FNOReg: Resolution-Robust Medical Image Registration Method Based on Fourier Neural Operator

This is an official implementation of paper FNOReg: Resolution-Robust Medical Image Registration Method Based on Fourier Neural Operator

Installation

git clone https://github.com/anac0der/fnoreg.git
cd fnoreg
python3 -m venv .
. bin/activate
pip install -r requirements.txt

Dataset downloading

We used the preprocessed version of OASIS-1 dataset from Adrian Dalca repository. To train our models on this dataset you need to do the following steps:

• Download 2D and 3D data from the specified source;
• Update paths to the datasets in corresponding configuration files (with regexp params_*.json, fields oasis_path for path to dataset folder and oasis_folder_path for path to file subjects.txt).

Launching experiments and reproducing the results

After installation of all required dependencies and downloading the data, you need to download model checkpoints in order to reproduce results from our paper. It can be done by running bash scripts ./download_ckpt.sh, ./download_ckpt_3d.sh.

To get further instructions about reproduction of metrics values in our paper, you can read instructions_to_reproduce.md.

Brief explanation of main commands (if you want to launch your own experiments):

Training scripts: train_*.py, in folder /deep_fourier_reg for FNO-based models, VoxelMorph-Large and Fourier-Net and in folder /baseline_models/transmorph for VoxelMorph, VoxelMorph-Huge and TransMorph.

To launch the model training, firstly fill in the corresponding config file and then run the following command:

python3 train_*.py --gpu_num gpu_num --size size

Here gpu_num is number of GPU device in your system and size is the size of smallest dimension of input data shape (160 for full resolution and 80 for halved resolution).

Evaluation scripts: files with pattern evaluate_*.py. After running of the experiment you can see its number in console output. To evaluate the experiment with number N, run the following command:

python3 evaluate_*.py --gpu_num gpu_num --exp_num N --ckpt_epoch ckpt_epoch

Here:

• gpu_num is the same as in the training scripts;
• ckpt_epoch is the number of epoch from which you want to download the checkpoint. This argument should be omitted if you want to evaluate the final model (model with weights after all epochs).

Acknowledgements

Hoopes et al. Learning the Effect of Registration Hyperparameters with HyperMorph - for providing preprocessed data collection;

Chen, Junyu, et al. TransMorph: Transformer for Unsupervised Medical Image Registration - we used the source code of this paper for training VoxelMorph and TransMorph.