pixel classification

[hackermd]

Let's create a new workflow step for pixel classification based on Ilastik. The step should be able to apply a classifier trained via the Ilastik GUI (saved to disk as .ilp file) to a set of images and output a probability map (.png file) for each image.

This should be implemented as a step of the image_preprocessing stage using Ilastik in headless mode via the command line. It is only a temporary solution, however. Ultimately, the classifier should be applied to an image directly in Python, which should be implemented as a Jterator module, see TissueMAPS/JtLibrary#4.

[hackermd]

After a longer discussion, we decided to implement the main pixel classification ourselves independent of the Ilastik framework.

At its core, Ilastik uses either vigra or skikit-learn for the actual classification. The rest of the program is concerned with GUI, lazyflow threaded workflow management (which we don't want because we want to run it on a cluster) and parsing of arguments. All of this creates an unnecessary overhead.

We rather want a very light-weight implementation that simply takes features and labels as inputs (as numpy arrays) and outputs the trained classifier.
The major tasks will be to:

• draw on the map in the TissueMAPS GUI, see Downloading Classifier Results TissueMAPS#16
• get the pixel coordinates and labels via the TissueMAPS GUI and load the corresponding images
• generate the features, i.e. filter the images, see skimage or OpenCV

For the actual classification, we will try to stick as close as possible to Ilastik's core classification functionality of the create_and_train() methods, e.g. VigraRfPixelwiseClassifier.

[hackermd]

In order to be able to load the corresponding images we need a mapping of pixel coordinate to image file. This could be implemented in the ChannelLayer class.

[hackermd]

@hafenr: When you click on the map can you also retrieve the name of the corresponding tile (.jpg file) or row/column coordinate of the pyramid at the highest level?

[riccardomurri]

This issue was moved to TissueMAPS/TissueMAPS#122