour. The default number of inner iterations is now only 1, giving improved speed.
  • Implemented a version of PDHG that supports multiple regularizers. However, I prefer the improved ADMM implementation.
  • Implemented the NDFT (Non-uniform Discrete Fourier Transform). This is incredibly slow compared to the NUFFT, but revealed some small innacuracies in the NUFFT implementation. These have now been fixed.

v0.11

  • Fixed many bugs that crept into v0.10.
  • Functionality to write out residual k-space/images
  • Small but important tweaks to how the ADMM algorithm works including more sensible defaults.
  • Added a through-time TV regularizer option for ADMM.
  • Added a tool to calcuate a basis set from temporal images.

v0.10

  • The admm command now supports TGV regularization. The separate tgv command, which used the form of PDHG in the original paper, has hence been retired. The admm version is superior as it properly supports preconditioning.
  • The log is now saved into the output .h5 file for late reference.
  • Iterative algorithms now scale the data. This makes regularization parameters (λ) behave sensibly. The default is to calculate a scaling factor by perfomring an adjoint NUFFT, using Otsu's method to identify foreground signal, and then taking the median value. If --scale=bart is used then the BART scaling of 90th percentile is used. A pre-calculated factor can also be specified with --scale=X.
  • Improvements to loading/saving data from Python thanks to Martin Krämer.
  • A basis can now be specified for several plotting functions.
  • The FFT wisdom is now saved to a path that includes the hostname and executable name. This gives more consistent behaviour in networked environments with hetergeneous server hardware.
  • Added a maximum entropy regularizer.

Before

The changelog was only started for version 0.10.