AFTER-QSM

Affine Transformation Edited and Refined (AFTER) Deep Neural Network for Quantitative Susceptibility Mapping

NeuroImage 2022  |  deepMRI collection

AFTER-QSM is a dedicated deep neural network for QSM reconstruction from highly oblique and anisotropic resolution scans (down to 0.6 mm isotropic), where standard networks trained on axial data fail due to orientation mismatch.

---

Requirements

• Python 3.7+, PyTorch
• NVIDIA GPU (8 GB+ VRAM recommended)

Tested on: Windows 11.

---

Overview

Framework

Fig. 1: Overview of the AFTER-QSM training process using affine transformation editing and refinement.

Simulation Results

---

Usage

Training

python train.py \
  --root <training_data_root> \
  --file_path <path_list_file> \
  --depth 4 \
  --recon_base 32 \
  --channel_num 16 \
  --refine_base 16 \
  --block_num 3 \
  --layer_num 4 \
  --batch_size 1 \
  --epoch 100 \
  --gpu_num 1 \
  --lr1 1e-4 \
  --lr2 1e-4

Implement your dataset by overriding path_composition in train.py.

Key training arguments:

Argument	Description
--root	Root directory of training dataset
--depth	U-net depth for QSM reconstruction
--recon_base	Base channel number for reconstruction U-net
--channel_num	Channels for successive refinement network
--refine_base	Base channels for refinement network
--block_num	Number of RDB blocks
--layer_num	Layers per RDB block
--use_dense	Dense architecture vs. Squeeze-and-Excitation
--padding	Zero-padding radius for rotation boundary

Testing

python test.py \
  --field_path <path_to_field_map.nii> \
  --vox <vx vy vz> \
  --z_prjs <x y z> \
  --checkpoint_path <path_to_checkpoint> \
  --save_path <output_dir>

Key testing arguments:

Argument	Description
--field_path	Path to testing data
--vox	Voxel size of testing data
--z_prjs	B0 direction (z-projections)
--use_GPU	Use GPU (boolean)
--GPU_NO	GPU device index
--save_path	Output save path
--checkpoint_path	Path to model checkpoint
--segment_num	Memory trade-off: 8 for <24 GB, 4 for <12 GB
--data_type	field for direct recon, otherwise forward calc

---

Citation

@article{afterqsm2022,
  title={Affine transformation edited and refined deep neural network for quantitative susceptibility mapping},
  journal={NeuroImage},
  year={2022},
  doi={10.1016/j.neuroimage.2022.119655}
}

---

⬆ top  |  deepMRI collection