Statistical Methods for Quantification and Segmentation of Material Microstructure from 3D X-ray Tomography Image Data

The structure of a material at the microscale is of great importance for its macroscopic properties. Accurate quantification of the microstructure is therefore crucial for the development of new and improved materials and devices. X-ray computed tomography (CT) is one popular technique for non-destructive investigation of the interior of a sample at the microscale, generating 3D image data. This thesis is focused on developing methods for microstructure quantification and segmentation of such image data.

A statistical approach to image analysis is investigated. A framework is presented for statistical modelling of image intensities and gradient magnitudes, enabling direct quantification of structural parameters like volume fractions and material density. Through an improved description of blurred interfaces and the distribution of gradient magnitudes, the proposed model enable quantification of additional parameters which are not accessible with existing statistical image models, including interface areas, image resolution and the average radius of image features. The methods presented in this thesis are thus extending the scope of statistical models for quantification.

Although a lot of information about the sample can be extracted directly from the image data, a segmentation may still be required in some cases. A method for image segmentation is presented, combining the benefit of an improved interface description and gradient information in the new model with Markov random fields (MRF).

With continuous development in both software and hardware, the spatial and temporal resolution of modern imaging techniques is steadily increasing, often resulting in terabytes of data to be analyzed. This thesis addresses this issue by presenting a fast quantification method which can handle large datasets, and provides accurate and reproducible results. With no manual tuning of parameters, the proposed framework has great potential to e.g. form the basis of an automatic image analysis pipeline in industry applications.

The flexible framework can easily be adapted to better suit the dataset at hand, and this way be applicable for a range of imaging techniques and modalities. In addition to 3D X-ray CT data, the method is demonstrated for 2D images from a scanning electron microscope (SEM). The implementation of the tool for quantification and segmentation is made publicly available, contributing to making the statistical approach to image based structural analysis more accessible.