Analysis approach for series of rodent-brain section image

Ilastik is a versatile image analysis tool specifically designed for the classification, segmentation, and analysis of biological images based on supervised machine-learning algorithms.

There are two main approaches for the analysis of rodent-brain section image:

1. pixel classification only (with two or more classes) and
2. pixel classification with two classes (immunoreactivity and background), followed by Object classification with two classes (objects of interest and artefact).

Which approach is best for my dataset?

As a general rule, pixel classification is suitable for images in which there are clear differences in the colour, intensity, and / or texture of the feature of interest (labelling) versus the background and other structures.  If there is non-specific labelling in the image that is very similar in appearance to the labelling of interest, object classification may allow the non-specific labelling to be filtered out on the basis of the object-level features such as shape and size. The best approach is determined by trial and error.

Pixel classification workflow

For a quick introduction, watch this video.

Basic steps

• Train the classifier with two classes (labelling and background).
• Apply the classifier to the rest of the images (batch processing).
• Export the probability maps in HDF5 format and simple_segmentation images in PNG format with the default settings.
• Review the results.

Object-classification workflow

There are three options on the ilastik start-up page for running "Object Classification." Choose the Object Classification with Raw Data and Pixel Prediction Maps as input.

• Save the object-classification file in the same folder as the raw images for analysis. If the images are moved after the ilastik file is created, the link between the ilastik file and the images may be lost, resulting in a corrupted file.
• In the "Input Data" applet, upload the original images and their respective probability maps in HDF5 format (output from the Pixel Classification).
• Train the classifier with two classes (labelling and artefacts).
• In the "Object Information Export" applet, export “Object Predictions” in PNG format.  Do not change the default export location.
• Review the results.