Python wrapper for CUDA-accelerated 3D deconvolution

    This package provides a python wrapper and convenience functions for cudaDecon, which is a CUDA/C++ implementation of an accelerated Richardson Lucy Deconvolution algorithm1.

    • CUDA accelerated deconvolution with a handful of artifact-reducing features.
    • radially averaged OTF generation with interpolation for voxel size independence between PSF and data volumes
    • 3D deskew, rotation, general affine transformations
    • CUDA-based camera-correction for sCMOS artifact correction


    The conda package includes the required pre-compiled libraries for Windows and Linux. See GPU driver requirements below

    conda install -c conda-forge pycudadecon

    macOS is not supported

    📖   Documentation

    GPU requirements

    This software requires a CUDA-compatible NVIDIA GPU. The underlying cudadecon libraries have been compiled against different versions of the CUDA toolkit. The required CUDA libraries are bundled in the conda distributions so you don't need to install the CUDA toolkit separately. If desired, you can pick which version of CUDA you'd like based on your needs, but please note that different versions of the CUDA toolkit have different GPU driver requirements:

    To specify a specific cudatoolkit version, install as follows (for instance, to use cudatoolkit=10.2)

    conda install -c conda-forge pycudadecon cudatoolkit=10.2
    CUDALinux driverWin driver
    10.2≥ 440.33≥ 441.22
    11.0≥ 450.36.06≥ 451.22
    11.1≥ 455.23≥ 456.38
    11.2≥ 460.27.03≥ 460.82

    If you run into trouble, feel free to open an issue and describe your setup.


    The pycudadecon.decon() function is designed be able to handle most basic applications:

    from pycudadecon import decon
    # pass filenames of an image and a PSF
    result = decon('/path/to/3D_image.tif', '/path/to/3D_psf.tif')
    # decon also accepts numpy arrays
    result = decon(img_array, psf_array)
    # the image source can also be a sequence of arrays or paths
    result = decon([img_array, '/path/to/3D_image.tif'], psf_array)
    # see docstrings for additional parameter options

    For finer-tuned control, you may wish to make an OTF file from your PSF using pycudadecon.make_otf(), and then use the pycudadecon.RLContext context manager to setup the GPU for use with the pycudadecon.rl_decon() function. (Note all images processed in the same context must have the same input shape).

    from pycudadecon import RLContext, rl_decon
    from glob import glob
    import tifffile
    image_folder = '/path/to/some_images/'
    imlist = glob(image_folder + '*488*.tif')
    otf_path = '/path/to/pregenerated_otf.tif'
    with tifffile.TiffFile(imlist[0]) as tf:
        imshape = tf.series[0].shape
    with RLContext(imshape, otf_path, dz) as ctx:
        for impath in imlist:
            image = tifffile.imread(impath)
            result = rl_decon(image, ctx.out_shape)
            # do something with result...

    If you have a 3D PSF volume, the pycudadecon.TemporaryOTF context manager facilitates temporary OTF generation...

     # continuing with the variables from the previous example...
     psf_path = "/path/to/psf_3D.tif"
     with TemporaryOTF(psf) as otf:
         with RLContext(imshape, otf.path, dz) as ctx:
             for impath in imlist:
                 image = tifffile.imread(impath)
                 result = rl_decon(image, ctx.out_shape)
                 # do something with result...

    ... and that bit of code is essentially what the pycudadecon.decon() function is doing, with a little bit of additional conveniences added in.

    Each of these functions has many options and accepts multiple keyword arguments. See the documentation for further information on the respective functions.

    For examples and information on affine transforms, volume rotations, and deskewing (typical of light sheet volumes acquired with stage-scanning), see the documentation on Affine Transformations

    1 D.S.C. Biggs and M. Andrews, Acceleration of iterative image restoration algorithms, Applied Optics, Vol. 36, No. 8, 1997.


    • 0.4.1

    Last updated:

    • 13 September 2023

    First released:

    • 10 August 2022


    Supported data:

    • Information not submitted

    Plugin type:

    GitHub activity:

    • Stars: 55
    • Forks: 11
    • Issues + PRs: 8

    Python versions supported:

    Operating system:

    • Information not submitted


    • numpy
    • tifffile
    • typing-extensions
    • importlib-metadata ; python_version < "3.8"