Robust Simultaneous Multislice MRI Reconstruction Using Slice-Wise Learned Generative Diffusion Priors
This is the codebase for Robust Simultaneous Multislice MRI Reconstruction Using Slice-Wise Learned Generative Diffusion Priors (ROGER), which is built upon openai/guided-diffusion with modifications for the training and sampling of image diffusion models and its application to MRI simultaneous multislice reconstruction.
Fig.1 Illustration of the simultaneous multislice (SMS) MRI sampling and reconstruction process. The acquisition involves multiple non-adjacent slices with CAIPI (controlled aliasing in parallel imaging) shift patterns and additional in-plane acceleration, resulting in sparse k-space and complex aliasing artifacts. The reconstruction of slices is difficult due to these strong aliasing artifacts and the absence of fully-sampled autocalibration signals in many SMS-accelerated sequences.
Fig.2 Schematic illustration of the proposed ROGER method. More details can be found in paper.
Prospectively accelerated dataset and generative model weight are released at Google Drive
Method for getting retrospectively accelerated brain dataset from fastMRI:
fastMRI_data_preprocess.ipynb
Before inference, the data should be saved in 'npz' format like:
readout_data = data['readout_data']
readout_calibration = data['readout_calibration']
readout_csm = data['readout_csm']
shifts = data['shifts']
python infer.py --input meas_MID00273_FID03217_TSE_SMS_334_SMS_data_slice0.npz --output recon_MB3R3.npz --MB 3 --R 3 --chk 384x384_ema_0.9999_200000.pt | Sampling setting | Mask | SMS Image | Recon | GT |
|---|---|---|---|---|
| MB4R1 |
|
|
|
|
| MB4R2 |
|
|
|
|
| MB4R3 |
|
|
|
If you have any questions, please contact Prof.Lyu (lvmengye@sztu.edu.cn) or raise an issue in the repository.
@article{huang2024robust,
title={Robust Simultaneous Multislice MRI Reconstruction Using Slice-Wise Learned Generative Diffusion Priors},
author={Shoujin Huang, Guanxiong Luo, Yunlin Zhao, Yilong Liu, Yuwan Wang, Kexin Yang, Jingzhe Liu, Hua Guo, Min Wang, Lingyan Zhang, Mengye Lyu*},
journal={Medical Image Analysis},
pages={103851},
year={2025},
publisher={Elsevier}
}