Eurographics Workshop on Visual Computing for Biology and Medicine (2010), pp. 1–8
Dirk Bartz, Charl Botha, Joachim Hornegger and Raghu Machiraju (Editors)
Dynamic Visualisation of Orbital Fat Deformation using
Anatomy-Guided Interaction
P.J. Schaafsma, S. Schutte, H.J. Simonsz, F.H. Post, C.P. Botha
Abstract
The human eye is a biomechanical system. Orbital fat plays an important rol in the working of this system, but
its behaviour during eye movement is not well understood. To give insight into this behaviour, visualisation is a
useful tool. This paper presents a complete pipeline for interactive particle-based visualisation and exploration of
orbital fat deformation from MRI data. Sensible 3D particle seeding is important in this type of visualisation. We
address that problem with a two-step process: Interactive, anatomy-guided slice positioning, and contour-based
region of interest specification. Since the deformation calculation is unlikely to be correct everywhere, we derive
and visualise an uncertainty measure based on deformed and original MRI data. We also performed a case study
evaluation to investigate the benefits of our approach towards orbital fat deformation visualisation.
Categories and Subject Descriptors (according to ACM CCS): I.3.6 [Computer Graphics]: Interaction techniques
1. Introduction
The human eye is a biomechanical system in which the eye-
ball, optic nerve and eye muscles work together to direct the
eye and capture images to be sent to the visual cortex. These
anatomical structures reside inside the orbita, or eye socket.
The orbital fat is the semiliquid tissue that fills the space be-
tween these structures and passively deforms according to
the shape and position of other tissue. This happens, for in-
stance, as the optic nerve is dragged trough it due to eyeball
rotation, or when it is being pushed away by the left eye
muscle rolling up onto the eyeball when gazing to the right.
The eye fat plays many roles, such as keeping the muscles
from pulling the eyeball back into the eye socket, lubricating
the muscles’ movement against the orbita wall, and dampen-
ing the rapid and sudden rotation of the eyeball, which can
quickly accelerate to upto 1000◦ per second.
The behaviour of orbital fat is not well understood. Im-
proving this is important because a better understanding
leads to more informed decisions during eye surgery. To this
end, visualisation is a valuable tool. Its goal is to give insight
into complex data. Deformation data is often presented as
low-level 3D time-dependent vector data, while human un-
derstanding responds better to a higher-level description as
patterns and features, such as pushing-away, dragging-along
and gliding-through. Visualisation helps to close the gap be-
tween these different levels of representation.
Figure 1: Particle visualisation of fat deformation in the
right orbita, posterior sideways view. A regular grid of par-
ticles was seeded behind the optic nerve, which in this view
is moving downwards. The image shows a deformed grid, in-
dicating fat movement. Particles are coloured according to
a deformation uncertainty measure.
In this paper, we present a complete pipeline for the dy-
namic visualisation and interactive exploration of orbital fat
deformation, see figure 1. After MRI data acquisition of an
axial skull section containing both orbitae, we use the Lu-
cas and Kanade optic flow algorithm to calculate deforma-
submitted to Eurographics Workshop on Visual Computing for Biology and Medicine
(2010)

2 P.J. Schaafsma, S. Schutte, H.J. Simonsz, F.H. Post, C.P. Botha / Orbital Fat Deformation
tion data as 3D time-dependent vector fields. An isosurface
rendering of relevant anatomical features provides a frame
of reference in which further visualisation and exploration
take place. Anatomy-guided region of interest specification
provides a seeding volume for particle-based visualisation
of deformation data. An uncertainty measure, derived from
comparison of deformed and original MRI data, is used to
qualify the particle visualisation. These techniques are com-
plemented by streamlines and pathlines, also seeded from
a specified region of interest. Furthermore, we evaluated
our approach trough a case study conducted with an eye
biomechanics scientist. A video demonstration of our work
is available online†.
With this paper, our main contributions are:
• Intuitive, anatomically relevant slice positioning and re-
gion of interest specification
• Focus+Context uncertainty visualisation for deformation
data
• Improved visualisation of orbital fat deformation though
tightly coupled dynamic visualisation and interactive ex-
ploration
The rest of this paper follows standard structure guide-
lines. After related work (section 2), we focus on the body of
our research: the visualisation and interaction pipeline (sec-
tion 3). Then, the results of a domain expert evaluation are
presented (section 4), and we conclude with analysis of our
results and mention promising avenues for future research
(section 5).
2. Related Work
Visualisation of orbital fat deformation was researched by
Botha et al. [BdGS∗07]. They use the Demons registration
algorithm to calculate deformation fields from a sequence
of MR images. While we only focus on general, interactive,
dynamic visualisation of this type of data, their work also ad-
dressed specific questions on the relative motion of fat tissue
to other orbital structures. The 3D Lucas and Kanade algo-
rithm we use to obtain orbital deformation data from MR
images was presented in detail by Abràmoff et al. [AV02].
While we follow their work in the calculation of deformation
data, we use a different approach for the uncertainty calcu-
lation that is closer to the data. Instead of deriving deforma-
tion data from MR images data, Schutte et al. [SvdBvK∗06]
worked on development of a finite element model that de-
scribes the eye’s behaviour. Schoemaker et al. [SHM∗06] de-
termined material properties of orbital tissue for use in such
an FEM model.
Fuchs et al. [FWH10] use an outline of the region of in-
terest (ROI) on the view plane to create an initial ROI. It-
erative operation on vertices and edges allows creation of
† http://graphics.tudelft.nl/publications/Schaafsma2010-
OrbitalFatDeformation.avi
complex polyhedra. This makes their technique general, al-
beit slower than ours. They also discuss and compare region
of interest specification techniques in general. Schiffman et
al. [SNB∗03] draw closed contours a brain surface to directly
specify a surface ROI.
Research on data-guided plane placement beyond stan-
dard orthogonal slices and freely rotating surfaces is limited.
Kreylos et al. [KBK08] fit primitives, such as cylinders and
cubes, to a point set. The point set is a result of interactive
brushing in point cloud rendering. Being able to fit many
types of shapes is useful, although it does require the extra
step of selecting the type of shape to fit.
Visualising general uncertainty in time-dependent vec-
tor fields was researched for texture-based visualisation by
Botchen et al. [BWE05]. They distinguish uncertainty levels
by parameterising visualisation style, but do not necessarily
guide user attention towards low uncertainty regions.
3. Deformation Visualisation Pipeline
In this section we present in relevant detail the complete
pipeline from MRI data acquisition to the visualisation re-
sults. The pipeline can be divided into three main sections:
data acquisition and processing, interaction, and visualisa-
tion, see figure 2. All three sections work together to create
an interactive environment for unrestricted exploration of or-
bital fat deformation.
3.1. Data acquisition and Processing
Our dataset consists of 15 axial MRI volumes contain-
ing both orbitae of a volunteer subject. Each 3D scalar
volume corresponds to a different gaze direction, rang-
ing from −30◦ (left) to +40◦ (right) axial rotation at
constant 5◦ increments. The dimensions of each volume
are 256×256×116 voxels (sagittal×coronal×axial), spaced
at 0.5469mm×0.5469mm×0.5000mm. During acquisition,
the subject was asked to bite down on a rigid landmark ob-
ject, which enabled accurate dataset realignment afterwards.
The raw scalar data is used in three subsequent process-
ing steps: segmentation, deformation calculation, and uncer-
tainty calculation. We will discuss the latter two in more de-
tail below.
We assume the rotation in each volume to be exactly the
specified, incremental angle, instead of having measured it
ourselves after acquisition, for instance by measuring the
orientation of the optic nerve and lens, both of which can be
accurately identified in the MRI data. We also assume that
the deformation of the orbital structures during high speed
rotation equals the static deformation of orbital structures at
fixed eyeball rotation angle, i.e. dynamic deformation equals
static deformation. Under this assumption, the actual defor-
mation depends only on view direction, not on angular ve-
locity of the eyeball. Also, while spatial resolution is dictated
submitted to Eurographics Workshop on Visual Computing for Biology and Medicine (2010)

P.J. Schaafsma, S. Schutte, H.J. Simonsz, F.H. Post, C.P. Botha / Orbital Fat Deformation 3
Figure 2: Three main components of our visualisation pipeline for orbital fat deformation. After MRI data acquisition (not
shown), data processing, interaction and visualisation work together to make interactive exploration of the dataset possible.
Note that the main purpose of the interaction pipeline component is to provide a point set for further visualisation.
by the acquisition process, we define the time frame between
consecutive images simply as unit time.
3.1.1. Deformation Calculation
Following Abràmoff [AV02], we used a 3D implementation
of the Lucas and Kanade (LK3D) [LK81] optic flow algo-
rithm to calculate dense deformation data for each of the
15 scalar volumes. LK3D produces per voxel deformation
data solely based on a least squares estimate from spatial and
temporal image derivatives over a small, Gaussian-weighted
region of influence. Aside from allowing generation of phys-
ically improbable data, it can also produce erratic deforma-
tion data. At the cost of sensitivity to small scale features,
this can be improved by either smoothing the original scalar
data or image derivatives, or by increasing the algorithm’s
region of influence. The image derivatives are calculated by
convolution of the MR images with a 4D Gaussian derivative
kernel.
The spatial and temporal image derivative were created
from MR image data by convolution with a spatially sym-
metric 4D Gaussian derivative kernel. The kernel width was
23 voxels with a variance of 1.0mm2 in each spatial di-
rection, and 3 voxels and 1.0 respectively in time. LK3D
weighted least squares minimisation also uses a Gaussian
kernel to weight the contributions of individual pixels. We
used a kernel width of 13 voxels and variance of 4.0mm2
in spatial direction. We did not minimise over multiple time
steps.
3.1.2. Uncertainty Calculation
The LK3D algorithm does not produce accurate results ev-
erywhere, to which several factors contribute. First, noise in
MRI data cascades though the derivative images to the de-
formation data. The sensitivity of LK3D to input noise is
highest in uniform intensity regions. Second, the quantised
character of the MRI data, both in time and space, allows
only estimates of the true image derivatives. Convolution of
the MRI data with a 4D Gaussian derivative kernel addresses
both these issues, but also smooths out small-scale image
patterns. Its variance determines the balance between sen-
sitivity to detail and noise robustness. Third, the Gaussian
variance of the weighted LK3D region of influence intro-
duces a similar source of error.
To calculate uncertainty for a single deformation image
we first place a regular, super-sampled grid of virtual parti-
cles in the domain. These particles are then advected over
unit time using the deformation data. Each particle carries
the linearly interpolated intensity value from its start loca-
tion. The intensity value of the image next in sequence at
the advected particle’s location is sampled and subtracted
from the carried intensity value. A large difference leads to
high uncertainty. The uncertainty value is then distributed
over eight voxels similar to the intensity value interpolation
at initial particle location. Since the last deformation field
has neither a next scalar nor deformation field, we duplicate
the second to last uncertainty volume. This way we have an
uncertainty field for each scalar and vector field.
3.2. Interaction
Trying to visualise too much deformation data at once will
lead to many of the problems common to the 3D visualisa-
tion challenge, mainly occlusion and cluttering. Specifying
a region of interest addresses these problems. We do this as
a two-step process. First, the user specifies a slice through
the dataset by drawing a set of points on a reference surface.
After fine-tuning slice position and orientation, a region of
interest is defined by drawing and extruding a contour on the
specified slice. Both steps are discussed in detail below.
submitted to Eurographics Workshop on Visual Computing for Biology and Medicine (2010)

4 P.J. Schaafsma, S. Schutte, H.J. Simonsz, F.H. Post, C.P. Botha / Orbital Fat Deformation
Figure 3: Slice positioning and adjustment in the right orbita, inferior axial view. Prominent features in the grey isosurface are
the eyeball (large spherical shape to the right) and optic nerve (cylindrical structure attached on the left side to the eyeball).
Left: A slice is determined by least squares fitting a plane to a point set (red) drawn on the eyeball. Centre: First and last drawn
point form the basis for a slice transformation widget. Inset shows the slice dragged through the volume along the plane normal.
Right: Repositioning of the widget by drawing a line on the slice through the optic nerve. Inset shows the slice after rotation
according to the new widget.
3.2.1. Anatomy-guided Slice Positioning
To provide an anatomical frame of reference for slice po-
sitioning, we observe that in MRI data, different anatomi-
cal structures are represented by different scalar ranges. Al-
though they overlap and MRI data is noisy, the main fea-
tures of the eye are clearly distinguishable. Isosurface ex-
traction and rendering is a simple yet sufficient technique to
start with when placing an anatomically relevant slice inside
a volume.
Drawing directly on the isosurface leads to a point set
from which we determine the position and orientation of a
plane using least squares estimation of the four plane equa-
tion parameters. This technique effectively gives the user the
opportunity to draw on a visible surface an approximation of
part of the prospective slice intersection with the surface, al-
though results are well-defined for almost any drawing pat-
tern. Another way of using this technique in an interactive
viewing environment, is to first reposition the camera such
that it would coincide with the plane. A single, straight draw-
ing gesture then quickly defines a plane of desired orienta-
tion and position. In our experience, this last method is eas-
ier and more reliable, especially in highly curved areas of the
reference surface.
Final adjustment or predictable repositioning of the slice
is required if first results are either unsatisfactory or planned
slice position and orientation are otherwise closely related
to a prior placement attempt. We have designed a three-
component, L-shaped widget that allows normal translation
and two degrees of rotational freedom of the slice, see fig-
ure 3. The widget is operated by adjusting the position of
the three spherical components. The orange corner compo-
nent allows translation in the plane normal direction. The
yellow in-plane component rotates the plane around the axis
normal to the widget and through the translation component.
The red out-of-plane component rotates the plane around the
in-plane axis of the widget. Initially, the widget’s location
and orientation are determined by the point set from which
the slice itself was derived. The orange and yellow compo-
nent correspond to the start and end drawing point projected
onto the plane. Simply redrawing on the plane repositions
the widget according to the new point set.
This interaction style for plane placement strikes a bal-
ance between complete freedom of orientation and position-
ing, and predictable, data-guided results. Note that the initial
placement using the second method already often provides
good results. Although we demonstrate this technique by us-
ing the plane as the basis for region of interest specification,
other techniques that rely on slices through in volumetric
data, such as Multi-Planar Reconstruction, also benefit from
a more intuitive, data-guided approach.
3.2.2. Region of Interest Specification
Once the interaction slice is in place, specifying a region
of interest by extrusion of a contour drawn on the slice
is straightforward. Contour drawing is easily distinguished
from widget repositioning by start and endpoint proxim-
ity testing. Having completed a contour, extrusion in the
slice normal direction using a widget similar to the trans-
lation component of the slice widget defines a closed re-
gion of interest volume, see figure 4. This region is used
during visualisation as a point seeding volume, to be used
in the particle-based visualisation as well as for streamline
and pathline generation. Contour drawing on the interac-
tion slice allows specification of custom regions of interest,
bound only by the requirement that they be an extrusion of a
non-selfintersecting contour of coplanar points.
Particles are initially seeded randomly over the data do-
main. They are also tagged as invalid and hence not dis-
played. During tracing, described in section 3.3.3, this is de-
tected and individual particles are reseeded inside the region
submitted to Eurographics Workshop on Visual Computing for Biology and Medicine (2010)

P.J. Schaafsma, S. Schutte, H.J. Simonsz, F.H. Post, C.P. Botha / Orbital Fat Deformation 5
Figure 4: Region of interest specification. Left: Direct placement of a slice through the optic nerve in the view direction.
Inset is the result of a camera change towards the bottom left. Centre: A zoomed-in, rotated view shows the drawing point set
distribution over the isosurface. Inset shows a contour being drawn. Right: A contour is extruded to form a region of interest
just below the optic nerve. Inset shows particles seeded inside the region of interest, just after animation has started.
of interest. For efficiency and speed, this process is split into
two phases. First, upon contour completion, initialisation of
a texture that separates the contour inside from the contour
outside. Second, when a particle is reseeded, remapping of
the particle position and a texture lookup to see whether the
new location is inside the region of interest. This separation
ensures, after initialisation, very fast region of interest in-
side/ouside testing for prospective particle reseed locations.
In detail, we take the following steps.
Intialisation of reseeding texture upon contour comple-
tion:
• Obtain an approximately minimal in-plane bounding rect-
angle around the contour using Principal Component
Analysis.
• Map the contour points onto the unit square. To do this,
express the contour points as a linear combination of the
scaled principal components, such that the mapped coor-
dinates are within [0.0, 1.0].
• Use an Advancing Front algorithm [PVMZ87] to create
an inside mesh for the mapped contour.
• Render the mesh as a unit quad to a texture of arbitrary
resolution. This texture will be available until another
contour is drawn.
Particle reseeding during particle tracing:
• Hash the particle coordinates into the unit cube, providing
some randomness.
• Do a texture lookup using hashed particle coordinates pro-
jected onto the unit square.
• If lookup indicates ’outside’, rehash the particle location
in a way that prevents repeated failure.
• If lookup indicates ’inside’, remap the hashed particle lo-
cation back to world coordinates, and validate the particle.
3.3. Visualisation
Immediately after specifying a seed volume, the user can
start the dynamic part of the visualisation, after which
the chosen deformation field representation and anatomical
structures are continuously animated over the available time
points. At any time, settings such as the resolution of seed-
ing can be adapted interactively, with immediate visual feed-
back. In the following subsections, the available visualisa-
tion techniques are discussed in more detail.
3.3.1. Anatomical Context Visualisation
The goal of anatomy visualisation as context to the defor-
mation data is little more than providing a frame of refer-
ence. This helps the viewer to form a mental image of the
eye fat deformation during exploration. It should not attract
attention, but be there to give meaning to the deformation
visualisation. We use both the original MR images and the
isosurfacing results to provide different levels of context.
The interaction plane is presented as an interpolated,
semi-transparent slice of the original MRI data. It provides
only minimal context, but avoids any occlusion problems
caused by the isosurface. Visualised together with the iso-
surface, the textured, semi-transparent plane does partly oc-
clude the reference surface, while against a plain back-
ground, scalar structures are clearly visible.
3.3.2. Uncertainty Visualisation in Deformation
We visualise uncertainty by adjusting the rendering style
of individual particles, an example of which is shown in
figure 5. The uncertainty is based on two measures: A
normalised, linearised value derived from deformation and
scalar data, see section 3.1.2, and a binary mask value
derived from segmentation indicating whether a particle
finds itself in the orbital fat. Borrowing ideas from Fo-
cus+Context, we use visibility, colour, and spatial frequency
to naturally avert user attention away from high uncertainty
regions.
We have designed an uncertainty transfer function that
combines both measures into a single function. First, a par-
ticle’s location is checked against the orbital fat mask, de-
submitted to Eurographics Workshop on Visual Computing for Biology and Medicine (2010)

6 P.J. Schaafsma, S. Schutte, H.J. Simonsz, F.H. Post, C.P. Botha / Orbital Fat Deformation
Figure 5: Uncertainty visualisation style. Top-Left: no un-
certainty visualisation, all particles rendered in the same
style. Top-right: Color as indication of uncertainty. Higher
saturation equals lower uncertainty. High spatial frequency
makes all particles attract almost equal attention during an-
imation. Botton-Left: Fading of sharp edges for high uncer-
tainty data naturally draws attention to other regions. This
is especially noticeable during animation. Out-of-mask par-
ticles on the right are not shown. Bottom-Right: Silhouette
visualisation of out-of-mask particles. Note that due to their
number, they are not clearly distinguishable. The rings also
attract too much attention due to their crisp edges.
ciding between out-of-mask and in-mask particle render-
ing. Second, the uncertainty value for in-mask particles is
mapped onto a perceptually linear colour scale. The render-
ing style for out-of-mask particles had to convey just enough
information to show that the particlesd are there, but not at-
tract user attention away from in-mask particles, however
high their uncertainty may be.
3.3.3. Particle-based Deformation Visualisation
Interactive particle-based visualisation can show global de-
formation patterns as well as small-scale local features, see
figure 6. Its insensitivity to noise, due to a few in thousands
of particles not detracting from global flow patterns, makes
it a good visualisation choice for our type of data. As an ex-
ploratory technique for time-dependent deformation data it
is useful as well, as it assumes nothing about the data and
captures small-scale temporal aspects of deformation. Com-
bined with the particle seeding strategy presented in this pa-
per, this technique delivers good results. Identifying larger-
scale temporal aspects or spatial patterns in the deformation
data is more difficult, as a particle has no memory.
Interactive adjustment of particle size is useful to match
the scale of the visualisation to the region of interest. Note
that the size of a particle is a visual attribute only; the parti-
cle is still advected over time as a dimensionless point. Us-
ing particle density to determine the number of particles to
be seeded inside the region of interest is an improvement
over using a fixed number of particles. The advantage is that
during exploration, specifying different regions-of-interest
of varying size, the particle density is not affected, leading
to a more uniform visualisation style between exploration
iterations.
3.3.4. Pathlines and Projected Streamlines
Figure 7: Streamlines and pathlines. Left: Projected Stream-
lines behind the opticus (dark grey patch on the bottom) as it
moves downwards. The projection of the streamlines onto the
interaction plane might give the idea that fat tissue is being
compressed. This is not the case, as the projection ignores
any motion perpendicular to the interaction plane. Right:
Pathlines seeded just below the bottom edge of the opticus
show that the fat initially travels downwards with the opti-
cus, but eventually gets pushed to the side. Color value indi-
cates time.
We also visualise the deformation data using pathlines and
streamlines, see figure 7. These techniques show the defor-
mation field in a different way than particle animation. Al-
though from simple pathlines, for instance, it is hard to get a
feel of how a deformation field behaves from one instant to
the next, when densely seeded they do indicate the volume
that the deforming matter in the seeding region has occu-
pied over time. Similarly, streamlines, being instantaneous
by nature, are not always suitable to give insight into time-
dependent data, but rather serve to inspect the characteristics
of the deformation field at a given instance.
4. Evaluation
Based on the principles and terminology set out by Yin for
case study research [Yin09], we performed an evaluation,
investigating the use of the visualisation tool by the second
author of this paper, a published expert in eye biomechan-
ics. Henceforth, we refer to this expert simply as “the user”.
The main study question was formulated as “How can the
visualisation tool assist biomechanics researchers in study-
ing the deformation of orbital fat under different directions
submitted to Eurographics Workshop on Visual Computing for Biology and Medicine (2010)

P.J. Schaafsma, S. Schutte, H.J. Simonsz, F.H. Post, C.P. Botha / Orbital Fat Deformation 7
Figure 6: Deformation of orbital fat. Left: Particles are laid out in a grid pattern just below the opticus. The eyeball is in the
back. Centre: As the opticus moves downwards, orbital fat (represented by particles) is starting to be pushed downward and to
the side, out of the way of the opticus. Right: As the opticus moves close to one of the eye muscles, fat is getting pushed more
and more to the side.
of gaze?” and the case was defined as “Use of the tool by an
expert in eye biomechanics.” We have chosen for this eval-
uation method as we wanted to study the application of the
complete tool in vivo, and as the potential user group is quite
small.
The evaluation was performed partly with the user operat-
ing the tool by himself, and partly by the first author demon-
strating. During the evaluation, we investigated a number of
case study propositions that we had previously defined in-
dependently of the user. The rest of this section consists of
the user’s feedback, structured along the study propositions
(printed in italics), followed by our conclusions.
The anatomy-guided plane interaction approach enables
the rapid definition of meaningful 2D planar reconstructions
of the 3D data. The user, having extensive experience work-
ing with CAD software, was in general quite impressed with
the anatomy-guided plane placement approach. Besides the
general utility of the method, the ability to rotate the plane
easily around its own axis was also appreciated. The follow-
ing points of criticism were discussed: 1) Due to the free
plane placement, as opposed to the more traditional orthog-
onal planes, maintaining a sense of orientation was difficult.
Adding orientation feedback, for example a glyph represent-
ing the gaze direction and the foot-head direction would al-
leviate these problems. 2) The plane manipulation spheres
can be confusing. These should be replaced with glyphs that
clearly show their purpose. 3) The effectivity of the plane
placement approach is of course dependent on the quality of
the anatomical surfaces.
The anatomy-guided plane interaction and volume of in-
terest specification enables the rapid definition of volumes
containing potentially interesting, according to expert ex-
pectation, fat deformations. After a few minutes of instruc-
tion, the user was able to independently explore the data
and rapidly specify anatomy-based regions of interest. It was
remarked that the method assumes good knowledge of the
anatomy under study.
The anatomy-guided interaction is significantly better
(easier / more intuitive / more effective) than other systems
with which the domain scientist has experience. The user has
experience with traditional CAD-software interfaces, but re-
marked that our interaction approach is indeed better in the
context of anatomy-driven selection.
The dense real-time particle tracing facilitates insight
into complex deformation patterns better than region-query
based techniques such as advection balls. (animation / dy-
namic; density) The user agreed strongly with this propo-
sition and added that the dense particle tracing was more
intuitive.
Confidence is clearly and intuitively represented. Initially
we used outline rendering for particles outside of the orbital
fat, and a perceptually linearised yellow-to-blue colour scale
to represent confidence or uncertainty in the rest of the de-
formation field. We asked the user to guess what the mean-
ing was of the different colouring, without giving him any
other information, upon which he speculated that the out-
line rendering meant that there was no data, or that it was
an extrapolation of other valid data. When we explained the
meaning, they commented that the blue-to-yellow scale was
not immediately clear in terms of what was high and low,
and that they found the grey-to-blue part of the colour scale
more suitable in this regard.
Visual representation of confidence / uncertainty aids in
the correct interpretation of the fat deformation, as it indi-
cates where the deformation model results should be more
carefully considered. The user remarked that due to all parti-
cles moving, it was hard to intuitively associate colour with
confidence and integrate its implications during the visual
interpretation of the data. He did admit that if he was to dis-
cover an interesting an interesting phenomenon in a low con-
fidence area, he would double-check the data. The remark
was also made that the outline representation was clearly
distinguished from the rest of the particles.
The visualisation reflects the expectations of the eye
submitted to Eurographics Workshop on Visual Computing for Biology and Medicine (2010)

8 P.J. Schaafsma, S. Schutte, H.J. Simonsz, F.H. Post, C.P. Botha / Orbital Fat Deformation
biomechanics domain scientist. During the evaluation, the
user completely indepedently specified a volume of inter-
est between the lateral eye muscle and the eyeball and man-
aged to create a comprehensive visualisation of the orbital fat
evacuating rapidly to above and below as the muscle rolls up
on the eye during gaze change. This behaviour is the topic of
great interest in the eye biomechanics community. Besides
this example, the user agreed that for a number of expected
behaviours the visualisation reflects his expectations.
The visualisation could be used as a component in the
validation of numerical models of the eye (compare model
behaviour globally to measured behaviour). There was em-
phatic agreement with this proposition, with the user adding
that the visualisation gives insight into the differences be-
tween simulation and measurement.
The visualisation could improve current research on or-
bital fat deformation (insight / understanding / hypotheses
/ practical applications for research results / compare FEM
behaviour to this behaviour?) The user agreed strongly with
the statement that it would facilitate understanding and com-
mented that research up to now was primarily hypothesis-
driven and region-query based, as opposed to explorative.
From these propositions, we conclude that the tool can
assist eye biomechanics researchers by offering them a com-
pletely new way to visually study the behaviour of orbital fat
as it deforms during eye movement, both through the dense
real-time particle tracing and through the flexible selection
of anatomically relevant regions of interest. In addition, a
number of suggestions for improvement can now be acted
upon, most important of which the anatomical orientation
feedback. The intriguing role of uncertainty visualisation re-
quires further study, although these first evaluation steps are
promising.
5. Conclusion and Future Work
In this paper, we have presented a complete pipeline for the
interactive visualisation of orbital fat deformation under dif-
ferent directions of gaze. Our work improves on the state of
the art in two ways: First, interactive anatomy-guided plane
placement and volume of interest specification allows for the
flexible exploration of deformation of data in an anatomical
context. Second, dense particle-based visualisation, accord-
ing to a domain expert, significantly improves on visuali-
sation styles employed by earlier work in this area. In ad-
dition, we investigated the role our application could play
in the study of orbital fat deformation through a case study
research evaluation conducted together with an eye biome-
chanics expert.
We plan to deploy and refine our tool in the further study
of orbital fat deformation. During the evaluation, previously
elusive aspects such as the vertical fat evacuation during eye
movement could be convincingly visualised and have led to
two of our domain scientist colleagues requesting a more
in-depth investigation of the clinical implications. Finally,
we would like to apply these techniques to other types of
anatomical deformation.
References
[AV02] ABRÀMOFF M. D., VIERGEVER M. A.: Compu-
tation and Visualization of Three-Dimensional Soft Tis-
sue Motion in the Orbit. IEEE Transactions on Medical
Imaging 21 (2002), 296–304.
[BdGS∗07] BOTHA C. P., DE GRAAF T., SCHUTTE S.,
ROOT R., WIELOPOLSKI P., VAN DER HELM F. C. T.,
SIMONSZ H. J., POST F. H.: MRI-based visualisation of
orbital fat deformation during eye motion. In Visualiza-
tion in Medicine and Life Sciences (VMLS) (2007), Linsen
L., Hagen H., Hamann B., (Eds.), pp. 221–236.
[BWE05] BOTCHEN R., WEISKOPF D., ERTL T.:
Texture-based visualization of uncertainty in flow fields.
In Proceedings of IEEE Visualization 2005 (2005), Cite-
seer, p. 647â ˘AS¸654.
[FWH10] FUCHS R., WELKER V., HORNEGGER J.: Non-
convex polyhedral volume of interest selection. Comput-
erized medical imaging and graphics : the official jour-
nal of the Computerized Medical Imaging Society 34, 2
(2010), 105–13.
[KBK08] KREYLOS O., BAWDEN G., KELLOGG L.: Ad-
vances in Visual Computing, vol. 5358 of Lecture Notes
in Computer Science. Springer Berlin Heidelberg, Berlin,
Heidelberg, 2008.
[LK81] LUCAS B. D., KANADE T.: An Iterative Image
Registration Technique with an Application to Stereo Vi-
sion, 1981.
[PVMZ87] PERAIRE J., VAHDATI M., MORGAN K.,
ZIENKIEWICZ O. C.: Adaptive remeshing for com-
pressible flow computations. Journal of Computational
Physics 72, 2 (1987).
[SHM∗06] SCHOEMAKER I., HOEFNAGEL P. P. W.,
MASTENBROEK T. J., KOLFF C. F., SCHUTTE S.,
VAN DER HELM F. C. T., PICKEN S. J., GERRITSEN A.
F. C., WIELOPOLSKI P. A., SPEKREIJSE H., SIMONSZ
H. J.: Elasticity, viscosity, and deformation of orbital
fat. Investigative ophthalmology & visual science 47, 11
(November 2006), 4819–26.
[SNB∗03] SHIFFMAN S., NG Y.-R., BROSNAN T. J.,
ELIEZ S., LINKS J. M., KELKAR U. V., REISS A. L.:
Interactive specification of regions of interest on brain sur-
faces. NeuroImage 20, 3 (November 2003), 1811–6.
[SvdBvK∗06] SCHUTTE S., VAN DEN BEDEM S. P. W.,
VAN KEULEN F., VAN DER HELM F. C. T., SIMONSZ
H. J.: A finite-element analysis model of orbital biome-
chanics. Vision research 46, 11 (2006), 1724–31.
[Yin09] YIN R. K.: Case Study Research: Design and
Methods, fourth ed. Sage, 2009.
submitted to Eurographics Workshop on Visual Computing for Biology and Medicine (2010)