This package provides a fast implementation of the Color Deconvolution algorithm by Ruifrok and Johnson [1].
For adjusting staining vectors, the package provides a dynamic tool that lets you adjust and inspect the quality of deconvolution results which bases on the work of Gilbert Bigras [2].
The ColorDeconvolution will happily work with even very large histological images.
Color Deconvolution is a method that extracts stain intensities from histological RGB images. So if you have an e.g. H&E stained sample, the algorithm aims to recover the Hematoxylin and the Eosin portions into separate images. In theory, the staining needs to follow the Beer-Lambert law that states that the absorbance of the single dyes are proportional to their concentration. However, in practice, it is not uncommon to use the algorithm even with dyes that don't follow this law like the DAB staining.
Therefore, a microscopic image of a stained histological section is regarded as the process of transmitting white light through the sample, where each dye absorbs a characteristic portion that depends on the concentration of the dye and its color. Using the characteristic color-vectors for a staining, this operation can be inverted and the algorithm is able to separate the concentration of each dye in the image. The result of this approach are intensity-images for each used dye.
To obtain excellent results, each step in the process of acquiring an image must be carefully revised to ensure
consistent lighting, white-balance, and correct camera settings. A basic tutorial can be found in
"How to correct background illumination in brightfield microscopy".
The acquired bright- and dark-images mentioned in the above article can used as an optional setting to the ColorDeconvolution
function that will then take care of correcting the background.
Install the package directly as paclet from the releases section on GitHub
PacletInstall[".."]I will soon include the package on our community paclet server but for now, download the repository and copy the contents
of the ColorDeconvolution sub-folder to a folder in the $Path of Mathematica.
The usual place to put packages in is
FileNameJoin[{$UserBaseDirectory, "Applications"}]The color deconvolution algorithm itself is rather simple, but it consists of two sub-functions functions that are computational expensive on large images. The first one is the computation of the optical density image from your original image and after that, the calculation of the deconvolution image. Both functions, however, can be applied in parallel on each pixel and the implementation here makes use of it.
For optimal performance, you should have a working C compiler on your system and you should ensure, that Mathematica can use it to create optimized code. Your compiler works and can be used by Mathematica if the following snippet runs without error
Compile[{}, 1, CompilationTarget -> "C"][]If a compiler is not available, the package will use Mathematica's virtual machine automatically. The parallel performance depends on the number of cores you have and how many parallel threads are used by Mathematica. You can check your number of cores by running
$ProcessorCount
(* 12 *)and the number of parallel threads by running
SystemOptions["ParallelOptions" -> "ParallelThreadNumber"]
(* {"ParallelOptions" -> {"ParallelThreadNumber" -> 12}} *)The following shows the running-time of the ColorDeconvolution on with an image of size 2048 x 2048 on a Mac Pro 2,7 GHz 12-Core Intel Xeon E5.
There are certain other implementations for color deconvolution. The plugin for ImageJ by Gabriel Landini is well known and its preset of color vectors is included in this package. Jakob Nikolas Kather provided a package for MATLAB and refers to two other MATLAB implementations whose sites seem to be down.