Stacks Image 1284

Ilastik: Machine Learning for Biologists


Microscopy isn’t all about the microscopes, it’s also about what to do with the images after they’ve been captured. So for a change of scenery, here’s a short introduction to the powerful world of machine learning. Don’t be put off if you have only a vague idea of its usefulness to your research, results so far have been very promising from this piece of open source software.

Machine learning suffers from a number of misconceptions, which include: it’s too hard for anyone not trained in programming; it’s the realm of engineers and biologists should not consider it as part of their image analysis toolbox; and it’s a slippery slope towards a dystopian world of Skynet and Terminator.

But to deny biologists a chance to use machine learning on the grounds it’s outside their usual experience is not how science works. Science has always been about collaboration and a very productive collaboration between programmers and biologists in the German city of Heidelberg has produced a program call Ilastik. (English speakers should pronounce it more like ‘elastic’ to replicate the German pronunciation.) It is expressly designed for biologists to make use of the advances in machine learning in their research.

Machine learning provides a more nuanced method of segmenting out regions or volumes of interest in images. Instead of simply using a pixel intensity range, as often used in programs such as Fiji, it also uses boundary changes, colour combinations and pattern recognition to determine which regions of an image represent the same regions of the underlying biology. It works very well with colour images as well as greyscale and once ‘taught’ the characteristics of an image, can batch process other similar images.

The output is generally a series of binary images, each containing the appropriate regions of the images, closer to what a human observer would choose. These are then quantified using programs such as Fiji.

Like Fiji, Ilastik is open source and free to use in your research. It will run on OS X, Linux and Windows, but in all cases it’s best to have a fast processor and plenty of RAM, as the computational demand is relatively high, particularly in 3D data sets.


Please contact Andrew McNaughton for further details and enquires about training how you might be able to use Ilastik in your research.
Stacks Image 1831
Stacks Image 1834
Stacks Image 1835
Stacks Image 1836
Stacks Image 1837
Stacks Image 1838
Stacks Image 1839
Stacks Image 1840
Stacks Image 1841
Further Reading

The links to the left will take you to the Ilastik website, or to a specific topic.
The program is under quite rapid development, so check back to see if other features have been added.
IlastikWhere to start and get a feel for what it can do, a few specific examples are linked below to get you started.
Pixel ClassificationThe workflow is especially suited if the objects of interests are visually (brightness, color, texture) distinct from their surrounding. The algorithm is applicable for a wide range of segmentation problems that fulfil these properties.
CarvingThis is suited to greyscale images which do not exhibit clearly delineated intensity zones, but have features separated by boundaries. TEM images, both 2D and 3D are good candidates for this method.
Animal Tracking
Digitally label animals in mazes and track/quantify where and when they go.

Confocal: How it Works

For most biological samples, a fluorescent marker(s) is required to use the confocal microscope.

Generally, if you've successfully stained your samples and viewed the results under an ordinary epi-fluorescence microscope, your samples will be suitable for the confocal. In other words, you've already done 90% of what's required.

Light Microscopy

We have a number of research grade light microscopes available for use. Most also have epi-fluorescence as well as bright field capabilities.

Other, more specialised microscopes, can obtain montages of large specimens or help you with stereology investigations.

Fiji and Analysis

Open source software. like ImageJ, can provide an excellent platform to base image analysis upon. There's a lot more to simply obtaining a 'pretty picture' with any of the equipment listed here.

You should be able to at least make semi-quantitative assessments of the data you have collected. This is where image analysis come in.

µCT: Sample Suitability

The µCT uses Computerised Axial Tomography (CAT) to produce a series of digital slices of solid objects. Best suited to relatively dense material like teeth and bone, the µCT can resolve details down to about 1µm.
If you're looking for electron microscopy techniques, go here: OCEM

Contact Details

Andrew McNaughton
03 479 7308
andrew.mcnaughton@otago.ac.nz

Room B01g Basement
Department of Anatomy
Lindo Ferguson Building
270 Great King Street
(Opposite main entrance to Public Hospital, Great King Street)
PO Box 913
Dunedin 9001
New Zealand