Volume Visualization

ARIE E. KAUFMAN
Computer Science Department, State University of New York at Stony Brook ^ari@[Link]&

Volume visualization is a method of ex- techniques for modeling, manipulation,

tracting information from volumetric and visualization, an approach called
datasets through interactive graphics volume graphics [Kaufman et al. 1993].
and imaging, and is concerned with the Volumetric data is typically a set S of
representation, manipulation, and ren- samples (x, y, z, v), representing the
dering of these datasets [Gallagher value v of some property of the data at a
1995; Kaufman 1991; Rosenblum 1994]. 3D location (x, y, z). If v is simply a 0 or
Volume data are 3D entities that may a 1, with 0 indicating background and 1
have information inside them, may not indicating the object, the data is called
consist of surfaces and edges, or may be binary data. The data may instead be
too voluminous to be represented geo- multivalued, with v representing some
metrically. Volume visualization encom- measurable property of the data, such
passes an array of techniques for peer- as density, color, heat, or pressure. The
ing inside the dataset and for value v may even be a vector, represent-
interactively extracting meaningful in- ing, for example, velocity at each loca-
formation from it using transforma- tion.
tions, cuts, segmentation, translucency, In general, the samples may be taken
measurements, and the like. The pri- at random locations in space, but in
mary sources of volume data are three: many cases S is isotropic, containing
sampled data of real objects or phenom-
samples taken at regularly spaced in-
ena, computed data produced by a com-
tervals along three orthogonal axes.
puter simulation, and modeled data
Since S is defined on a regular grid, a
generated from a geometric model. Ex-
3D array (called volume buffer, cubic
amples of applications generating sam-
pled data are medical imaging (e.g., CT, frame buffer, 3D raster) is typically used
MRI), biology (e.g., confocal microsco- to store the values. S is therefore
py), geoscience (e.g., seismic measure- referred to as the array of values
ments), industry (e.g., nondestructive S(x, y, z), which is defined only at grid
inspection), and chemistry (e.g., elec- locations. A function may be defined to
tron density maps) [Kaufman 1991]. describe the value at any continuous
Some examples of applications generat- location by approximating v at a loca-
ing computed datasets, typically by run- tion (x, y, z) using some interpolation
ning a simulation on a supercomputer, function to S, such as zero-order (near-
are meteorology (e.g., storm prediction), est-neighbor), piecewise function known
computational fluid dynamics (e.g., wa- as first-order (trilinear), or higher-order
ter flow), and materials science (e.g., interpolation. The region of constant
new materials). Recently, many tradi- value that surrounds each sample in
tional computer graphics applications, zero-order interpolation is known as a
such as computer-aided design and volume cell (voxel for short), with each
flight simulation [Cohen and Shaked voxel being a rectangular cuboid having
1993; Kaufman et al. 1993], have been six faces, twelve edges, and eight cor-
exploiting the advantages of volumetric ners. The terms, voxel, grid location,

Copyright © 1996, CRC Press.

ACM Computing Surveys, Vol. 28, No. 1, March 1996

166 • Arie E. Kaufman

and sample points, are often used inter- Volume rendering can be divided into
changeably. three main approaches: object-order,
In addition to regular grids, rectilin- image-order, and domain methods. In
ear, curvilinear, and unstructured grids object-order methods the contribution of
are employed. In a rectilinear grid the each voxel to the screen pixels is calcu-
cells are axis-aligned, but grid spacings lated and the combined contribution
along the axes may be arbitrary. When yields the final image (e.g., Kaufman
such a grid has been nonlinearly trans- [1991, pp. 125–134 and 154–159]). One
formed while preserving the grid topol- such method is splatting [Kaufman
ogy, the grid becomes curvilinear. Usu- 1991, pp. 144–153], in which an image-
ally, the rectilinear grid defining the plane footprint for each voxel is used to
logical organization is called computa- spread the voxel energy onto a neigh-
tional space and the curvilinear grid is borhood of pixels. In image-order meth-
called physical space. In an unstruc- ods such as ray casting, rays are cast
tured grid, there is no explicit or im- from the screen pixels into the volume,
plicit grid topology. Cells may be, for and the contributions of voxels along a
example, tetrahedra, hexahedra, pyra- ray are calculated and used to color the
mids, or prisms, where tetrahedral corresponding pixel (e.g., Kaufman
grids are particularly popular. Unstruc- [1991, pp. 135–143, 111–115, and 160–
tured grids are common for scattered 168]). In domain methods the spatial
data, finite-element/volume analysis, data is transformed into an alternative
and computational fluid dynamics. domain, such as compression (e.g., Yeo
Over the years many techniques have and Liu [1995]), wavelet (e.g., Muraki
been developed to visualize volumetric [1993]), or frequency domain (e.g.,
data. Most of the early methods (e.g., Malzbender [1993]), from which a pro-
marching cubes [Kaufman 1991, pp. 64– jection is directly generated.
71]) involved approximating a surface Although volumetric representation
contained within the data using geomet- and visualization seem more natural for
ric primitives, since methods for dis- sampled or computed datasets, their ad-
playing geometric primitives were al- vantages have also been attracting tra-
ready well-established. When such ditional computer graphics applications
surface rendering is used, a dimension that deal with synthetic scenes repre-
of information is essentially lost. In ad- sented by geometric models. In this
dition, adequate approximations may emerging approach, called volume
require an excessive number of primi- graphics, the geometric model is voxel-
tives. Also, amorphous phenomena, ized (3D scan-converted) into a set of
such as clouds, fog, and fire, cannot be voxels that “best” approximate the
represented adequately using surfaces. model. Each of these voxels is stored in
In response to these, volume rendering the volume buffer along with the voxel
techniques have been developed, at- pre-computed view-independent at-
tempting to capture the entire 3D data tributes, such as texture, anti-aliasing,
in a single 2D image by projecting a 2D normal vector, and so on. The voxelized
image directly from the 3D volumetric model can be either binary (see Kauf-
data. Volume rendering conveys more man [1991, pp. 280–301 and 302–310])
information than surface rendering, but or volume-sampled, which is an alias-
at the cost of increased algorithm com- free filtered voxelization of the model.
plexity and consequently increased ren- In many applications involving sampled
dering times. To improve interactivity or computed data, such as medical im-
in volume rendering, many optimization aging, the data need to be visualized
methods and several special-purpose along with synthetic objects, such as
volume-rendering machines [Kaufman scalpels, prosthetic devices, mirrors,
1991, pp. 311–378] have been devel- and radiation beams. These geometric
oped. objects can be voxelized and intermixed

ACM Computing Surveys, Vol. 28, No. 1, March 1996

 Volume Visualization • 167

with the sampled object in the volume memory systems, coupled with the de-
buffer. sire to reveal the inner structures of
Volume graphics is concerned with volumetric objects, suggests that vol-
the synthesis, manipulation, and ren- ume visualization and volume graphics
dering of volumetric geometric objects may develop into major trends in com-
[Kaufman et al. 1993]. Unlike volume puter graphics. Just as raster graphics
visualization, which focuses primarily in the seventies superseded vector
on sampled and computed datasets, vol- graphics for visualizing surfaces, vol-
ume graphics is concerned primarily ume visualization and volume graphics
with modeled geometric scenes and par- have the potential to supersede surface
ticularly with those that are repre- graphics for handling and visualizing
sented in a volume buffer. As an ap- volumes as well as surfaces.
proach, volume graphics has the
potential to greatly advance the field of
3D graphics by offering a comprehen- REFERENCES
sive alternative to traditional surface
graphics. Volume graphics has advan- COHEN, D. AND SHAKED, A. 1993. Photo-realistic
imaging of digital terrain. Comput. Graph.
tages over surface graphics in being
Forum 12, 3 (Sept.) 363–374.
viewpoint-independent and insensitive
GALLAGHER, R. S., (ED.) 1995. Computer Visual-
to scene and object complexities, and it ization. CRC Press, Boca Raton, FL.
lends itself to block operations, con- KAUFMAN, A. 1991. Volume Visualization. Tuto-
structive solid modeling, and hierarchi- rial, IEEE Computer Society Press, Los
cal representations. It is suitable for the Alamitos, CA.
representation of sampled or computed KAUFMAN, A., COHEN, D., AND YAGEL, R.
datasets and their intermixing with 1993. Volume graphics. IEEE Comput. 26, 7
geometric objects, and it supports the (July), 51– 64. In Japanese, Nikkei Comput.
Graph. 1, 88, 148 –155 and 2, 89 (1994), 130 –
visualization of internal structures and 137.
amorphous phenomena. The problems MALZBENDER, T. 1993. Fourier volume rendering.
associated with the volumetric repre- ACM Trans. Graph. 12, 3 (July), 233–250.
sentation, such as memory size, process- MURAKI, S. 1993. Volume data and wavelet
ing time, aliasing, and lack of geometric transform. IEEE Comput. Graph. Appl. 13, 4
information, echo problems encountered (July) 50 –56.
when raster graphics emerged as an ROSENBLUM, L. ET AL., EDS. 1994. Scientific Visu-
alternative technology to vector graph- alization: Advances and Challenges. Academic
Press, London.
ics and can be alleviated in similar
YEO, B. AND LIU, B. 1995. Volume rendering of
ways. DCT-based compressed 3D scalar data. IEEE
The progress so far in volume visual- Trans. Visual. Comput. Graph. 1, 1 (Mar.)
ization techniques, in hardware, and in 29 – 43.

ACM Computing Surveys, Vol. 28, No. 1, March 1996