Flowdec

*Note*: This library is no longer actively maintained and requires older versions of Python and TensorFlow to run. If you have point spread functions already, then cucim.skimage.restoration.richardson_lucy is another implementation worth considering. If not, then the utilities in this library for generating them may still be useful.

---

Flowdec is a library containing TensorFlow (TF) implementations of image and signal deconvolution algorithms. Currently, only Richardson-Lucy Deconvolution has been implemented but others may come in the future.

Flowdec is designed to construct and execute TF graphs in python as well as use frozen, exported graphs from other languages (e.g. Java).

Here are a few other features, advantages, and disadvantages of the project currently:

Highlights

• Support for Windows, Mac, and Linux - Because TensorFlow can run on these platforms, so can Flowdec.
• Client Support for Java, Go, C++, and Python - Using Flowdec graphs from Python and Java has been tested, but theoretically they could also be used by any TensorFlow API Client Libraries.
• Point Spread Functions - PSFs can be defined as json configuration files to be generated dynamically during the deconvolution process using a Fast Gibson-Lanni Approximation Model (which can also create Born & Wolf kernels as a degenerate case).
• GPU Accleration - Executing TensorFlow graphs on GPUs is trivial and will happen by default w/ Flowdec if you meet all of the TensorFlow requirements for this (i.e. CUDA Toolkit installed, Nvidia drivers, etc.).
• Performance - There are other open source and commercial deconvolution libraries that run with partial GPU acceleration, which generally means that only FFT and iFFT operations run on GPUs while all other operations run on the CPU. For example, on a roughly 1000x1000x11 3D volume with a PSF of the same dimensions this means that execution times look like:
  • CPU-only solutions: 10 minutes
  • Other solutions with FFT/iFFT GPU acceleration: ~40 seconds
  • Flowdec/TensorFlow with full GPU acceleration: ~1 second
• Signal Dimensions - Flowdec can support 1, 2, or 3 dimensional images/signals.
• Multi-GPU Usage - This has yet to be tested, but theoretically this is possible since TF can do it (and this Multi-GPU Example is a start).
• Image Preprocessing - A trickier part of deconvolution implementations is dealing with image padding and cropping necessary to use faster FFT implementations -- in Flowdec, image padding using the reflection of the image along each axis can be specified manually or by letting it automatically round up and pad to the nearest power of 2 (which will enable use of faster Cooley-Tukey algorithm instead of the Bluestein algorithm provided by Nvidia cuFFT used by TF).
• Visualizing Iterations - Another difficulty with iterative deconvolution algorithms is in determining when they should stop. With Richardson Lucy, this is usually done somewhat subjectively based on visualizing results for different iteration counts and Flowdec at least helps with this by letting observer functions be given that take intermediate results of the deconvolution process to be written out to image sequences or stacks for manual inspection. Future work may include using Tensorboard to do this instead but for now, it has been difficult to get image summaries working within TF "while" loops.

Disadvantages

• No GUIs - Flowdec is intended for use by those familiar with programming but some future work might include an ImageJ plugin (if there's interest in that). For those looking for something more interactive, imagej-ops is likely your best bet which currently supports the same PSF generation model used in Flowdec as well as Richardson Lucy deconvolution. At the moment it doesn't include full GPU acceleration but that may be on the way as part of imagej-ops-experiments. See this github issue for more details.
• No Blind Deconvolution - Currently, nothing in this arena has been attempted but since much recent research on this subject is centered around solutions in deep learning, TensorFlow will hopefully make for a good platform in the future.

Basic Usage

Here is a basic example demonstrating how Flowdec can be used in a single 3D image deconvolution:

See full example notebook here

%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
from skimage import exposure
from scipy import ndimage, signal
from flowdec import data as fd_data
from flowdec import restoration as fd_restoration

# Load "Purkinje Neuron" dataset downsampled from 200x1024x1024 to 50x256x256
# See: http://www.cellimagelibrary.org/images/CCDB_2
actual = fd_data.neuron_25pct().data
# actual.shape = (50, 256, 256)

# Create a gaussian kernel that will be used to blur the original acquisition
kernel = np.zeros_like(actual)
for offset in [0, 1]:
    kernel[tuple((np.array(kernel.shape) - offset) // 2)] = 1
kernel = ndimage.gaussian_filter(kernel, sigma=1.)
# kernel.shape = (50, 256, 256)

# Convolve the original image with our fake PSF
data = signal.fftconvolve(actual, kernel, mode='same')
# data.shape = (50, 256, 256)

# Run the deconvolution process and note that deconvolution initialization is best kept separate from 
# execution since the "initialize" operation corresponds to creating a TensorFlow graph, which is a 
# relatively expensive operation and should not be repeated across multiple executions
algo = fd_restoration.RichardsonLucyDeconvolver(data.ndim).initialize()
res = algo.run(fd_data.Acquisition(data=data, kernel=kernel), niter=30).data

fig, axs = plt.subplots(1, 3)
axs = axs.ravel()
fig.set_size_inches(18, 12)
center = tuple([slice(None), slice(10, -10), slice(10, -10)])
titles = ['Original Image', 'Blurred Image', 'Reconstructed Image']
for i, d in enumerate([actual, data, res]):
    img = exposure.adjust_gamma(d[center].max(axis=0), gamma=.2)
    axs[i].imshow(img, cmap='Spectral_r')
    axs[i].set_title(titles[i])
    axs[i].axis('off')

As a more realistic use case, here is an example showing how a point spread function configuration can be used in a headless deconvolution:

See full deconvolution script here

# Generate a configuration file containing PSF parameters (see flowdec.psf module for more details)
echo '{"na": 0.75, "wavelength": 0.425, "size_z": 32, "size_x": 64, "size_y": 64}' > /tmp/psf.json

# Invoke deconvolution script with the above PSF configuration and an input dataset to deconvolve.
# If flowdec has been installed, you may run the “deconvolution” command.
python examples/scripts/deconvolution.py \
--data-path=flowdec/datasets/bars-25pct/data.tif \
--psf-config-path=/tmp/psf.json \
--output-path=/tmp/result.tif \
--n-iter=25 --log-level=DEBUG
> DEBUG:Loaded data with shape (32, 64, 64) and psf with shape (32, 64, 64)
> INFO:Beginning deconvolution of data file "flowdec/datasets/bars-25pct/data.tif"
> INFO:Deconvolution complete (in 7.427 seconds)
> INFO:Result saved to "/tmp/result.tif"

Examples

Python

• Neuron - Deconvolution of a natural 3D image with synthetic point spread function
• C. Elegans - Deconvolution of 712x672x104 acquisition for 3 separate channels
• Astronaut - Dealing with artifacts in deconvolved images
• Monitoring Iterations - Tracking deconvolution progress by visualizing results at each iteration
• Hollow Bars - Deconvolution of downsampled 64x64x32 synthetic volume (as a CPU-friendly example)
• Hollow Bars GPU Benchmarking - Testing running times on full 256x256x128 volume with GPU-enabled system
• DeconvolutionLab2 Comparison - Comparing execution times between DeconvolutionLab2 and Flowdec
• Graph Export - Defining and exporting TensorFlow graphs
• Command Line Interface - CLI for executing single deconvolutions with either a pre-defined or dynamically generated point spread function

Java

• Multi-GPU Example - Prototype example for how to be able to execute deconvolution against multiple GPUs in parallel (not tested yet -- waiting for the use case to come up though it is very likely possible to do)

Installation

The project can be installed, ideally in a python 3.6 environment (though it should work in 3.5 too), by running:

pip install flowde