Joint EUROGRAPHICS - IEEE TCVG Symposium on Visualization (2002)
D. Ebert, P. Brunet, I. Navazo (Editors)
Interactive Previewing for Transfer Function Speci cation in
Volume Rendering
Charl P. Botha and Frits H. Post
Data Visualisation Group
Delft University of Technology, The Netherlands
{c.p.botha,f.h.post}@its.tudelft.nl
http://visualisation.tudelft.nl/
Abstract
This paper presents a new technique for supplying meaningful visual feedback during direct volume rendering
transfer function speci cation. The technique uses meta-data calculated during a pre-processing step to generate
interactively an approximate volume rendering that is voxel-registered with a single user-selected slice. Because
of the registration, this preview can easily be alpha-blended with a grey-scale image of the data that is being
volume rendered. In this way, the user gets real-time visual feedback on her transfer function speci cation with
regards to both the expected composited optical properties and the delity (how closely the rendering matches
the original data) of the resulting rendering.
1. Introduction
Direct Volume Rendering12 (DVR) is an important and
useful technique for visualising structures in volumetric
data. However, the dif culty in specifying the DVR trans-
fer function3 can be prohibitive to the deployment of this
technique in practical visualisation applications. The transfer
function is a critical component of the volume rendering pro-
cess that speci es the relation between scalar data (e.g. com-
puterised tomography Houns eld units), as well as deriva-
tive values (e.g. the gradient volume of an MRI dataset), and
optical characteristics (e.g. colour and opacity).
Our work attempts to give the user full control over the
transfer function speci cation and to supply real-time visual
feedback in such a fashion that the system’s reliance on the
user’s visualisation expertise is mimimised and the leverage
of the user’s application-speci c (e.g. clinical) knowledge is
maximised. The visual feedback is supplied with a special
volume rendering preview technique.
The preview technique has been designed to give a good
indication of the composited optical properties in the resul-
tant volume rendering. In addition, our technique yields ex-
plicitly data-registered feedback, meaning that the delity
of the resultant rendering (and the transfer function) can be
checked. On a higher level, the one-to-one correspondence
between data and volume rendered structures is accentuated.
In section 2 we discuss existing methods for nding trans-
fer functions. Section 3 documents our new scheme and in
section 4 we show some examples that demonstrate its ef-
fectivity. We summarise our work and point out possible av-
enues for future research in section 5.
2. Related Work
Finding good transfer functions has been listed among the
top ten problems in volume visualisation3. Consequently,
much effort has recently been spent on improving this sit-
uation. Existing schemes range from fully manual to fully
automatic techniques for nding transfer functions.
Probably the oldest method of nding a transfer function
is by trial and error. This usually involves manipulating a
transfer function whilst periodically checking the resulting
volume rendering. If special volume rendering hardware that
can do this at interactive rates is not available, this can be a
very laborious and time-consuming process.
Another very interesting technique that is based on the de-
sign galleries paradigm4 generates many volume renderings
simultaneously, each representing a different con guration
c© The Eurographics Association 2002.

Botha and Post / Interactive Previewing for Transfer Function Speci cation in Volume Rendering
of the transfer function. The user selects the renderings that
satisfy her requirements and thus implicitly optimises the
transfer function. The considerable challenges here are to
generate automatically the different transfer functions that
are going to generate a wide spread of dissimilar output
renderings5 and to present these renderings to the user in an
effective way.
As potentially hundreds of different renderings have to be
made, this technique does rely on fast rendering hardware
being available to reach its full potential as an interactive
method.
Also take into account that many of the optimised
software volume rendering techniques such as shear-
warp factorization6 require pre-calculated transfer function
lookups, which makes them less applicable to this problem
where the transfer functions are being continuously modi-
ed.
The work of K nig and Gr ller combines elements of the
design galleries and trial-and-error techniques with the re-
quired use of real-time raycasting hardware7. They present
transfer function speci cation as a simpli ed three step pro-
cess: First the user indicates scalar ranges of interest, then
she assigns colours to these ranges and nally she assigns
opacities. Numerous feedback renderings are performed dur-
ing this process in order to guide the user’s choices. This
technique simpli es the speci cation to quite an extent.
In a previous paper8, we proposed a fast feedback-based
method that can be implemented without real-time raycast-
ing facilities. It transforms the current slice with the cur-
rent transfer function (as it is being edited) and then alpha-
blends a mapping of this transformed slice with a greyscale
slice of the data that is about to be rendered. Because it is
essentially visualising instantaneous optical properties (i.e.
not accumulated as in real volume rendering) it is less effec-
tive at previewing lower opacities. A user-tunable exponen-
tial opacity compensation is utilised to remedy this problem.
Although this works satisfactorily, we required a mathemat-
ically defensible foundation. The work presented in this pa-
per is based on the same overlay idea, but adds a predictive
raycasting accumulation and optional shading.
Bajaj’s contour spectrum9 consists of metrics that are
computed over a scalar eld. This spectrum, presented as
a user interface component, can be used to assist the user in
nding a suitable transfer function. It offers an alternative
and condensed way of examining the volume of data that
reveals global characteristics which can be very helpful in
creating a transfer function.
The semi-automatic method of Kindlmann10 is a highly-
regarded technique for generating transfer functions from
volumetric data. This method makes the reasonable assump-
tion that the features of interest in the data are the boundary
regions between areas of relatively homogeneous material.
Bajaj’s and Kindlmann’s methods are designed for creat-
ing transfer functions with minimal or no user-interaction.
According to our references, they do not address the prob-
lem of providing meaningful feedback during manual trans-
fer function speci cation or ne-tuning. Because our method
focuses on providing meaningful and fast feedback, it does
not replace these schemes, but could be used very pro tably
in augmenting them.
The design galleries approach, trial-and-error scheme and
their derivatives can be considered as being focused on pro-
viding feed-back in a user-driven transfer function speci ca-
tion process and this is where our scheme excels.
Even if interactive volume rendering facilities are avail-
able that can cope with continuous changes in the transfer
function, it is often dif cult to identify the fuzzy structures
in a feedback volume rendering that can result from a partic-
ular transfer function and accurately relate them to the struc-
tures in the data that are being rendered. This makes it very
dif cult to quantify the impact that a small modi cation to
the transfer function has.
It is desirable to have some kind of directed search process
where the user is gradually but purposefully moving towards
an optimal transfer function. The dif culty in quantifying
the impact of a transfer function change, especially with re-
gards to structures visible in the unprocessed data, hinders
this process and necessitates a high level of visualisation ex-
perience from the user. Our method yields explicit and real-
time feedback on the relationship between the structures in
the data that are being rendered and the structures that can
be expected in the resultant volume rendering.
3. The Data-Registered Interactive Previewing
Technique
The central idea of this technique is to predict the optical
result of casting a ray through a voxel and parallel to the
view direction, through all the material represented by that
particular voxel. The term material is de ned as the set
of optical properties assigned to a given scalar value by the
DVR transfer function.
The amount of a speci c material intersected by a ray
can be estimated for each voxel position by making use of
fast run-time processing of pre-calculated per-ray scalar fre-
quency distributions (i.e. histograms of scalar values). The
estimated amount, measured as a number of voxels, can be
used in a simpli ed form of the volume-rendering integral to
predict the optical outcome.
This prediction is done for all voxels in a particular slice
of the raw data and all predicted results are then alpha-
blended with that slice.
What the user sees is an estimate of what the volume ren-
dering would look like with her currently con gured transfer
functions, super-imposed on a grey-scale slice image of the
raw data. When the user makes any changes to the transfer
c© The Eurographics Association 2002.