MoDIP

Quantitative Susceptibility Mapping through Model-Based Deep Image Prior

NeuroImage 2024  |  deepMRI collection

MoDIP performs QSM reconstruction using a model-based deep image prior (DIP) framework, combining a forward physical model with a data-driven prior without requiring training data.

---

Overview

---

Requirements

• Python 3.7+, PyTorch
• NVIDIA GPU

---

Installation

git clone https://github.com/sunhongfu/MoDIP.git
cd MoDIP

conda create -n MoDIP python=3.8
conda activate MoDIP
pip install torch nibabel numpy scipy

---

Quick Start

python recon.py \
  --data_path sample.nii \
  --save_path ./result \
  --z_prjs 0 0 1 \
  --vox 1 1 2 \
  --input_type phi

---

Arguments

Input:

Argument	Description
--data_path	Path to input NIfTI file
--is_fieldmap	Set if input is a field map (not raw phase)
--input_type	Input type: phi (raw phase), pure (axial field map), or noise

DFO (Data Fidelity Optimization):

Argument	Description
--iter	Number of DFO iterations
--alpha	DFO step size

DIP (U-net):

Argument	Description
--depth	U-net depth
--base	Base channel number
--decoder_block_num	Number of decoder blocks

Other:

Argument	Description
--z_prjs	B0 direction (z-projections), e.g. 0 0 1 for axial
--vox	Voxel size in mm, e.g. 1 1 2
--crop_background	Crop background (mask background to 0 beforehand)
--padding_mode	Padding mode for QSM forward calculation

Memory tip: Decrease --base (channel number) and increase --iter if GPU memory is insufficient.

---

Citation

@article{modip2024,
  title={Quantitative susceptibility mapping through model-based deep image prior},
  journal={NeuroImage},
  year={2024},
  doi={10.1016/j.neuroimage.2024.120...}
}

---

⬆ top  |  deepMRI collection