Patch Denoising Methods

This repository implements patch-denoising methods, with a particular focus on local-low rank methods.

The target application is functional MRI thermal noise removal, but this methods can be applied to a wide range of image modalities.

It includes several local-low-rank based denoising methods (see the documentation for more details):

1. MP-PCA
2. Hybrid-PCA
3. NORDIC
4. Optimal Thresholding
5. Raw Singular Value Thresholding

A mathematical description of these methods is available in the documentation.

Installation

$ pip install patch-denoise

patch-denoise requires Python>=3.12

Apptainer / Singularity

A container recipe with the GPU extras (PyTorch, Triton) is provided in patch-denoise.def. Build it once, from a checkout of this repository (no root needed with --fakeroot):

$ apptainer build --fakeroot --build-arg VERSION=$(git describe --tags | sed s/^v//) \
      patch-denoise.sif patch-denoise.def

and run it like the regular command-line. Use --nv to expose the host GPU, and --bind for data that lives outside your home directory:

$ apptainer run --nv --bind /path/to/data patch-denoise.sif input.nii output.nii --mask="auto"

The image bundles CUDA 13 user-space libraries, so the host NVIDIA driver must support CUDA 13 (nvidia-smi reporting driver >= 580). Without --nv the CPU implementation is used.

On a Slurm cluster:

$ srun --gres=gpu:1 --cpus-per-task=8 --mem=32G --time=01:00:00 \
      apptainer run --nv --bind /path/to/data patch-denoise.sif input.nii output.nii --mask="auto"

Quickstart

After installing you can use the patch-denoise command-line.

$ patch-denoise input_file.nii output_file.nii --mask="auto"

See patch-denoise --help for detailed options.

Documentation and Examples

Documentation and examples are available at https://paquiteau.github.io/patch-denoising/

Development version

$ git clone https://github.com/paquiteau/patch-denoising
$ pip install -e patch-denoising[optional] --group dev

Citation

If you use this package for academic work, please cite the associated publication, available on HAL

@inproceedings{comby2023,
  TITLE = {{Denoising of fMRI volumes using local low rank methods}},
  AUTHOR = {Pierre-Antoine, Comby and Zaineb, Amor and Alexandre, Vignaud and Philippe, Ciuciu},
  URL = {https://hal.science/hal-03895194},
  BOOKTITLE = {{ISBI 2023 - International Symposium on Biomedical Imaging 2023}},
  ADDRESS = {Carthagena de India, Colombia},
  YEAR = {2023},
  MONTH = Apr,
  KEYWORDS = {functional MRI ; patch denoising ; singular value thresholding ; functional MRI patch denoising singular value thresholding},
  PDF = {https://hal.science/hal-03895194/file/isbi2023_denoise.pdf},
  HAL_ID = {hal-03895194},
  HAL_VERSION = {v1},
}

Related Packages

• https://github.com/paquiteau/retino-pypeline

  For the application of the denoising in an fMRI pipeline using Nipype