Skip to content

Python wrapper for cudaDecon - GPU accelerated 3D deconvolution for microscopy

License

Notifications You must be signed in to change notification settings

tlambert03/pycudadecon

Repository files navigation

pycudadecon

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

Install

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

GPU requirements

This software requires a CUDA-compatible NVIDIA GPU.

The libraries available on conda-forge 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 may specify cuda-version as follows:

conda install -c conda-forge pycudadecon cuda-version=<11 or 12>

You should also ensure that you have the minimum required driver version installed for the CUDA version you are using.

Usage

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. https://doi.org/10.1364/AO.36.001766