2013-04-22 - Trainable Segmentation Hackathon

For three days last week, LOCI in Madison hosted Dr Ignacio Arganda-Carreras of Trainable Segmentation fame for a mini-hackathon. Just in case you do not know what the Trainable Segmentation is: it provides an intuitive, graphical way to apply advanced Machine Learning techniques to image segmentation.

During this hackathon, Ignacio held a well-received workshop introducing the underlying technique and teaching how to use it in advanced ways by using one of Fiji's scripting languages.

As a lot of scientists rely on the Trainable Segmentation (I was impressed that it is apparently used already in dozens of published papers, even before the Trainable Segmentation paper is out, point in favor of making your research available openly, to as many people as possible!), we added a few unit tests. These tests are implemented as Java code that generates test images and compares the results to known-good ones, and they get executed by the ImageJ Jenkins server whenever Fiji's source code changes. That way, we can be reasonably sure that the plugin still does what it is expected to, even if we make a ton of changes.

We also made progress on a completely new idea: making the Trainable Segmentation extensible. What does this mean, exactly?

The Trainable Segmentation plugin works by calculating a number of features for every pixel. Such a feature might be the Gaussian convolution with radius 5, or the derivative in the x-direction. But really, a feature can be any image processing filter. For every coordinate these numbers -- plus the original pixel value -- form a so-called feature vector. Given a bunch of such feature vectors for, say, background, cell membrane and cell interior, Machine Learning tries to find clusters in the high-dimensional feature space which correspond to the classes.

Typically, combinations of the most common image processing filters form a good basis for a feature space to discern between the different classes of interest.

Sometimes, however, you need to augment that feature space with highly specialized filters, e.g. Tubeness or a non-standard wavelet filter to make it easier to segment your data reliably.

This is where our work comes in: similar to ImageJ2 commands, programmers will be able to implement their own features for use with the Trainable Segmentation.

All it needs is to conform to a simple interface and to add an annotation (no worry, once this is finalized, we will provide a very easy tutorial on how to implement your own feature). Then, just drop in your .jar file into the jars/ folder (you can even offer these on your own update site!). No need for configuration files, let alone changes in the Trainable Segmentation itself.

That way, if you come up with a powerful algorithm that makes classification of your objects of interest much easier, just a few minutes of work will enable you to include that algorithm as a feature (or multiple features) in the Trainable Segmentation.

Something else we started to work on: we identified the places to change for ImgLib2 support.

Due to ImgLib2's interface-drive design, it is very easy to wrap any data structure (e.g. Icy's, Alida's, ImageJ 1.x', etc) to conform with ImgLib2's interfaces, so all of the algorithms implemented to work with ImgLib2 interfaces can run on all of aforementioned data structures.

Hence, once Trainable Segmentation is ported to work on ImgLib2 structures, it will be much more useful: it will no longer be limited to ImageJ 1.x!

All in all, a very productive hackathon!