Improved segmentation of meteorite micro-CT images using local histograms

Abstract

In micro-CT images of meteorites individual components such as matrix, chondrules, Ca,Al-rich inclusions (CAIs), and opaque phases (metal and sulfide) are visually distinguishable. Automated classification of the components is desirable to deal with the large amount of data in a 3-D CT image. Classification by pixel intensity achieves a performance only 25% of the way from baseline to perfect. The poor performance is explained by an overlap in the range of intensities present in the different components. An improved method of semiautomated classification is presented, based on local histograms of the intensity. This achieves a performance 60% of the way from baseline to perfect.

Introduction

X-ray computerized tomography (CT) is a radiographic imaging modality that produces 3-D digital images of volumetric objects. The 3-D images are reconstructed from multiple 2-D images, each acquired by projecting X-rays through the object at a different orientation. X-ray CT images are composed of scalar-valued pixels (called voxels when the 3-D nature is important to stress) each of which records the reconstructed X-ray transmissibility of a location within the object. X-ray CT is widely used in medical and biological fields. In medicine the images encompass entire human organs, and have voxel dimensions around 1 mm. Micro-CT is the term used for smaller scale imaging setups, typically with a field-of-view dimension around 1 cm, and a voxel dimension of down to 1 μm. Micro-CT has already been extensively applied in medicine, biological sciences, materials science, engineering, and bio-engineering and has been recently extended to the field of geosciences and meteoritics (Van Geet et al., 2001, Ketcham and Carlson, 2001, Ebel and Rivers, 2007, Spinsby et al., 2008, Friedrich, 2008, Friedrich et al., 2008a, Friedrich et al., 2008b). When applied to rock samples the pixel values are determined by the proportions and distribution of different mineral types, and by porosity.

Asteroids are the building blocks of planets; hence, they predate the formation of Earth. Meteorites are fragments from asteroids that orbit the sun between Mars and Jupiter. One group of meteorites, called “primitive chondrites,” contain micrometer- to millimeter-sized components that formed in the protoplanetary disk of our Solar System. Study of these components is informative about the formation and evolution of the early Solar System before planet formation. Primitive chondrites contain the following: (i) ∼20–80% by volume of micrometer- to millimeter-sized, roundish “chondrules” primarily consisting of silicate minerals; (ii) ∼5–80% by volume of matrix primarily formed of micrometer-sized silicates; (iii) 0.1–3% by volume of Ca,Al-rich inclusions consisting of silicates and oxides; (iv) sulfides and metals (collectively opaques) that occur as minor abundances in chondrules and matrix (Brearley and Jones, 1998, McSween, 1999, Hezel et al., 2008, Hezel et al., 2010). These components formed separately under different conditions in the nebula, and were then aggregated in the meteorite parent bodies, some of which survived as asteroids.

Information on meteorite internal structure, such as proportions, spatial distribution, and relative orientation of components, is pertinent to a number of important problems in meteoritics (e.g., Ebel and Rivers, 2007). This information was for a long time accessible only through slow, difficult, and destructive methods. Since micro-CT can image the interior of a meteorite it has the potential to nondestructively access the required 3-D information (e.g., Ebel et al., 2008, Friedrich et al., 2008a, Friedrich et al., 2008b, Uesugi et al., 2010). Whether this potential can be realized depends on whether the different components of the meteorite can be accurately distinguished in the X-ray tomographic image. The experience of observers is that the components can be visually distinguished by careful inspection. An example is shown in Fig. 1.

Since the components are visually distinguishable, manual segmentations, such as that shown in Fig. 1, are possible but they are laborious. A single slice will take between 1 and 10 min even with bespoke software, a graphics tablet, and an experienced observer. A 3-D micro-CT image will typically have thousands of slices, making full manual segmentation of 3-D micro-CT images impractical.

The fact that manual segmentation is possible implies that there is information available in the images that could be used to automate the segmentation task. Many methods of automated image segmentation have been developed. Different methods are effective for different tasks. Edge-based methods focus on detecting particular image structures, for example, intensity discontinuities, that signal the boundaries of regions; while region-based methods focus on delineating regions with homogeneous textural properties (Haralick and Shapiro, 1985). Model-based approaches that make use of a strong prior expectation of the geometry of the desired segmentation (McInerney and Terzopoulos, 1996), while pixel classification approaches impose no geometric constraints on the produced segmentation, but simply consider each pixel in turn and independently classify it according to features computed from the local neighborhood (Harsanyi and Chang, 1994, Phung et al., 2005).

Pixel classification approaches are the simplest and fastest, and so are preferred if adequate to the task. The simplest feature that can be used as the basis for the classification is pixel intensity. There are three reasons why this feature would be suspected to be of little effectiveness for segmenting meteorite components in X-ray CT. First, in a simplified idealized meteorite, each component would exhibit a unique X-ray absorbance, and would therefore give rise to a unique image intensity. However, even with such an idealized meteorite, the image would exhibit a greater variety of pixel values than one per component due to the partial volume effect, which occurs when the spatial extent of a voxel contains a mixture of different components. Second, each component of a real meteorite exhibits a range of X-ray absorbances due to variations in mineral content, composition, and porosity. Crucially, the ranges for the different components overlap significantly, as shown in Fig. 2. Third, errors (known as imaging artifacts) in reconstructing the 3-D X-ray image from the 2-D X-ray projection images may lessen the accuracy with which image voxel intensities record the X-ray absorptions of locations in the sample (Remeysen and Swennen, 2006).

It is the aim of this study to present a new technique that in the long run and in combination with other techniques will allow accurate semiautomated component segmentation. As a first step we focus on segmenting matrix from what we will call “large components.” These comprise chondrules, CAIs, and dark inclusions. Chondrules are by far the most abundant of the large components. Large components are very similar in their densities and, hence, other techniques than we described here are required to further segment the large components. Opaques are not considered, as these are easily thresholded out from silicate and oxid material. We refer to the method as semiautomated, rather than automated, as the algorithms parameters need to be tuned on a subset of the data for any given meteorite.

This article is structured as follows. In Section 2 we introduce the image and ground truth data we use to assess different methods of semiautomated classification. In Section 3 we describe these methods. In Section 4 we report results on their effectiveness. In Section 5 we give a summary description of the winning method (local histogram-based classification) in a form suitable for reimplementation. In Section 6 conclusions are drawn.