Smart Brain Data Analyses — Machine Learning approach for a background segmentation of the 3D image volume of a brain tissue block
The project is based on a database of around 700 pictures that show consecutive tissue sections of the human brain – like those typically used in high-resolution 3D-modelling. Such 3D models are part of current top-level research projects at the research center in Jülich, such as the EU-flagship project “Human Brain Project”.
The brain, underlying our image dataset, was frozen at -80°C and cut into 70µm-thin sections in a cryotome. Before conducting the next cut and fixating the particular tissue section on a glass substrate, each sectional surface of the tissue block was scanned with an industrial camera (blockface-image). Later on, these blockface-images are important as reference images for the correct shape of the tissue because during the fixation on the glass substrate the tissue sections often get deformed and also damaged sometimes. However, it is possible to revise later taken pictures of the fixed tissue sections by using the blockface-images so that the original shape of the tissue can be reproduced as correctly as possible.
Machine learning algorithms (e.g. Support Vector Machines) need datasets for automated learning processes (“Labelled Data”). For that reason, a very time-consuming manual marking of the image information has already taken place once. The automation is currently studied by means of the SDIL platform and available IBM solutions.
The task is to mask the 3D stack of the blockface images in such a way that
solely the brain tissue of the section surface is visible
the 3D surface is extracted as correctly as possible
The high-resolution images of the tissue sections are transferred back into an anatomically consistent 3D-volume. In the case of the highly complex image registration procedures that are the basis of this project, the blockface-images are used as a reference. For example, volume, structure, and shape of the brain can be visualized like this. In addition, it also allows to map and explore the boundaries of brain regions.
“Together with SDIL, we have ventured into the depths of the human brain” says Timo Dickscheid from Forschungszentrum Jülich. “We used more than 700 images of successive tissue sections of the human brain to investigate new approaches for efficient image segmentation, based on machine learning algorithms.”
Data Innovation Community
FZ Jülich, KIT, IBM
Timo Dickscheid, email@example.com