Characterizing the microstructure of energy materials based on X-ray CT results in large amounts of data. The research within analysis methods includes techniques for efficiently handling and analyzing large-scale image data. Specifically the research aims at methods for managing large-scale data, fast and efficient computation, reconstruction, segmentation, and quantification techniques.

Because of the large amounts of data acquired using X-ray CT imaging techniques, the data handling requires special attention. Dedicated grid-based data-management solutions where data can be stored, accessed and processed rapidly are focus of our research using the Minimum Intrusion Grid.

An analysis pipeline often used for analyzing X-ray CT image data involves CT reconstruction, segmentation, characterization in the form of structure quantification, and modeling. Our research is focused on techniques that are easy to understand and intuitive to use, and enables the use of prior geometric information. Specifically we have based our research on two methodologies – deformable meshes and dictionaries of image patches. 

Deformable meshes provide a highly useful data-structure applicable to both CT-reconstruction and segmentation. Preserving a high quality mesh-structure is essential in order to retain robust mesh-based modeling, and therefore we employ the so-called deformable simplicial complex method (DSC).   

In the CINEMA project we are investigating the possibilities of DSC-based CT-reconstruction, and an initial result is shown in Figure…). Typically, a reconstruction is based on a voxel grid, containing a billion voxels or more, and from a modeling point of view, all voxels are parameters that must be estimated. This is a time consuming even with modern hardware accelerated techniques3D hamster single. Using a mesh-based approach drastically reduces the number of parameters to a set of mesh nodes. Our initial results show that the method is highly robust to noise and sparsely sampled data, which is promising for the materials applications in CINEMA.

DSC is also a well-suited framework for image segmentation in both 2D and 3D as shown in the example here. Our experiments have shown that DSC based segmentation is robust to noise while being computational efficient. This mesh-based segmentation makes additional modeling, using simulation methods like finite elements, a natural next step in our research. 

Some segmentation problems benefit from contextual information utilizing the texture surrounding a pixel. We have explored an example of this when segmenting individual fibers in a dense fiber composite material as shown in the figure below.

Fiber Diagram

This problem requires methods that can separate the individual fibers that have no visible boundary between them. In CINEMA we employ so called dictionary-based methods that utilizes contextual information by learning a discriminative dictionary from user input. Providing input to a segmentation algorithm can be tedious, and therefore we investigate semi-supervised methods for obtaining high-quality segmentation results with limited user input. The obtained segments are used for individual fiber detection and characterization.

The developed analysis methods within CINEMA have general applicability, but are motivated by specific applications. Our experience is that this close link between methodology and application result in robust and user-friendly solutions. The developed methods are described in scientific publications, but we also plan to make implementations available for use in specific applications or further development.
19 JANUARY 2018