Analysis of 3D Microscopy Images

3D image analysis is one of the core research areas within the Visual Computing section at DTU Compute. Our research ambition is to develop new methods for 3D image analysis for fast and accurate quantification of 3D microstructures from large 3D volumes. Much of our research is in close collaboration with scientists developing and using 3D imaging. Together, we futher advance 3D imaging as a reliable tool for measuring microstructure.

Thanks to the construction of the large-scale facilities MAX IV and ESS in Lund, Sweden, there is now a fantastic opportunity for 3D imaging at extremely high spatial and temporal resolutions close to DTU. An obstacle in fully utilizing this opportunity is a time consuming image analysis, which is challenged by the size and the complexity of the acquireded data sets. There is therefore a great need for expertise and new analysis tools for 3D imaging. Developing this expertise and associated tools is our research focus.

We have established the QIM: The Center for Quantification of Imaging Data, which is a collaboration between DTU, KU, and Lund University and also a part of the DTU 3D Imaging Center - 3DIM, which is our X-ray µCT laboratory that is also tied closely to the DANMAX beamline at MAX IV. Much of our research is done with collaborators in QIM and 3DIM.

Research core

Quantitative characterization of 3D images involves measuring the geometric and intensity characteristics of volumetric images. Often, the geometric features are of interest, because they determine the properties of a sample. This gives a particular focus on methods for geometric characterization that includes methods for segmentation, morphology, and other features that describe geometry.

Segmentation based on machine learning

Segmentation based on machine learning is a central part of our research. Machine learning is based on the principle that model parameters are learned from data, which for image segmentation typically is learning from labeled images.

Dictionary-based segmentation

Our work on dictionary-based segmentation is based on machine learning principles. In our work, the dictionary is made of a set of characteristic image features that have associated labels. This enables automatic image segmentation. Both the dictionary and labels are learned from data, and the learning is very fast, which makes it possible to obtain real-time interactive segmentation. An outcome of this is the InSegt segmentation tool, that is particularly strong for segmenting repetitive structures as shown below.

Dictionary-based interactive segmentation with InSegt

Graph-based segmentation

Our work on graph-based segmentation involves the development of novel methods (see for example our SLG work, or NOS work), implementation of efficient graph-based algorithms (see Parallel PQBP and our comparison paper), and the use of graph-based models for analysing large 3D data (see for example our analysis of peripheral nerves and muscle fibres).