See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/242597932

Computational methods for quantitative analysis of three-dimensional features

in geological systems

Article in Geosphere · January 2005

DOI: 10.1130/GES00001.1

CITATIONS READS
348 397

1 author:

Richard Ketcham
University of Texas at Austin
307 PUBLICATIONS 19,148 CITATIONS

SEE PROFILE

All content following this page was uploaded by Richard Ketcham on 21 May 2014.

The user has requested enhancement of the downloaded file.

 Computational methods for quantitative analysis of three-dimensional
features in geological specimens

Richard A. Ketcham
Department of Geological Sciences, University of Texas at Austin, Austin, Texas 78712, USA

ABSTRACT ing technologies such as high-resolution X-ray tain conditions are traced and described. As
computed tomography (HRXCT) (Carlson and with standard blob analysis, the primary dif-
Volumetric imaging techniques, such as Denison, 1992; Denison et al., 1997; Ketcham ficulty comes when a single region contains
high-resolution X-ray computed tomogra- and Carlson, 2001), synchrotron-based com- multiple contacting or impinging objects that
phy, allow three-dimensional measure- puted microtomography (Ikeda et al., 2004), must be separated to allow each to be consid-
ments of discrete objects inside solid sam- magnetic-resonance imaging (Gingras et al., ered and measured individually. Although
ples. This paper introduces a new computer 2002; Herrmann et al., 2002), and confocal some automated procedures have been devel-
program called BLOB3D that is designed microscopy (Bozhilov et al., 2003). Further- oped for relatively straightforward cases such
to allow efficient measurement of up to more, once such data have been acquired, an- as vesicles in basalt (Proussevitch and Sahag-
thousands of features in a single sample, alyzing them to obtain desired measurements ian, 2001), in general the variety of materials,
such as porphyroblasts, sediment grains, has often required surmounting an entirely textures, and forms in geological samples
clasts, and voids. BLOB3D implements an new set of challenges and obstacles: in many complicates development of a strictly algorith-
original suite of software methods, divided cases investigators need to exert significant ef- mic approach. Thus, BLOB3D is designed to
into three modules, which respectively en- fort to invent dedicated approaches and tools give the program operator primary control of
able the tasks of segmenting those regions that are often applicable only within narrow how the data are interpreted and measure-
in the data volume that correspond to the circumstances of data type, computer config- ments made, so that geological experience and
material of interest; separating touching or uration, and geological problem. insight can guide the analytical process.
impinging objects; and extracting measure- The barriers to routine acquisition and uti- The BLOB3D software was developed with
ments from the interpreted volume. Pro- lization of three-dimensional data describing the primary motivation of analyzing HRXCT
gram functions are demonstrated and ver- geological specimens are gradually being imagery of porphyroblastic metamorphic
ified with a set of phantoms used to test overcome as three-dimensional imaging tech- rocks, as described in the companion to this
determinations of particle size distributions nologies increase in capability and diminish in paper. However, it was also written to be ap-
and particle-particle contact orientations. cost. Many of these methods are nondestruc- plicable generally to any three-dimensional
tive, leaving samples intact for posterity or for data set containing identifiable discrete fea-
Keywords: X-ray tomography, quantitative subsequent analysis. Advances in computer tures; other applications to date have included
textural analysis, image processing, crystal- processing power and memory allow even HRXCT data of vesicles in meteoritic basalts
lization mechanisms, porphyroblasts. desktop machines to handle substantial data (Benedix et al., 2003; McCoy et al., 2002),
volumes, and there are many sources of soft- gold grains in ores (Kyle and Ketcham, 2003),
INTRODUCTION ware dedicated to three-dimensional visuali- and aggregate clasts in asphalt concretes (Ket-
zation. However, software tools that allow ex- cham and Shashidhar, 2001).
Petrographic inspection of rocks has largely traction of quantitative information of
been restricted to two-dimensional observa- geological interest are still required. HRXCT
tion, whether on the outer surface of an out- This paper and its companion1 respectively
crop or hand sample or from analysis of thin describe and utilize new software intended to HRXCT is the industrial adaptation of med-
sections. While two-dimensional measure- fill some of this gap, permitting efficient and ical CAT scanning. By using X-ray sources
ments can sometimes be used to obtain three- accurate identification and measurement of up with higher energies and smaller focal spots,
dimensional information using stereological to thousands of discrete objects in a volumet- smaller or finer-resolution detectors, and lon-
methods, model-based assumptions are often ric digital data set. The software, BLOB3D, ger acquisition times, industrial scanners
required, and the data are primarily general- implements a three-dimensional version of the achieve one to three orders of magnitude bet-
izations, describing rock-wide properties rath- image-processing tool of blob analysis, in ter resolution than medical devices while im-
er than particular features. Acquisition of di- which contiguous sets of pixels meeting cer- aging denser objects (Ketcham and Carlson,
rect three-dimensional information has 2001). The gray levels in HRXCT images cor-
1
See ‘‘Improved methods for quantitative analysis
typically been labor intensive and somewhat of three-dimensional porphyroblastic textures,’’ by
respond to linear attenuation coefficient,
expensive, whether by serial sectioning (Dan- R.A. Ketcham et al., Geosphere, v. 1, p. 42–59, doi: which is strongly correlated with density and
iel and Spear, 1999; Johnson, 1993) or imag- 10.1130/GES00002.1. is also a function of atomic number and X-ray

Geosphere; August 2005; v. 1; no. 1; p. 32–41; doi: 10.1130/GES00001.1; 6 figures; 1 table.

For permission to copy, contact [email protected]

32 q 2005 Geological Society of America
 COMPUTATIONAL ANALYSIS OF THREE-DIMENSIONAL FEATURES

energy. Individual computed tomography individual objects. The third, Extract, performs materials due to impurities, zoning, inclu-
(CT) images are termed slices, as they corre- a variety of measurements on the individual ob- sions, and microporosity, the grayscale for a
spond to what would be observed if a slice jects that have been separated, including size, given mineral will tend to have a range of val-
were extracted from the object being scanned, shape, orientation, and contact relationships. ues which may overlap with the ranges for
analogous to a slice from a loaf of bread or a Each module is described in detail below, and other phases with similar attenuation.
petrographic thin section. Because each slice an example sequence of operations for a set of In standard industrial CT practice, the prop-
image represents a finite thickness of material, porphyroblasts is shown in Figure 1. er grayscale value for defining a boundary be-
the pixels in HRXCT images are referred to tween two phases is the average of their two
as voxels, or volume elements. By acquiring Segment mean end-member grayscales (ASTM [Amer-
a contiguous series of slices, data describing ican Society for Testing and Materials], 1992).
an entire volume can be obtained. The Segment module performs the image- In a polymineralic rock, simple thresholding
processing task of segmentation, determining based on this value is not practical when other
BLOB3D DATA PROCESSING which voxels in the data set correspond to a phases have overlapping ranges of grayscales
particular material of interest (Figs. 1A and (Fig. 2). Three grayscale-to-binary filters were
Previous textural studies using HRXCT 1B). Three classes of algorithms are available, created to achieve accurate segmentation by
(Carlson and Denison, 1992; Denison and grouped by the form of their input and output utilizing connectivity information. The Seeded
Carlson, 1997; Hirsch, 2000) included the cre- data. The first class is grayscale-to-grayscale Threshold filter first selects all voxels that lie
ation of software tools to process scan imag- algorithms that may be used to clarify the im- within a specified ‘‘seed’’ grayscale range, fol-
ery using a two-stage procedure to measure age data through noise reduction or edge en- lowed by all voxels that lie in a second, ex-
the size and location of garnet porphyroblasts. hancement. The second class, grayscale-to- panded grayscale range and are connected to
Each slice image was first interpreted individ- binary algorithms, transform the processed seed voxels via a path of any length. The Ex-
ually by superposing a circle on each porphy- data into binary (two-component) data, defin- panding Threshold filter is similar, except that
roblast to represent its location and cross- ing the material of interest versus everything it limits the voxels selected in the second stage
sectional area. After all of the slices were else. A third set of binary-to-binary algo- to those within a finite distance of first-stage
processed in two dimensions, the three- rithms can be used for clean-up tasks such as voxels. The Expanding Seeded Threshold
dimensional porphyroblast volumes were re- removing small isolated groups of voxels. All combines these by first executing a seeded
constructed by stacking circles from succes- Segment filters are implemented as three- threshold and then allowing a final expansion
sive slices and approximating them as spheres dimensional operators. For example, a Gauss- step. These algorithms allow selection of
(Denison et al., 1997; Hirsch, 2000). Data pro- ian3D filter (a grayscale-to-grayscale algo- phases based on their end-member grayscales,
cessing with these tools was very labor inten- rithm) uses a Gaussian n 3 n 3 m kernel to and then allow limited expansion into their
sive, typically requiring weeks per sample. In smooth the volume data, where n is the kernel blurred boundaries.
addition, data quality and flexibility were lim- dimension in the slice plane and m is the
ited by assuming an exclusively spherical between-slice dimension, which may be the Separate
crystal geometry and only offering a single same as n or different if the within-slice pixel
planar view of the data to inform interpreta- spacing is not equal to the spacing between The Separate module proceeds by tra-
tion. Publicly available freeware image pro- slices. The filters are implemented using object- versing the data volume and presenting the
cessing programs, such as NIH Image and oriented software design, allowing new mod- program operator with each set of three-
ImageJ, have also been effectively used to in- ules to be written based on a template and dimensionally connected voxels (blob) en-
terpret two-dimensional data and provide lim- added to the program as needs arise. countered (Figs. 1C and 1D; Fig. 3). The
ited three-dimensional information (e.g., Hig- The grayscale-to-binary step in Segment operator then decides whether one or more
gins, 2000; Koeberl et al., 2002). exerts the greatest influence on the overall re- objects are represented, based on descriptive
BLOB3D allows this processing to be done sults. For polymineralic rocks, specialized al- statistics, the shape of the blob, and refer-
in three dimensions, increasing both efficiency gorithms were created to account for the finite ence to the original image data.
and accuracy. The program begins by assem- resolution of HRXCT data, which has three
bling a contiguous series of slice images, typ- consequences. First, the partial-volume effect Separation Methods
ically a set of TIFF files, into a three-dimen- causes blurring of phase boundaries: because Several tools are available for subdividing
sional matrix representing the entire data voxels are discrete, they may contain more a voxel set into its component objects. The
volume (this is also known as a data brick or than one phase, resulting in a grayscale value simplest, but most labor intensive, is manual
voxel brick). Data dimensions are defined in that is a weighted average between the ideal definition of one or more cutting planes using
terms of the pixel and slice spacing. These end-member values for each phase. Second, mouse clicks on the blob surface. Three other
need not be equal, as the program will appro- the averaging brought about by the nonzero tools work by removing certain voxels to
priately account for noncubic data elements. size of the X-ray focal spot and/or detectors achieve separate objects and then replacing
The data-processing task is divided into three causes material in one voxel to affect sur- the removed voxels while maintaining the sep-
steps, each of which is implemented in its own rounding ones. As a result, the grayscale tran- aration. The Erode/Dilate function is suited to
program module. The first, Segment, is used to sition between two adjacent objects of differ- cases where the cross-sectional area of a con-
define which voxels in the data volume corre- ing X-ray attenuation, rather than being sharp, tact is less than the maximum cross-sectional
spond to each material of interest. The second, characteristically spans several voxel widths area of the objects (for example, two slightly
Separate, takes all three-dimensionally con- (typically 2–4 in this study). Third, because of impinging spheres). Successive erosion oper-
nected sets of segmented voxels of a particular the limited signal-to-noise ratio of HRXCT ations, removing shells of voxels from the
material and allows them to be cut apart into data and the frequent variability of geological outer surface, can cause the contact to neck

Geosphere, August 2005 33

 R.A. KETCHAM

Figure 1. Example of BLOB3D processing for cluster of garnet porphyroblasts. Because garnet is the only porphyroblast, it can be
segmented using a simple grayscale threshold (A, B). When inspected in two dimensions (B) and three dimensions (C) it is apparent
that the cluster consists of four individual crystals. Separation was accomplished (D) using erosion-dilation operations. To estimate the
extended volume and center of nucleation for lower-right crystal (E), a sphere primitive was fit to outer surface (F) by excluding from
consideration points in contact with other garnet.

34 Geosphere, August 2005

 COMPUTATIONAL ANALYSIS OF THREE-DIMENSIONAL FEATURES

Figure 2. Illustration of seeded and expand-

ing threshold algorithms, as applied to a
garnet-kyanite schist (MD from companion
paper [see text footnote one]). Part A shows
a subsection of computed tomography (CT)
data, below which a graph shows CT gray-
scale values along traverse. Mineral phases,
from brightest to darkest, are garnet (grt),
kyanite (ky), biotite (bio), muscovite, pla-
gioclase, and quartz. For this data set the
appropriate threshold value for correct di-
mensional measurement of garnet is 200,
the midpoint between garnet (240) and ma-
trix (160 overall). Although this value over-
laps with kyanite and occasionally biotite,
appropriate segmentation is accomplished
using connectivity information. Garnet is
segmented by first defining a seed range of
218–255 (light gray in graph, light pink in
B) unique to garnet, followed by an unlim-
ited neighborhood expansion from seed
voxels into a mean threshold range of 212–
217 (medium gray in graph, medium pink
in C) and finally a limited one-voxel expan-
sion into the range 200–211 (dark gray in
graph, dark pink in D) to capture boundary
material. Uppermost image is interpolated
for better viewing, lower images show ac-
tual voxel data, upon which calculations are
performed.

and eventually break, separating the objects

(Proussevitch and Sahagian, 2001). The erod-
ed material is then added back to each object
using dilation operations. The Grayscale Seed
1 Dilate function is similar, but replaces the
erosion step with a search for local grayscale
maxima (or minima) and uses each isolated
grouping as a seed from which to commence
dilation. This method is intended for situations
where objects are touching or in close prox-
imity but not impinging, and thus tend to have
a slight grayscale barrier between them. Fi-
nally, the Grayscale Sea Level Fall function
implements a version of the watershed algo-
rithm. A threshold value (‘‘sea level’’) is ini-
tially set to encompass all of the data, and is
slowly lowered to first uncover clusters of
similar voxels (‘‘islands’’) that grow as the
threshold continues to fall. It is intended for
clusters of small grains whose peak gray lev-
els are variably affected by blurring.

Manual versus Automated Separation

Whereas the three-dimensional processing
described by Proussevitch and Sahagian
(2001) for interpreting HRXCT data of vesi-
cles in basalt is largely automated, every de-
cision in the Separate module is made by the

Geosphere, August 2005 35

 R.A. KETCHAM

Figure 3. Example of typical processing operations in Separate module. Part A shows a sample of computed tomography (CT) data of
a cluster of spheres; field of view is 70 mm. A three-dimensional volume searching algorithm finds all voxels in contact, up to limits
imposed by computer memory (B); faces truncated by the algorithm are marked in red. ‘‘Stair steps’’ on upper and lower portions of
spheres are caused by 3:1 voxel aspect ratio (inter-slice vs. inter-pixel spacing). An erosion/dilation operator successfully separates most
of spheres (C). Those contacting truncated faces have their processing postponed (D), allowing interior spheres to be processed.

program operator. This design choice was sands of objects are involved. However, the characterized or well tested by the measure-
made because of the difficulty of defining a speed of data processing is an order-of- ments that choices are based on, and for the
set of prearranged operations and criteria that magnitude improvement over the methods de- resulting poor data to be undetected and
fully account for the range of textures en- scribed by Denison et al. (1997). interpreted.
countered in natural geological samples with Further time savings are anticipated as im-
the reliability of a human petrologist. Features provements in automation are made. As cur- Example of Separate Operation
such as crystal habit, alteration and resorption, rently envisioned, we anticipate adding the ca- An example of the Separate module applied
fractures, inclusions, and zoning affect the ap- pability to make certain easy decisions by to a phantom created for verification of par-
pearance of porphyroblasts in HRXCT data default, based on passing a customizable (and ticle size distribution (grading) is shown in
and their three-dimensional shapes, detracting perhaps learnable) set of criteria: for example, Figure 3. The phantom consists of a set of
from the ideal forms that purely algorithmic always accepting nearly perfect spheres, while glass spheres of various sizes, bound with as-
procedures may be based upon. Because of its leaving harder decisions to the user. Some au- phalt. Figure 3A shows a typical slice of the
reliance on a human operator, Separate is the tomated separation processing may also be at- scan data. Because the network of contacting
most time-consuming BLOB3D step, typically tempted. Ultimately, the danger to be avoided spheres spans the entire sample, the set of
requiring several seconds per individual object is having the software make poor decisions three-dimensionally connected voxels that
separated, and thus several hours when thou- because the imagery includes features not well comprises them likewise spans the entire data

36 Geosphere, August 2005

 COMPUTATIONAL ANALYSIS OF THREE-DIMENSIONAL FEATURES

set, which would overtax computer resources. center of nucleation and the extended volume by taking the vector sum of the normal to each
To solve this problem, the volume-searching (i.e., the volume in absence of impingement contacting voxel face, weighted by face sur-
algorithm automatically limits the size of the by neighboring porphyroblasts). Image pro- face area; and by fitting a plane to the set of
growing blob region; those faces that are cut cessing described by Denison et al. (1997) and points on the midpoints of each contacting
off are denoted with red lines (Fig. 3B). A Hirsch (2000) accounted for impingement by voxel face. These two approaches are com-
good separation into individual objects is ob- allowing the operator to adjust the size and pared below using the orientation of the plane
tained from two erosion steps followed by di- location of the circle superposed over each normal to the line connecting the center of
lation (Fig. 3C). In Figure 3D, the operator porphyroblast in each slice to compensate for mass of each object, which is also calculated.
has postponed the processing of all voxels in missing material. A weakness of this method is
subblobs intersected by the cutoff planes, that detection and delineation of impingement Software Availability
leaving only spheres that are complete or trun- is not always obvious on a two-dimensional
cated by the edge of the data set. The operator section. To account for impingement in The BLOB3D software is written in the
would then accept, reject, or further process BLOB3D, the Extract module includes two IDL (Interactive Data Language) program-
these blobs, after which the traversal algo- three-dimensional algorithms that are based ming language (Research Systems, Inc., Boul-
rithm would find the next set of unprocessed on primitive fitting: finding idealized shapes der, Colorado). Programs developed in IDL
three-dimensionally connected voxels. Voxels (spheres or ellipsoids) that best match the por- have the potential to run on any platform
that are postponed are returned to the pool phyroblast surface based on modified least- (Windows, Macintosh, Linux, Unix), although
available for searching, so objects that were squares criteria. The Impinged Surface algo- minor incompatibilities may need to be re-
truncated will eventually appear complete and rithm excludes all surface points that are in solved; the current version is maintained on
suitable for processing. contact with another porphyroblast of the Windows and Linux. Most current desktop
Despite efforts to create an accurate seg- same type, or are at the edge of the data vol- computers are capable of running it, with the
mentation during the first stage of processing, ume. The Incomplete Surface algorithm is in- principal requirement being sufficient com-
sometimes unwanted voxels may be included. tended for cases where other factors impact a puter memory to store multiple copies of the
When such voxels are recognized during pro- porphyroblast surface, such as uneven resorp- volumetric data set; typically 1–2 Gb are re-
cessing with Separate they can be rejected by tion and alteration, and it works by restricting quired. Persons interested in obtaining or us-
the operator, returning them to the pool of un- the deleterious effects of a percentage of the ing the software should contact the author.
defined voxels that may be classified as a dif- worst-fitting points. The maximum percentage
ferent phase during subsequent processing. of restricted points inside and outside the VERIFICATION
Thus, the Separate step also constitutes a final primitive boundary, and the maximum sum-
segmentation filter. of-squares values that these points can A series of tests were done to verify that
A volume may be processed for up to seven achieve, are parameters selected by the oper- the program produces reliable results and to
components by successive iterations of run- ator. An example of a garnet described using provide examples of various practical matters
ning the Segment and Separate modules. Be- a sphere with the Impinged Surface algorithm of program operation. The first test verified
cause the edges of high-attenuation phases is shown in Figures 1E and 1F. that the program could reproduce the particle-
will have grayscale values that overlap with Of the primitive-fitting methods, the Im- size characteristics of a sample, including ac-
less attenuating materials, processing optimal- pinged Surface algorithm applied to spheres commodating large-aspect-ratio data voxels
ly proceeds from greatest to least departure tends to give the best-constrained results. The and utilizing primitive fitting to correct for
from mean matrix grayscale (for example, transition from sphere to ellipsoid primitive particle cut-off. Another pair of tests verified
from brightest to dimmest phase). corresponds to a jump from four to nine de- the calculations of contact orientation.
grees of freedom (from center position and ra-
Extract dius to center position, three radii, and three Particle Size and Gradation
angular displacements to align the major
Once one or more phases are segmented axes), greatly increasing the range of possible The phantom for the first test consists of a
and separated, the Extract module is used to solutions. The Incomplete Surface algorithm mixture of glass spheres of various sizes and
extract data of interest for each object. A va- uses no a priori information to select the parts known grading bound with asphalt (Ketcham
riety of measurements is available, including of a surface to exclude. These factors, indi- and Shashidhar, 2001). The manufacturer-
center of mass, volume, surface area, aspect vidually and particularly in concert, can oc- specified sphere sizes and the proportions used
ratio, and orientation. The first three are cal- casionally result in nonintuitive primitive fits. to create the phantom are given in Table 1. A
culated simply from the set of voxels com- In practice, an operator inspects the results of section of the phantom was scanned at the
prising an object, while shape and orientation several primitive fits, adjusting settable pa- HRXCT facility at the University of Texas at
are estimated by least-squares fitting of an el- rameters as appropriate, and then allows the Austin (UTCT), using the high-resolution sub-
lipsoid to the object outer surface. The Extract entire data volume to be processed with a sin- system described by Ketcham and Carlson
module also measures contacts between par- gle set of parameters. (2001). Slices were acquired at 0.4 mm inter-
ticles, including contact normal orientations vals with a 1024 3 1024 reconstruction of a
and surface areas. Contact Measurements 140 mm field of view, resulting in an inter-
Contacts can be measured for any touching pixel spacing of 0.137 mm. The data were
Primitive Fitting pair of objects. The areas of surfaces are mea- thus much lower resolution in the z dimension
Two more specialized quantities necessary sured by summing the areas of the voxel faces than in x-y (e.g., Figures 3B–3D), in part to
for textural analysis applied to study of crystal that make up the boundary between them. test the program’s ability to deal with inequant
nucleation and growth mechanisms are the Contact orientation is calculated in two ways: voxels. An additional 1.18-mm-size fraction

Geosphere, August 2005 37

 R.A. KETCHAM

TABLE 1. RESULTS OF ANALYSIS OF GRADING PHANTOM be judged using Figure 4, which shows the
Manufacturer Original Enclosed Side-intersecting Total relationship between measured volume (i.e.,
specified grading grading number of voxels subtended) and primitive
diameter (wt%) N Measured Grading N Fitted diameter (wt%)
(mm) diameter (wt%) (mm) volume. Enclosed spheres fall on the 1:1 line,
(mm) while truncated spheres fall to the left. Among
2.3–2.6 18.2 5660 2.22 6 0.11 25.0 862 2.34 6 0.19 19.5 the large sphere classes, the size distribution
5 36.4 946 4.77 6 0.05 41.5 359 4.87 6 0.23 35.9 remains unchanged until the truncated volume
10 25.4 43 9.61 6 0.10 15.4 43 9.69 6 0.43 17.3
12.5 20 24 12.30 6 0.41 18.1 57 12.62 6 0.48 27.2 falls below ;100 mm3, or 10%–20% of the
full volume. Below this point the spread in
values increases, with a slight bias toward
was also included in the phantom, but the slice also be due to conservative segmentation, as higher values, although in some cases appar-
spacing was insufficient to allow reliable de- parameters were chosen to minimize the num- ently reasonable fits are obtained for spheres
tection and measurement of it, so it was ex- ber of unresolvable 1.18 mm spheres that were as much as 99% truncated. These trends are
cluded from the analysis. included. Although a large statistical sample also observed among the smaller size classes,
A 512 3 512 subvolume of the complete was taken, the confined volume leads to a bias with the onset of increased variation occurring
once a sphere is more than 50% truncated.
CT data set was processed by first segmenting in the estimate of grading, as larger spheres
The classification of truncated spheres by
and separating the 5-mm-size class, which had are more likely to be truncated by the edges
primitive fitting allows their volumes to be in-
a slightly higher attenuation coefficient than of the subvolume, and are thus underrepre-
cluded in the grading analysis (Table 1, col-
the other classes. The two larger size classes sented in the grading based on complete
umn 8). The two small size classes match the
were processed as a single set and later dif- spheres (Table 1, Column 5).
original grading to within 1.3%. The larger
ferentiated by size; the final component to be Spheres that were truncated by one or more
two size classes do not match the grading,
processed was the ;2.5-mm-size class. In all, faces of the subvolume were processed using
probably because of the limited sample size
8356 spheres were measured. Of these, 6673 the impinged-sphere primitive fitting method
and consequent small number of such spheres
were entirely enclosed in the subvolume, re- described previously. A total of 1321 addi-
analyzed. However, their combined grading
sulting in the size and grading listed in col- tional spheres was successfully fitted; the re-
matches the original grading of these two
umns 3 through 5 of Table 1. Measured sizes maining 362 were omitted from further con-
phases to within 0.9%.
are slightly smaller than the manufacturer- sideration, as they comprised a volumetrically
specified sizes. Caliper measurements made insignificant fraction (,0.4% of the subvolu- Particle-Particle Contact Orientations
afterwards indicated that the spheres are often me). Results for each population are shown in
slightly irregular and slightly smaller than the columns 6 and 7 of Table 1, and the effec- Particle-particle contact orientations are not
specifications. Part of the discrepancy may tiveness of the primitive fitting procedure can a widely used type of information for geolog-

Figure 4. Volume of spheres from grading phantom, entirely enclosed and truncated by edge of data volume. Volumes of the latter were
estimated using Impinged Sphere algorithm.

38 Geosphere, August 2005

 COMPUTATIONAL ANALYSIS OF THREE-DIMENSIONAL FEATURES

Figure 5. Cubic (A) and hexagonal (B) packing phantoms, and representative computed tomography (CT) slices. CT slice fields of view:
154 mm for cubic phantom (C), 134.5 mm for hexagonal phantom (D). Spheres are roughly 18.5 mm in diameter.

ical investigations, perhaps in part because ing whether porphyroblasts may behave as hexagonal packing each sphere may be in con-
they have not been easy to obtain in three di- rigid particles in a deforming schistose matrix tact with up to twelve others, six along the hor-
mensions. However, such data have a wide undergoing asymmetric shear (Ketcham, izontal plane 608 apart, and three each 54.78
range of possible uses, including: evaluating 2005). above and below the plane, each 1208 apart.
the mechanical load-bearing properties of The phantoms used for these tests were made Because the spheres were not perfectly uniform
matrix-supported conglomerates (e.g., Ket- by gluing relatively uniformly sized (;9.25 in their size and shape, geometric configura-
cham and Shashidhar, 2001); estimating the mm radius) glass spheres into simple cubic and tions were not perfect, as shown in the scan
gravitational vector direction in differentiating hexagonal close packing arrangements (Fig. 5). images, and contacts range from genuinely
metal-sulfide assemblages in extraterrestrial In cubic packing, each sphere is in contact with touching to being separated by some thickness
impact melts (Benedix et al., 2003); and test- up to six others in the orthogonal directions. In of glue. However, the spheres are close enough

Geosphere, August 2005 39

 R.A. KETCHAM

Figure 6. Equal-angle stereonets of determinations from contact analysis of cubic (A–C; N 5 538) and hexagonal (D–F; N 5 793)
packing phantoms. Bubble size reflects contact surface area.

to uniform that the center-to-center vector of a indistinguishable results (Figs. 6B and 6C). In lations on the horizontal plane once again
sphere to its neighbor provides a reasonable test comparison with the center-to-center data, the show less dispersion off of the plane than
of the results provided by contact measurement contact-normal measurements have a relative within it, and populations near orthogonal
methods. The phantoms were scanned on the lack of dispersion off of the horizontal plane, planes have slight biases toward those planes.
UTCT high-resolution subsystem, with 1.0- and away from the vertical direction. Also, the The three horizontal clusters should be spaced
mm-thick slices spaced at 0.8 mm, with each clusters on the horizontal planes are shifted 2– at 608, whereas the measured cluster spacings
512 3 512 image encompassing a field of view 3 degrees toward the orthogonal axes. These are at 558, 608, and 658. The two algorithms
of 134.5 mm for hexagonal packing and 154 reflect the biasing effects of the orthogonal are similarly affected; the median error (in-
mm for cubic packing. Once again, the thick faces of the data voxels, which can affect sep- ferred as net angular divergence from the
slices lead to a voxel aspect ratio of ;3:1. arations during BLOB3D processing in much center-to-center vector) among nonhorizontal
Processing results for the cubic packing the same way as a cleavage plane. This effect orientations is 3.58 for the best-fit plane al-
phantom (Figs. 6A–6C) show the expected is exacerbated by the voxel anisotropy, which gorithm and 3.28 for the mean face normal
configuration of contact orientations, with all makes stepping from one horizontal plane to determination. The nonhorizontal clusters are
populations 908 apart. The spread evident in another less likely. also biased somewhat toward the vertical di-
center-to-center vectors (Fig. 6A) is a conse- The hexagonal packing phantom (Figs. 6D– rection, although the best-fit plane algorithm
quence of imperfections in the construction of 6F) shows similar effects. Again, the results has less divergence (4.28 versus 7.18 median
the phantom. The two contact measurement are in general accordance with expectation, error). The small-surface-area outliers in Fig-
algorithms (the vector average of normals to with the spread in center-to-center vectors at- ures 6E–6F reflect cases where spheres are
contacting voxel faces and the best-fit plane tributable to phantom imperfections. Among connected only by a small and irregular inter-
to the center points of these faces) give nearly the contact normal measurements, the popu- face of glue.

40 Geosphere, August 2005

 COMPUTATIONAL ANALYSIS OF THREE-DIMENSIONAL FEATURES

CONCLUSIONS ods and analytical procedures: Philadelphia, ASTM, textures [Ph.D. thesis]: Austin, Texas, The University
p. 690–713. of Texas, 184 p.
Benedix, G.K., Ketcham, R.A., McCoy, T.J., and Wilson, Ikeda, S., Nakano, T., Tsuchiyama, A., Uesegi, K., Suzuki,
The new tools introduced here greatly im- L., 2003, Vesiculation in ordinary chondrites due to Y., Nakamura, K., Nakashima, Y., and Yoshida, H.,
impact melting: The ‘‘PAT’’ 91501 answers, in Lunar 2004, Nondestructive three-dimensional element-
prove the quality and ease of acquisition of concentration mapping of a Cs-doped partially molten
and Planetary Science Conference XXXIV, abs. 1947,
three-dimensional measurements from volu- CD-ROM. granite by X-ray computed tomography using syn-
metric data sets. While making heavy use of Bozhilov, K.N., Green, H.W.I., and Dobrzhinetskaya, L.F., chrotron radiation: American Mineralogist, v. 89,
2003, Quantitative 3D measurement of ilmenite abun- p. 1304–1312.
operator decision-making, BLOB3D’s pro- dance in Alpe Arami olivine by confocal microscopy: Johnson, S.E., 1993, Unravelling the spirals; a serial thin-
cessing is fast enough (typically ,1–10 sec- Confirmation of high-pressure origin: American Min- section study and three-dimensional computer-aided
onds per object) that it can measure thousands eralogist, v. 88, p. 596–603. reconstruction of spiral-shaped inclusion trails in gar-
Carlson, W.D., and Denison, C., 1992, Mechanisms of porphy- net porphyroblasts: Journal of Metamorphic Geology,
of objects in a matter of hours, providing ex- roblast crystallization: Results from high-resolution com- v. 11, no. 5, p. 621–634.
cellent statistical sampling. Furthermore, it puted X-ray tomography: Science, v. 257, Ketcham, R.A., and Carlson, W.D., 2001, Acquisition, op-
makes many three-dimensional measurements p. 1236–1239. timization and interpretation of X-ray computed to-
Daniel, C.G., and Spear, F.S., 1999, The clustered nucle- mographic imagery: Applications to the geosciences:
possible that heretofore may have been con- ation and growth processes of garnet in regional meta- Computers & Geosciences, v. 27, p. 381–400, doi:
sidered impossible or impractical. Although morphic rocks from north-west Connecticut, USA: 10.1016/S0098-3004(00)00116-3.
optimized specifically for HRXCT imagery, Journal of Metamorphic Geology, v. 17, p. 503–520, Ketcham, R.A., and Shashidhar, N., 2001, Quantitative analysis
doi: 10.1046/j.1525-1314.1999.00213.x. of 3-D images of asphalt concrete, in Transportation Re-
BLOB3D can potentially be used with any Denison, C., and Carlson, W.D., 1997, Three-dimensional search Board 80th Annual Meeting, Washington, D.C.,
type of volumetric data. quantitative textural analysis of metamorphic rocks preprint CD-ROM, paper no. 01-0321.
using high-resolution computed X-ray tomography: Ketcham, R.A., 2005, Three-dimensional grain fabric mea-
Part II. Application to natural samples: Journal of surements using high-resolution X-ray computed to-
ACKNOWLEDGMENTS mography: Journal of Structural Geology (in press).
Metamorphic Geology, v. 15, no. 1, p. 45–57, doi:
10.1111/j.1525-1314.1997.00007.x. Koeberl, C., Denison, C., Ketcham, R.A., and Reimold,
This research was supported by National Science Denison, C., Carlson, W.D., and Ketcham, R.A., 1997, W.U., 2002, High-resolution X-ray computed tomog-
Foundation grants EAR-9902682 and EAR- Three-dimensional quantitative textural analysis of raphy of impactites: Journal of Geophysical Research,
0113480, and by National Research Council Trans- metamorphic rocks using high-resolution computed X- v. 107, no. E10, p. 5089–5097, doi: 10.1029/
portation Research Board IDEA Program contract ray tomography; Part I. Methods and techniques: Jour- 2001JE001833.
number NCHRP-64. Programming contributions at nal of Metamorphic Geology, v. 15, no. 1, p. 29–44, Kyle, J.R., and Ketcham, R.A., 2003, In-situ distribution of
early stages were made by D. Hirsch, and recent doi: 10.1111/j.1525-1314.1997.00006.x. gold ores using high-resolution X-ray computed to-
enhancements were made by W.P. Watson. Program Gingras, M.K., MacMillan, B., Balcom, B.J., Saunders, T., mography: Economic Geology, v. 98, p. 1697–1701.
and Pemberton, S.G., 2002, Using magnetic resonance McCoy, T.J., Ketcham, R.A., Benedix, G.K., Carlson, W.D.,
testing and utilization for data acquisition were per- and Wilson, L., 2002, Vesicular basalts from asteroids:
imaging and petrographic techniques to understand the
formed by S. Anderson, T. Eiting, and C. Meth. Il- textural attributes of porosity distribution in Macaron- Clues to processes in their parent magmas, in Lunar
lustrations were made with help from C. Meth. The ichnus-burrowed sandstone: Journal of Sedimentary and Planetary Science XXXIII, abs. 1213; CD-ROM.
manuscript was improved thanks to helpful reviews Research, Section A, Sedimentary Petrology and Pro- Proussevitch, A., and Sahagian, D., 2001, Recognition and
by M. Rivers and M. Higgins. cesses, v. 72, no. 4, p. 552–558. separation of discrete objects within complex 3D vox-
Herrmann, K.-H., Pohlmeier, A., Gembris, D., and Vereecken, elized structures: Computers & Geosciences, v. 27,
H., 2002, Three-dimensional imaging of pore water dif- p. 441–454, doi: 10.1016/S0098-3004(00)00141-2.
REFERENCES CITED fusion and motion in porous media by nuclear magnetic
resonance imaging: Journal of Hydrology, v. 267, MANUSCRIPT RECEIVED BY THE SOCIETY 15 DECEMBER 2004
ASTM (American Society for Testing and Materials), 1992, p. 244–257, doi: 10.1016/S0022-1694(02)00154-3. REVISED MANUSCRIPT RECEIVED 3 JUNE 2005
Standard guide for computed tomography (CT) im- Higgins, M.D., 2000, Measurement of crystal size distri- MANUSCRIPT ACCEPTED 14 JUNE 2005
aging, ASTM designation E 1441–92a, in 1992 annual butions: American Mineralogist, v. 85, p. 1105–1116.
book of ASTM standards, section 3 metals test meth- Hirsch, D.M., 2000, Quantitative studies of porphyroblastic Printed in the USA

Geosphere, August 2005 41

View publication stats