Universal Segmentation Framework (UniSeF)

Deutsches Elektronen-Synchrotron DESY


Helmholtz Zentrum Geesthacht (HZG)

Synchrotron radiation based computed tomography (SRCT) enables to study a wide spectrum of samples at a high spatial, temporal, and density resolution. Segmentation of the reconstructed tomograms is essential for almost all projects. However, this task is usually done manually or in a semi-automatic fashion, since routine methods for automatic segmentation often fail or deliver unsatisfying results.
Convolutional neural networks, in particular U-Nets and derivatives, have already been applied to segment 2D slices of 3D SRCT data. Here, we will develop a dedicated segmentation framework which addresses the particular challenges of SRCT data. This includes very large volume sizes in the order of (10k)3 voxels, low contrast, high noise, image artefacts, and the exploitation of the available three dimensional information. The proposed developments comprise the automatic selection of the most suitable deep learning segmentation architecture, the segmentation of identical objects (instance segmentation), a guided interactive and iterative strategy for the annotation of training data, and the deployment of a browser-based service.
Instance segmentation of 3D volumetric data is not well studied, and the challenge to precisely segment hundreds of instances in billions of voxels is only to a limited extent comparable with published instance segmentation methods. We expect the resulting methods to be applicable to various other imaging techniques.

Coordination at DESY:

Deutsches Elektronen-Synchrotron
Dr. Philipp Heuser
Notkestr. 85
D-22607 Hamburg

e-mail: philipp.heuser@desy.de
Tel.: 040 8998 4622

Weiterführende Links