This is the code for paper "Coil-Agnostic Attention-Based Network for Parallel MRI Reconstruction" accepted at ACCV2022.
-
create Anaconda environment via conda create -n myenv
-
install torch, e.g. with cuda=10.2: conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=10.2 -c pytorch
-
install other dependencies, e.g. pip package:
- h5py
- matplotlib
- pathlib
Multi-coil knee MRI samples from the FastMRI Dataset [1] are used in our experiments. Data are saved in h5 files with keywords csm and multi_coil, representing the sensitivity maps and multi-coil images. The structure of one instance is shown as follows:
- file['section0']['csm'][()] of shape (15, 2, 256, 256),
- file['section0']['multi_coil'][()] of shape (15, 2, 256, 256),
where 15 is the number of coils, 2 channels are used to handle complex-valued data, and (256, 256) refers to the spatial size. The dataset and dataloader can be conveniently customized by tracking the specified data size in codes.
run the code to train the model:
python multicoil_knee_baseline.py --mask ./mask/brain_mask_0_125.pt --dataroot path/to/training/dataset --testroot path/to/test/samples --dataset_name multi_coil_DACB --n_epochs 35 --epoch 0
Checkpoints will be saved in ./saved_models/multi_coil_DACB and samples in ./images/multi_coil_DACB.
- Zbontar, J.*, Knoll, F.*, Sriram, A.*, Murrell, T., Huang, Z., Muckley, M. J., ... & Lui, Y. W. (2018). fastMRI: An Open Dataset and Benchmarks for Accelerated MRI. arXiv preprint arXiv:1811.08839.