SegmentationTrainer - A Robust and User-Friendly Machine Learning Image

Segmentation Solution

Nicolas Piche1, Isabelle Bouchard1, Mike Marsh2

1. Object Research Systems. Montreal, Canada.
2. Object Research Systems. Denver, CO, USA.

For many experimental imaging techniques such as FIB-SEM nanotomography, image

segmentation remains a bottleneck that impedes efficient quantitative analysis and,
consequently, it slows many imaging-based research projects. A number of advanced
segmentation techniques have been advanced, both manual and automated, but they are
usually highly-specialized and tuned to meet the needs of highly constrained image analysis
problems. To address the shortcoming, we present a generalized machine learning method that
is suitable for a large class of research imaging applications.

Incorporated into our commercial image processing platform, Dragonfly, our Segmentation
Trainer is a machine learning engine that utilizes textural details in the image to discriminate
between material phases. We will present the paradigm whereby users can build Classifiers that
are capable of automatically segmenting images. Classifiers are highly configurable finite-state
machines that take one or more input image channels, extract pre-defined textural details, and
then use a configurable learning machine (e.g. Random Forest, Deep Neural Network) to label
all pixels in the image to which phase they best match. In order to add robustness to the
texture encoding, we incorporate a Superpixel-based aggregation of statistics.

We will show examples from materials, life science, and geoscience. We will discuss
integration into more complicated workflows and integrations with macro recording to apply
sequential actions in a reproducible but flexible way. Finally, we will also present the newly-
launched online repository for sharing Classifiers with others in the community for broader use
and continued refinement to improve accuracy.