Dual-Channel Segmentation Model for Dynamic Shim

Implementation of paper "Reliable Off-Resonance Correction in High-Field Cardiac MRI Using Autonomous Cardiac B_0 Segmentation with Dual-Modality Deep Neural Networks"

Getting started

Installation

# if conda already installed
conda create --name dynamicShim python=3.9
conda activate dynamicShim

# install the dependencies for preprocess
pip install torchio natsort

# install the dependencies for segmentation
# to the nnUNet folder
cd nnUNet
pip install -e .

Also set the environment PATH as described in nnUNet or specify that in nnUNet/nnunetv2/path.py

Usage

data requirement, predict single subject each time. Folder structure

subject/
├── subject_0000.nii (magnitude map)
├── subject_0001.nii (phase map)

move to matlab folder and run the data_prepare.m to prepare the data. Two UI will pop up to select the input folder and output folder.
Run the segmentation.

nnUNetv2_predict -i <input folder> -o <outputput folder> -d 301 -c 3d_fullres --save_probabilities -chk checkpoint_latest.pth -device cpu --verbose

Matlab

data_prepare.m: Load the original data and mask out the air using automask.
main_preprocess.m: Load the data and prepare the 3D volumes for segmentation.
main_shim.m: Load the B0field and Bz field. Calculate the shim coil current.
B0Bzpreprocess.m: save every single 3D volume in .mat file as a Nifti file.
get_img_params.m: get the image parameters from B0 mat file.
B0Bzmap.m: Resolution matbch between the B0 and Bz maps.
dynamic_shim.m: Prepare the ROI of B0 and Bz for dynamic shim.
solveDC.m: Apply the lsqlin to solve the coil current.

Python

This implementation is based on this paper:

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2020). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 1-9.