Graphic STI
logo EPFL
Text EPFL
english only
Biomedical Imaging Group
IP-LAB: Image Processing Laboratories
BIG > Teaching > IP-LAB > DeconvolutionLab 18

3D Microscopy Deconvolution

In order to perform the deconvolution of 3D images acquired on modern microscopes, we use the complete software package DeconvolutionLab2 instead of using the framework of ImageAccess as usual.

DeconvolutionLab2 is an open-source software package that includes several standard algorithms of deconvolution. It can be run in two modes:

Here, we mainly use the second option to facilitate the reproducibility of the experiments. Nevertheless, all the operations can be done through the GUI with the plugin DeconvolutionLab2 Lab.

DeconvolutionLab2 has a bunch of functionalities that allows one to validate the algorithms by running convolution and deconvolution on known synthetic datasets and synthetic PSF.

1. Introduction to the DeconvolutionLab2 Java library

Below, we provide an example of code. Test this method on siemens-star-3D using the plugin Experiment Example.

public static void example() {
	// Get a RealSignal object r from a open image of ImageJ
	RealSignal r = Lab.getImage("siemens-star-3D.tif");
	if (r == null) {
		IJ.error("No open image.");
		return;
	}
	// Create a RealSignal object, here a Gaussian function with 3 different sigmas (sigma_x, sigma_y, sigma_z)
	RealSignal h = new Gaussian(1, 2, 4).generate(r.nx, r.ny, r.nz);
	// Display h as 3 orthogonal section from the center
	Lab.showOrthoview(h, "h");

	// Create a operator Simulation = convolution and additive gaussian noise of mean = 0 and standard deviation = 5
	Simulation sim = new Simulation(0, 5, 0);
	// Run the simulation with h as PSF. The method .noPopup() hides the final stack ouput
	RealSignal y = sim.noPopup().run(r, h);
	// Display y as a 2D planar montage of all z-slices
	Lab.showPlanar(y, "y");

	// Run the deconvolution operator RIF, regularization factor = 0.001
	RegularizedInverseFilter rif = new RegularizedInverseFilter(0.001);
	RealSignal xrif = rif.noPopup().run(y, h);
	Lab.showOrthoview(xrif, "RIF");
	IJ.log("SNR RIF " + Assessment.snr(xrif, r));

	// Run the deconvolution operator Landweber, 10 iterations, step = 1
	// The methods .setReference(r).setStats() enables a table of the SNR after every iteration
	Landweber lw = new Landweber(10, 1);
	lw.setReference(r).setStats();
	RealSignal xlw = lw.run(y, h);
	Lab.showOrthoview(xlw, "LW");
	IJ.log("SNR LW " + Assessment.snr(xlw, r));
}

Orthoview of the "siemens-star-3D" image

2. Inverse Crime

A naive way to evaluate the deconvolution chain is to convolve (multiplication in the Fourier domain) an image with a PSF and then deconvolve it using a simple stabilised division in the Fourier domain. This last operation is the role of the Naive Inverse Filter (NIF).

Using the same input image r siemens-star-3D and the same PSF h (Gaussian(1, 2, 4)) of the Question 1, code the method inverseCrime() that generates the 3 following outputs:

This method has to also display the 3 outputs as orthoview and computes the SNR with respect to the ground-truth image r. Conclude on the Naive Inverse Filter and fill the report.

3. Wiener Inverse Filter

In presence of noise, a solution is to use a Regularized Inverse Filter where a hyper-parameter λ controls the regularisation regime.

In this experiment, we ask to find a good regularization factor λ for different levels of noise n using the algorithm TikhonovRegularizedInverseFilter(λ) depending of the noise of the input image. The input image yn is artificially created with the following commands:

	RealSignal r = new Cube(10, 0.01).generate(36, 36, 36);
	RealSignal h = new Gaussian(1, 1, 1).generate(r.nx, r.ny, r.nz);
	Simulation sim = new Simulation(0, n, 0);
	RealSignal yn = sim.noPopup().run(r, h);
	Lab.showPlanar(yn, "y" + n);

Code the method inverseFilter() which loops over λ from 0 to 1 (step = 0.05) to find λ given the best SNR(r, xn).

This method has to also display the output xn as planar and computes the SNR with respect to the ground-truth image r. Run the experiment for n = 0, 0.25, 0.5, 0.75 and 1 using the plugin Inverse Filter. Fill the report.

4. Comparison of algorithms

In this experiment, we ask to compare five algorithms of deconvolution and find their parameter given the best SNR.

Algorithm Symbol Type Parameter Constraint
Regularized Inverse Filter RIF Direct λ
Laplacian regularization factor
None
Tikhonov Regularized Inverse Filter TRIF Direct λ
Tikhonov regularization factor
None
Richardson Lucy RL Iterative Number of iteration (niter) None
Landweber LW Iterative Number of iteration (niter)
The step γ is always set to 1.8
None
LandweberPositivity LW+ Iterative Number of iteration (niter)
The step γ is always set to 1.8
Non-negative

The input image y is artificially created with the following commands:

	RealSignal r = Lab.getImage("siemens-star-3D.tif");
	RealSignal h = new BesselJ0(1, 2, 0.001, 0.00001).generate(96, 96, 96);
	Simulation sim = new Simulation(0, 0.1, 5);
	RealSignal y = sim.noPopup().run(r, h);
	Lab.showOrthoview(y, "y");

Code the method comparison() which search the best λ (a power of 10 is enough precision) for RIF and TIRF and the best niter (maximum 300 iterations) for RL, LW and LW+. In order to compare algorithms, the method .setReference(r).setStats(); computes automatically the SNR of the deconvolved image and the reference image r (after every iterations for iterative algorithm).

5. Save Sherlock Holmes!


Fragment of the picture

The question can be done using the ImageJ commands and the GUI of DeconvolutionLab, no need to code.

Sherlock Holmes is not satisfied! Dr. Watson took a picture with hand-motion blur. Help Sherlock Holmes to retrieve the hidden quote in the image quote-motion-blur.

6. Deconvolution of a real biological specimen


Maximum intensity projection of the input image

The question can be done using the ImageJ commands and the GUI of DeconvolutionLab, no need to code.

Try to solve the deconvolution task as well as you can on the real biological sample. Because in real life the ground-truth is unknown, the only quality criterion is your visual appreciation.

The acquired data (one fluorescence FITC channel, green) is data-FITC and the microscope PSF is PSF-FITC. Fill the report.


webmaster.big@epfl.ch • 23.05.2018