2026 Group-patch joint compression: compressing dynamic B0 and static RF spatial modulations across k-space subregion groups for highly accelerated MRI

This is a package of example codes and data for the group-patch joint compression paper, including a Pulseq sequence example (Wave-CAIPI) and offline MATLAB reconstruction, under BSD 3-Clause License.

We propose a group-patch joint compression technique to reduce the computational burden of MRI image reconstruction associated with rapid B0 modulations and readout oversampling.

For instance, during the frequency-encoding phase of FLASH/GRE sequences, rapid B0 field modulation and readout oversampling can be enabled to collect more k-space information per unit time.

Depending on the linearity of the rapid B0 field, this approach corresponds to bunched-phase encoding or Wave-CAIPI (linear), or FRONSAC/local B0 coil modulations (nonlinear).

While combining readout oversampling with multiple RF receivers accelerates acquisition, it significantly slows down reconstruction—particularly for iterative compressed sensing (CS) and neural network-based approaches.

To address this, our proposed method jointly compresses data encoded by both the RF receivers and the rapid B0 fields within small k-space patches.

Because the rapid B0 modulation is periodic along the k-space readout, the small compression matrices required for a single modulation period only need to be calculated once. 

These matrices can then be efficiently reused to compress the entire k-space patch-by-patch across different compression groups.

The joint compression substantially surpasses the conventional RF array compression method, when rapid B0 field modulation and readout-oversampling are used for highly-accelerated MRI sampling.

Please download the .zip file, unzip it, and check the multi-shot EPI sequence generation and the related offline reconstruction. A Pulseq package is also included, which was used to generate sequences and support the reconstruction.

Note that, the 3D CS-Wave package in this version was tested using a high memory compute node with 1T RAM. The current implementation is not memory-efficient and is only for proof-of-concept purpose. If your computer does not support such high RAM, you may have to optimize the recon script (e.g., reusing a small readout encoding matrix with one rapid B0 modulation period, rather than explicitly storing the entire readout encoding matrix). Anyway, the benefits of joint compression should be universally applicable to different implementations of CS reconstructions.

Please cite the following paper:

It was also published in the conference:

Tian R, Scheffler K. k-space subregion-wise joint compression: compressing dynamic B0 and static RF field modulations for accelerated MRI. Abstract #01000. Oral presentation. The annual meeting of ISMRM, Cape Town, South Africa, 2026.

For further questions, please contact Rui Tian, rui.tian@tuebingen.mpg.de/ruitianwater@outlook.com

The example codes are also available in GitHub:

https://github.com/ruitianwater/GroupPatchJCS_v1.0

This package also includes third-party components to support running the code:
- Pulseq (directory: pulseq-master), with its own license.