Deformable organisms for automatic medical image analysis.
ABSTRACT We introduce a new approach to medical image analysis that combines deformable model methodologies with concepts from the field of artificial life. In particular, we propose "deformable organisms", autonomous agents whose task is the automatic segmentation, labeling, and quantitative analysis of anatomical structures in medical images. Analogous to natural organisms capable of voluntary movement, our artificial organisms possess deformable bodies with distributed sensors, as well as (rudimentary) brains with motor, perception, behavior, and cognition centers. Deformable organisms are perceptually aware of the image analysis process. Their behaviors, which manifest themselves in voluntary movement and alteration of body shape, are based upon sensed image features, pre-stored anatomical knowledge, and a deliberate cognitive plan. We demonstrate several prototype deformable organisms based on a multiscale axisymmetric body morphology, including a "corpus callosum worm" that can overcome noise, incomplete edges, considerable anatomical variation, and interference from collateral structures to segment and label the corpus callosum in 2D mid-sagittal MR brain images.
[show abstract] [hide abstract]
ABSTRACT: We propose to use statistical models of shape and texture as deformable anatomical atlases. By training on sets of labelled examples these can represent both the mean structure and appearance of anatomy in medical images, and the allowable modes of deformation. Given enough training examples such a model should be able synthesise any image of normal anatomy. By finding the parameters which minimise the difference between the synthesised model image and the target image we can locate all the modelled structure. This potentially time consuming step can be solved rapidly using the Active Appearance Model (AAM). In this paper we describe the models and the AAM algorithm and demonstrate the approach on structures in MR brain cross-sections.02/2000;
IEEE Trans. Pattern Anal. Mach. Intell. 01/2000; 22:85-106.
Conference Proceeding: Controlled Shape Deformations via Medial ProfilesVision Interface (VI);
Deformable Organisms for Automatic Medical Image Analysis
Ghassan Hamarneh1,2, Tim McInerney3,2, and Demetri Terzopoulos4,2
1 Dept. of Signals and Systems, Chalmers University of Technology, Göteborg 41296, Sweden
2 Dept. of Computer Science, University of Toronto, Toronto M5S 3H5, Canada
3 School of Computer Science, Ryerson University, Toronto M5B 2K3, Canada
4 Courant Institute, New York University, New York NY 10003, USA
Abstract. We introduce a new paradigm for automatic medical image analysis that
adopts concepts from the field of Artificial Life. Our approach prescribes
deformable organisms, autonomous agents whose objective is the segmentation and
analysis of anatomical structures in medical images. A deformable organism is
structured as a ‘muscle’-actuated ‘body’ whose behavior is controlled by a ‘brain’
that is capable of making both reactive and deliberate decisions. This intelligent
deformable model possesses an ‘awareness’ of the segmentation process, which
emerges from a conflux of perceived sensory data, an internal mental state,
memorized knowledge, and a cognitive plan. We develop a class of deformable
organisms using a medial representation of body morphology that facilitates a
variety of controlled local deformations at multiple spatial scales. Specifically, we
demonstrate a deformable ‘worm’ organism that can overcome noise, incomplete
edges, considerable anatomical variation, and occlusion in order to segment and
label the corpus callosum in 2D mid-sagittal MR images of the brain.
The automatic segmentation and labeling of anatomical structures in medical images is a
persistent problem that continues to defy solution. There is consensus within the medical
image analysis research community that the development of general-purpose automatic
segmentation algorithms will require not only powerful bottom-up, data-driven
processes, but also equally powerful top-down, knowledge-driven processes within a
robust decision-making framework that operates across multiple levels of abstraction
. Deformable models, one of the most actively researched model-based segmentation
techniques , feature a potent bottom-up component founded in estimation theory,
optimization, and physics-based dynamical systems, but their top-down processes have
traditionally relied on interactive initialization and guidance by knowledgeable users.
Attempts to fully automate deformable model segmentation methods have so far been
less than successful at coping with the enormous variation in anatomical structures of
interest, the significant variability of image data, the need for intelligent initialization
The time has come to shift our attention to what promises to be a critical element in
any viable, highly automated solution: the decision-making framework itself. Existing
decision-making strategies for deformable models are inflexible and do not operate at an
appropriate level of abstraction. Hierarchically organized models, which shift their focus
from structures associated with stable image features to those associated with less stable
features, are a step in the right direction [4,9]. However, high-level contextual
knowledge remains largely ineffective because it is intertwined much too tightly with the
low-level optimization-based mechanisms. It is difficult to obtain intelligent, global (i.e.,
over the whole image) model behavior throughout the segmentation process from such
mechanisms. In essence, current deformable models have no explicit awareness of where
they (or their parts) are in the image or what their objectives are at any time during the
It is our contention that we must revisit ideas for incorporating knowledge that
were explored in earlier systems (e.g., ), and develop new algorithms that focus on
top-down reasoning strategies which may best leverage the powerful bottom-up feature
detection and integration abilities of deformable models and other modern model-based
medical image analysis techniques. We further contend that a layered architecture is
appropriate, where the high-level reasoning layer has knowledge about and control over
the low-level model (or models) at all times. The reasoning layer should apply an active,
explicit search strategy that first looks for the most stable image features before
proceeding to less stable image features, and so on. It should utilize contextual
knowledge to resolve regions where there is a deficiency of image feature information.
To achieve these goals, we introduce a new paradigm for automatic medical image
analysis that adopts concepts from the emerging field of Artificial Life. In particular, we
develop deformable organisms, autonomous agents whose objective is the segmentation
and analysis of anatomical structures in medical images. A deformable organism is
structured as a ‘muscle’-actuated ‘body’ whose behavior is controlled by a ‘brain’ that is
capable of making both reactive and deliberate decisions. This intelligent deformable
model possesses a non-trivial ‘awareness’ of the segmentation process, which emerges
from a conflux of perceived sensory data, an internal mental state, memorized
knowledge, and a cognitive plan. By constructing deformable organisms in a layered
fashion, we are able to separate the knowledge-driven model-fitting control functionality
from the data-driven, local image feature integration functionality, exploiting both for
1.1 Artificial Life Modeling
The Artificial Life (ALife) modeling approach has been applied successfully to produce
realistic computer graphics models of plants and animals . Artificial animals are
relevant to deformable organisms. Autonomous agents known as “artificial fishes” 
serve to illustrate the key functional components of artificial animals: bodies that
comprise muscle actuators, sensory organs (eyes, etc.) and, most importantly, brains
consisting of motor, perception, behavior, learning and cognition centers. Controllers in
the motor center coordinate muscle actions to carry out specific motor functions, such as
locomotion and sensor actuation. The perception center employs attention mechanisms
to interpret sensory information about the dynamic environment. The behavior center
realizes an adaptive sensorimotor system through a repertoire of behavior routines that
couple perception to action in meaningful ways. The learning center in the brain enables
the artificial animal to learn motor control and behavior through practice and sensory
reinforcement. The cognition center enables it to think.
To manage their complexity, artificial animal models are best organized
hierarchically, such that each successive modeling layer augments the more primitive
functionalities of lower layers. At the base of the modeling hierarchy (see Fig 1a), a
geometric modeling layer represents the morphology and appearance of the animal.
Next, a physical modeling layer incorporates biomechanical principles to constrain the
geometry and simulate biological tissues. Further up the hierarchy is a motor control
layer that motivates internal muscle actuators in order to synthesize lifelike locomotion.
Behavioral and perceptual modeling layers cooperate to support a reactive behavioral
repertoire. At the apex of the modeling pyramid is a cognitive modeling layer, which
simulates the deliberative behavior of higher animals, governs what an animal knows
about itself and its world, how that knowledge is acquired and represented, and how
automated reasoning and planning processes can exploit knowledge to achieve high-
1.2 An Artificial Life Modeling Paradigm for Medical Image Analysis
Viewed in the context of the artificial life modeling hierarchy (Fig. 1a), current
automatic deformable model-based approaches to medical image analysis include
geometric and physical modeling layers only (in interactive deformable models, such as
snakes, the human operator is relied upon to provide suitable behavioral level and
cognitive level support). At the physical level, deformable models interpret image data
by simulating dynamics or minimizing energy terms, but the models themselves do not
monitor or control this optimization process except in a most primitive way. At the
geometric level, aside from a few notable exceptions , deformable models are not
generally designed with intuitive, multi-scale, multi-location deformation ‘handles’.
Their inability to perform global deformations, such as bending, and other global
motions such as sliding and backing up makes it difficult to develop reasoning or
planning strategies for these models at the correct level of abstraction .
In more sophisticated deformable models, prior information is used to constrain
shape and appearance, as well as the statistical variation of these quantities [1,10];
however, these models have no explicit awareness of where they are and, consequently,
the effectiveness of these constraints is dependent upon model starting conditions. The
lack of awareness also prevents the models from knowing when to trust the image
feature information and ignore the constraint information and vice versa. The constraint
information is therefore applied arbitrarily. Furthermore, because there is no active,
explicit search for stable image features, the models are prone to latching onto incorrect
features  simply due to their proximity and local decision-making. Once this latching
prior knowledgeprior knowledge
Plan or schedulePlan or schedule
and limbsand limbs
other creaturesother creatures
medial based Shape
Muscle actuation causes
shape deformationshape deformation
medial based Shape
Muscle actuation causes Perceptual attention
Fig. 1. (a) The ALife modeling pyramid (adapted from ). (b) A deformable
organism: The brain issues ‘muscle’ actuation and perceptual attention commands.
The organism deforms and senses image features, whose characteristics are
conveyed to its brain. The brain makes decisions based on sensory input, memorized
information and prior knowledge, and a pre-stored plan, which may involve
interaction with other organisms.
occurs, the lack of control of the fitting procedure prevents the model from correcting
the misstep. The result is that the local decisions that are made do not add up to
intelligent global behavior.
To overcome the aforementioned deficiencies while retaining the core strengths of
the deformable model approach, we add high-level controller layers (a ‘brain’) on top of
the geometric and physical (or deformation) layers to produce an autonomous
deformable organism (Fig. 1b). The intelligent activation of these lower layers allows
the organism to control the fitting/optimization procedure. The layered architecture
approach allows the deformable organism to make deformation decisions at the correct
level of abstraction.
The perception system of the deformable organism comprises a set of sensors that
provide information. Any type of sensors can be incorporated, from edge strength and
edge direction detectors to snake ‘feelers’. Sensors can be focused or trained for specific
image features and image feature variation in a task-specific way; hence, the organism
can disregard sensory information superfluous to its current behavioral needs.
Explicit feature search requires powerful, flexible and intuitive model deformation
control. We achieve this with a set of ‘motor’ (i.e. deformation) controllers, which are
parameterized procedures dedicated to carrying out a complex deformation function,
such as successively bending a portion of the organism over some range of angles or
stretching part of the organism forward some distance.
The organism is ‘self-aware’ (i.e. knows where it and its parts are and what it is
seeking) and therefore it effectively utilizes global contextual knowledge. The organism
begins by searching for the most stable anatomical features in the image before
proceeding to less stable features. Once stable features are found and labeled, the
organism uses neighboring information and prior knowledge to determine the object
boundary in regions known to provide little or no feature information.
Because the organism carries out active, explicit searches for object features, it is
not satisfied with the nearest matching feature but looks further within a region to find
the best match, thus avoiding local minimum solutions. Furthermore, by carrying out
explicit searches for features we ensure correct correspondence between the model and
the data. If a feature cannot be found, the organism flags the situation. Subsequently, if
multiple plans exist, another plan could potentially be selected and the search for the
missing feature postponed until further information is available.
2 A Deformable Organism for 2D MR Brain Image Analysis
To demonstrate the potential of our framework for medical image analysis, we have
developed a deformable “worm” organism that can overcome noise, incomplete edges,
considerable anatomical variation, and occlusion in order to segment and label the
corpus callosum (CC) in 2D mid-sagittal MR images of the brain. We will now describe
in detail the layered architecture for this particular deformable organism.
2.1 Geometric Representation
As its name suggests, the deformable worm organism is based on a medial
representation of body morphology  that facilitates a variety of controlled local
deformations at multiple spatial scales. In this shape representation scheme, the CC
anatomical structure is described with four shape profiles derived from the primary
medial axis of the CC boundary contour. The medial profiles describe the geometry of
the structure in a natural way and provide general, intuitive, and independent shape
measures. These profiles are: a length profile
left (with respect to the medial axis) thickness profile
and N is the number of medial nodes. The
length profile represents the distances between consecutive pairs of medial nodes, and
the orientation profile represents the angles of the edges connecting the pairs of nodes.
The thickness profiles represent the distances between medial nodes and their
corresponding boundary points (Fig. 2, Fig. 3)1.
L m , an orientation profile
Tm , and a right thickness
O m , a
L mL m
0 50 100
0 50 100
0 50 100
Fig. 2. (a) CC anatomical feature labels overlaying a
reconstruction of the CC using the medial shape profiles shown
in Fig. 3. (b) Diagram of shape representation.
Fig. 3. Example medial shape
orientation, (c) left and (d)
right thickness profiles.
2.2 Motor System
Shape Deformation Actuators. In addition to affine transformation abilities (translate,
rotate, scale), we control organism deformation by defining deformation actuators in
terms of the medial shape profiles (Fig. 4). Controlled stretch (or compress), bend, and
bulge (or squash) deformations are implemented as deformation operators acting on the
length, orientation, or thickness profiles, respectively. Furthermore, by utilizing a
hierarchical (multiscale) and regional principal component analysis to capture the shape
variation statistics in a training set , we can keep the deformations consistent with
prior knowledge of possible shape variations. Whereas general, statistically-derived
shape models produce global shape variation modes only [1,10], we are able to produce
spatially-localized feasible deformations at desired scales, thus supporting our goal of
intelligent deformation planning.
Several operators of varying types, amplitudes, scales, and locations can be applied
to any of the length, orientation, and thickness shape profiles (Fig. 5a-d). Similarly,
multiple statistical shape variation modes can be activated, with each mode acting at a
specified amplitude, location and scale of the shape profiles (Fig. 5e-h). In general,
operator- and statistics-based deformations can be combined (Fig. 5i) and expressed as
1 Currently we construct medial profiles only from the primary medial axis and have not
considered secondary axes. This may prevent the CC worm organism from accurately representing
highly asymmetrical (with respect to the primary axis) parts of some corpora callosa. We also
realize that our medial shape representation needs improvement near the end caps. We are
currently exploring these issues, as well as issues related to the extension of our model to 3D, and
we intend to make full use of the considerable body of work of Pizer et al [6,7,8] on these topics.
dddls dlsdlst dlst
where p is a shape profile, d is a deformation type (stretch, bend, left/right bulge), i.e.
mL m O m Tm Tm
, p is the average shape profile, k is an
operator profile (with unity amplitude), l and s are the location and scale of the
deformation, t is the operator type (e.g. Gaussian, triangular, flat, bell, or cusp), ? is
the operator amplitude, the columns of M are the variation modes for a specific d , l ,
and s , and w contains variation mode weights. Details can be found in .
0 1020 30
0 1020 30
Fig. 4. Introducing a bulge on
the upper boundary of the CC
by applying a deform-ation
operator on the upper thickness
Tm . (a)
before and (c) after applying
reconstructed shape before and
(d) after the operator.
Deformation (Motor) Controllers. The organism’s low-level motor actuators are
controlled by motor controllers. These parameterized procedures carry out complex
deformation functions such as sweeping over a range of rigid transformation parameters,
sweeping over a range of stretch/bend/thickness amplitudes at a certain location and
scale, bending at increasing scales, moving a bulge on the boundary etc. Other high-level
deformation capabilities include, for example, smoothing the medial/left/right
boundaries, interpolating a missing part of the thickness profile, moving the medial axis
to a position midway between the left and right boundaries, and re-sampling the model
by including more medial and boundary nodes.
Fig. 5. Examples of controlled deformations: (a)-(c) Operator-
based bulge deformation at varying locations/amplitudes/scales.
(d) Operator-based stretching with varying amplitudes over entire
CC. (e)-(g) Statistics-based bending of left end, right end, and
left half of CC. (h) Statistics-based bulge of the left and right
thickness over entire CC. (i) From left to right: (1) mean shape,
(2) statistics-based bending of left half, followed by (3) locally
increasing lower thickness using operator, followed by (4)
applying operator-based stretch and (5) adding operator based
bend to right side of CC.
2.3 Perception System
Different parts of the organism are dynamically assigned sensing capabilities and thus
act as sensory organs (SOs) or receptors. The locations of the SOs are typically confined
to the organism’s body (on-board SOs) such as at its medial or boundary nodes, at
curves or segments connecting different nodes. In our implementation, the SOs are made
sensitive to different stimuli such as image intensity, image gradient magnitude and
direction, a non-linearly diffused version of the image, an edge detected (using Canny’s
edge detector) image, or even the result of a Hough transform. In general, a wide variety
of image processing/analysis techniques can be applied to the original image.
2.4 Behavioral/Cognitive System
The organism’s cognitive center combines sensory information, memorized information,
and instructions from a pre-stored segmentation plan to carry out active, explicit
searches for object features by activating ‘behavior’ routines. Behavior routines are
designed based on available organism motor skills, perception capabilities, and available
anatomical landmarks. For example, the routines implemented for the CC worm
organism include: find-top-of-head, find-upper-boundary-of-CC, find-genu, find-
rostrum, find-splenium, latch-to-upper-boundary, latch-to-lower-boundary, find-fornix,
thicken-right-side, thicken-left-side, back-up. The behavior routines subsequently
activate the deformation controllers to complete a stage in the plan and bring the
organism closer to its intention of object segmentation.
The segmentation plan provides a means for human experts to incorporate global
contextual knowledge. It contains instructions on how best to achieve a correct
segmentation by optimally prioritizing behaviors. If we know, for example, that the
corner-shaped rostrum of the CC is always very clearly defined in an MRI image, then
the find-rostrum behavior should be given a very high priority. Adhering to the
segmentation plan and defining it at a behavioral level affords the organism with an
awareness of the segmentation process. This enables it to make effective use of prior
shape knowledge – it is applied only in anatomical regions of the target object where
there is a high level of noise or known gaps in the object boundary edges, etc. In the next
section we describe the segmentation plan for the CC organism to illustrate this ability to
harness global contextual knowledge.
When a CC deformable worm organism is released into a 2D sagittal MRI brain image,
it engages in different ‘behaviors’ as it progresses towards its goal. Since the upper
boundary (Fig. 2a) of the CC is very well defined and can be easily located with respect
to the top of the head, the cognitive center of the CC organism activates behaviors to
first locate the top of the head and then move downwards (through the gray and white
matter) in the image space to locate the upper boundary (Fig. 6.1-5). Next, the organism
bends to latch to the upper boundary and activates a find-genu routine, causing the CC
organism to stretch and grow along this boundary towards the genu (Fig. 6.6-7). Once
the genu is located, the find-splenium routine is activated and the organism stretches and
grows in the opposite direction (Fig. 6.11). The genu and splenium are easily detected by
looking for a sudden change in direction of the upper boundary towards the middle of
the genu is well defined so it backs up and latches to the lower boundary (Fig. 6.8). It
then activates the find-rostrum behavior that tracks the lower boundary until it reaches
the distinctive rostrum (Fig. 6.8-10). At the splenium end of the CC, the organism backs
up and finds the center of a circle that approximates the splenium end cap (Fig. 6.12).
The lower boundary is then progressively tracked from the rostrum to the splenium
while maintaining parallelism with the organism’s medial axis in order to avoid latching
to the potentially occluding fornix (Fig. 6.13-14). However, the lower boundary may still
dip towards the fornix, so a successive step is performed to locate where, if at all, the
fornix occludes the CC, by activating the find-fornix routine (making use of edge
strength along the lower boundary, its parallelism to the medial axis, and statistical
thickness values). Thus, prior knowledge is applied only when and where required. If the
fornix does indeed occlude the CC, any detected dip in the organism’s boundary is
repaired by interpolating neighboring thickness values. The thickness of the upper
boundary is then adjusted to latch on to the corresponding boundary in the image (Fig.
6.15-17). At this point the CC organism has almost reached its goal; however, the medial
axis is not in the middle of the CC organism (Fig. 6.18), hence the medial axis is re-
parameterized by positioning the medial nodes halfway between the boundary nodes
(Fig. 6.19-20). Finally the lower and upper boundaries are re-located again to obtain the
final segmentation result (Fig. 6.21).
In addition, Fig. 7 demonstrates the detection and repairing of the fornix. Fig. 8
demonstrates the organism’s self-awareness. Fig. 9 shows other segmentation results and
several validated examples are also shown in Fig. 10.
Once the genu is found, the organism knows that the lower boundary opposite to
(4) (5) (6)
(1) (2) (3) (7) (8) (9)
(10) (11) (l2) (13) (14) (15)
(16) (17) (18) (19) (20) (21)
Fig. 6. Intelligent CC organism progressing through a sequence of behaviors to segment the CC.
(a) (b) (c)
Fig. 7. (a) Before and (b) after detecting and repairing the fornix dip. (c) The gradient magnitude.
Fig. 8. The CC organism’s self-awareness
makes it capable of identifying landmark parts.
Fig. 9. Example segmentation results.
Fig. 10. Example segmentation results (top), also shown (in black) over manually segmented
(gray) CC (bottom).
Robust, automatic medical image analysis requires the incorporation and intelligent
utilization of global contextual knowledge. We have introduced a new paradigm for
medical image analysis that applies concepts from artificial life modeling to meet this
requirement. By architecting a deformable model-based framework in a layered fashion,
we are able to separate the ‘global’ model-fitting control functionality from the local
feature integration functionality. This separation allows us to define a model-fitting
controller or ‘brain’ in terms of the high-level anatomical features of an object rather
than low-level image features. The layered-architecture approach also provides the brain
layer with precise control over the lower-level model deformation layer. The result is an
intelligent organism that is continuously aware of the progress of the segmentation,
allowing it to effectively apply prior knowledge of the target object. We have
demonstrated the potential of this approach by constructing a Corpus Callosum “worm”
organism and releasing it into MRI brain images in order to segment and label the CC.
Several interesting aspects of our approach are currently in consideration for
further exploration. These include extending our model to 3D, designing a motion
tracking plan and releasing an organism into time-varying image ‘environments’ (i.e. 4D
images), exploring the use of multiple plans and plan selection schemes, and exploring
the application of learning algorithms, such as genetic algorithms, to assist human
experts in the generation of optimal plans. Another potentially important research
direction is the use of multiple organisms that intercommunicate contextual image
information (i.e. are ‘aware’ of one another).
GH was funded in part by the Visual Information Technology (VISIT) program,
Swedish Foundation for Strategic Research (SSF). Dr. Martha Shenton of the Harvard
Medical School generously provided the MRI data.
1. Cootes, T., Beeston, C., Edwards, G., Taylor, C.: A Unified Framework for Atlas Matching
using Active Appearance Models. Image Processing in Medical. Imaging (1999) 322-333.
Duncan, J., Ayache, N.: Medical Image Analysis: Progress Over Two Decades and the
Challenges Ahead. IEEE Trans. PAMI 22(1) (2000) 85 -106.
Hamarneh, G., McInerney, T.: Controlled Shape Deformations via Medial Profiles. Vision
Interface (2001) 252-258.
McInerney, T., Kikinis, R.: An Object-based Volumetric Deformable Atlas for the Improved
Localization of Neuroanatomy in MR Images. MICCAI (1998) 861-869.
McInerney, T., Terzopoulos, D.: Deformable Models in Medical Image Analysis: A Survey.
Medical Image Analysis 1(2) (1996) 91-108.
Pizer, S., Fletcher, P., Fridman, Y., Fritsch D., Gash, A., Glotzer, J., Joshi, S., Thall A.,
Tracton, G., Yushkevich, P., Chaney, E.: Deformable M-Reps for 3D Medical Image
Segmentation. Submitted to Medical Image Analysis (2000).
Pizer, S., Fritsch, D., Low, K., Furst, J.: 2D & 3D Figural Models of Anatomic Objects from
Medical Images. Mathematical Morphology and Its Applications to Image Processing,
Kluwer Computational Imaging and Vision Series, H.J.A.M. Heijmans, J.B.T.M. Roerdink,
Eds. Amsterdam, (1998) 139-150.
Pizer, S., Fritsch, D.: Segmentation, Registration, and Measurement of Shape Variation via
Image Object Shape. IEEE Transactions on Medical Imaging 18(10) (1999) 851-865.
Shen, D., Davatzikos, C.: An Adaptive-Focus Deformable Model Using Statistical and
Geometric Information. IEEE Trans. PAMI 22(8) (2000) 906 -913.
10. Szekely, G., Kelemen, A., Brechbuehler C., Gerig, G.: Segmentation of 3D Objects From
MRI Volume Data Using Constrained Elastic Deformations of Flexible Fourier Surface
Models. Medical Image Analysis 1(1) (1996) 19-34.
11. Terzopoulos, D., Metaxas, D.: Dynamic 3D Models with Local and Global Deformations:
Deformable Superquadrics. IEEE Trans. PAMI 13(7) (1991) 703-714.
12. Terzopoulos, D., Tu, X., Grzeszczuk, R.: Artificial Fishes: Autonomous Locomotion,
Perception, Behavior, and Learning in a Simulated Physical World. Artificial Life 1(4) (1994)
13. Terzopoulos, D.: Artificial Life for Computer Graphics. Commun. ACM 42(8) (1999) 32-42.
14. Tsotsos, J.K., Mylopoulos, J., Covvey, H.D., Zucker, S.W.: A Framework for Visual Motion
Understanding. IEEE Trans. PAMI 2(6) (1980) 563-573.