Advanced Settings

Likelihood

Visible for Fast Iterative and Constrained Iterative algorithms.
Selects which likelihood calculation you want to work with.

–

Poisson (Meinel)

Only visible for the Fast Iterative algorithm.
The calculation according to Meinel works with one convolution per iteration and converges very fast, normally in 4-5 iterations. This method can also produce artifacts, however.

–

Poisson (Richardson-Lucy)

Only visible for the Fast Iterative algorithm.
Accelerated version of Richardson-Lucy, based on a Poisson likelihood and Biggs acceleration (https://doi.org/10.1364/AO.36.001766), allowing for fast and artifact free reconstructions.

–

Poisson

Only visible for the Constrained Iterative algorithm.
Computation assuming a Poisson noise distribution, this is normally the correct noise model for microscopic images.

–

Gauss

Only visible for the Constrained Iterative algorithm.
Computation assuming a Gaussian noise distribution. If detector noise is dominant over sample noise, using a Gaussian noise model can be advantageous, however, this is rarely the case with modern microscopy systems.

Regularization

Selects which frequencies in the image are taken into account during regularization.

–

None

Only visible for the Fast Iterative and Constrained Iterative algorithms.
No regularization is performed.

–

Zero Order

Regularization based on G-difference, modeled on Tikhonov, but accelerated.

–

First Order

Only visible for the Regularized Inverse Filter and Constrained Iterative algorithms.
Regularization based on Good's roughness. Under certain circumstances, more details are extracted from noisy data. It may be better suited to the processing of confocal data sets.

–

Second Order

Only visible for the Constrained Iterative algorithm.
Regularization according to Tikhonov-Miller. Here higher frequencies are penalized more than in the case of Good's roughness. Results have a tendency to become overly smoothed.

–

Total Variation

Only visible for the Fast Iterative algorithm.
Total Variation regularization denoises the input data but protects the edges of structures. This can improve results for noisy data but increases the computation time quite significantly.

Optimization

Visible for Fast Iterative and Constrained Iterative algorithms.

–

Analytical (Newton Raphson)

Only visible for the Constrained Iterative algorithm.
Here an attempt is made to optimize the iterations analytically. It is an optimization method to find the step size in each of the Constrained Iterative iterations. This option is usually faster but may also be somewhat less precise.

–

Line Search

Only visible for the Constrained Iterative algorithm.
Searches rigorously and comprehensively for the minimum. It is therefore more robust, but the calculation takes longer. It is an optimization method to find the step size in each of the Constrained Iterative iterations. Line search is recommended for confocal data sets, especially if they are noisy as this can enforce convergence even for noisy and sparsely sampled data.

–

Numerical Gradient

Only visible for the Fast Iterative algorithm.
If selected, an attempt is made to determine the trend of the iterations in advance and extrapolate this to the entire calculation. This can significantly speed up the calculation.

First Estimate

Visible for Fast Iterative and Constrained Iterative algorithms.

–

Input Image

The input image is used as the first estimate of the target structure (default).

–

Last Result Image

The result of the last calculation is used to estimate the next calculation. This can speed up a calculation that is repeated using slightly different parameters.

–

Mean of Input

No estimate is made, the mean gray level of the input image is being used. This is the most rigid application of deconvolution. It should be chosen for confocal images, where the data sampling can be quite sparse. The computation time will increase, but missing information can be recovered from the PSF.

–

Zero Values

Only visible for the Constrained Iterative algorithm.

Maximum Iterations

Visible for Fast Iterative and Constrained Iterative algorithms.
Sets the maximum permitted number of desired iterations. In the case of Richardson-Lucy, you should allow significantly more iterations here.

Quality Threshold

Only visible for the Fast Iterative and Constrained Iterative algorithms.
Defines the quality level at which you want the calculation to be stopped. The percentage describes the difference in enhancement between the last and next-to-last iteration compared with the greatest difference since the start of the calculation. 1% is the default value. Lowering this can bring about small improvements in quality.

GPU Acceleration

Only visible if a suitable (NVIDIA, CUDA based) graphics card is installed in your PC. The checkbox is then activated by default.

Activated: Uses GPU processing.
Deactivated: Uses CPU processing.

GPU Tiling

Only available for very large images that exceed the available graphic card memory.

Activated: With this function the image is split up in smaller portions which fit into the memory of the graphic card. The function automatically determines into how many tiles the image must be split to allow maximum usage of the graphics card. The resulting tiles are automatically stitched together for the final output result.

Deactivated: No tiling is performed, however, in this case only certain sub-functions of deconvolution can run on the graphics card and the speed increase compared to CPU processing will be lower. The image quality might be higher than with tiling because there is no need for stitching.