Generalized Deep Diffusion-Weighted Imaging (DeepDWI) Reconstruction Framework

Introduction

DeepDWI implements self-supervised learning image reconstruction techniques for DWI.

Prospectively self-gated diffusion-weighted image reconstruction on a iEPI data acquired with 0.7 mm isotropic resolution. Self-gated zero-shot self-supervised learning (ZSSSL) enables much cleaner delineations of diffusion contrasts than locally-low rank (LLR) regularized reconstruction.

Installation

Let's use conda to install DeepDWI:

1. create a new conda environment named ('deepdwi', you can use other names as you like):

   conda create -n deepdwi python=3.10

2. activate the environment:

   conda activate deepdwi

3. download and install deepdwi:

   git clone https://github.com/ZhengguoTan/DeepDWI.git

   and then cd to the DeepDWI directory,

   python -m pip install -e .

4. download and install sigpy:

   git clone https://github.com/ZhengguoTan/sigpy.git

   and then cd to the sigpy directory,

   python -m pip install -e .

That's it! Enjoy deepdwi!

Features

1. The repo implements several state-of-the-art algorithm unrolling for high-resolution high-dimensional diffusion-weighted imaging reconstruction, incl.

   (1) Variational Network (VarNet);

   (2) Model-based deep learning architecture (MoDL);

   (3) Alternating Direction Method of Multipliers (ADMM).

• This implementation is also applicable to other high-dimensional MRI reconstruction!

Data

Before running the scripts in DeepDWI, you need to download the following data to the /data/ folder:

Note	Spatial Resolution (mm3)	Diffusion Mode	Acceleration (in-plane x slice)	Shots	Navigator	Link
vol3	0.7 x 0.7 x 0.7	MDDW 20 directions with b-value of 1000 s/mm2	2 x 2	3	No
vol2	0.7 x 0.7 x 0.7	MDDW 20 directions with b-value of 1000 s/mm2	2 x 2	3	No
vol1	0.7 x 0.7 x 0.7	MDDW 20 directions with b-value of 1000 s/mm2	2 x 2	3	Yes
	1.0 x 1.0 x 1.0	MDDW 20 directions with b-value of 1000 s/mm2	1 x 3	4	No
	1.0 x 1.0 x 1.0	3-shell 126 directions with b-values up to 3000 s/mm2	3 x 3	2	No

Examples

General Example

1. load the data

   cd data/
   
   python load.py --records 10781347 --file 0.7mm_21-dir_R2x2_md5sum.txt
   
   python load.py --records 10781347 --file 0.7mm_21-dir_R2x2_kdat_slice_000.h5 --md5sum 0.7mm_21-dir_R2x2_md5sum.txt
   
   python load.py --records 10781347 --file 0.7mm_21-dir_R2x2_coils.h5 --md5sum 0.7mm_21-dir_R2x2_md5sum.txt

2. run the zsssl training

   cd examples/
   python run_zsssl.py --mode train

3. run the zsssl testing

   python run_zsssl.py --mode test --slice_idx 0 --checkpoint /examples/path/to/zsssl_best.pth

Specific Examples

• Please follow the sub-folders in /figures/ for specific experiments and examples.

• To plot /figures/ without running all experiments, we provide saved results here: . Please download them into the /data/ folder.

References

If you find the open-source codes/data useful, please cite:

@Article{tan_2025_sg_ssl,
    Title     = {{High-Resolution Diffusion-Weighted Imaging with Self-Gated Self-Supervised ADMM Unrolling}},
    Author    = {Tan, Zhengguo and Liebig, Patrick A and Hofmann, Annika and Laun, Frederik B and Knoll, Florian},
    Journal   = {},
    Year      = {},
    Volume    = {},
    Pages     = {},
    doi       = {}
}

@Article{tan_2024_diff7t,
    Title     = {{Accelerated Diffusion Weighted Magnetic Resonance Imaging at 7 T: Joint Reconstruction for Shift-Encoded Navigator-based Interleaved Echo Planar Imaging (JETS-NAViEPI)}},
    Author    = {Tan, Zhengguo and Liebig, Patrick A and Heidemann, Robin M and Laun, Frederik B and Knoll, Florian},
    Journal   = {Imaging Neuroscience},
    Year      = {2024},
    Volume    = {2},
    Pages     = {1-15},
    doi       = {10.1162/imag_a_00085}
}