Learning Center
Plans & pricing Sign in
Sign Out



									          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

                                                             Research Plan

                                                     A. Specific Aims (R21 Phase)

The overarching goal of this proposal is to develop a new set of tools that enable surgeons to easily design
and fabricate custom-fitting biocompatible implant scaffolds to support regrowth in regions of missing bone.
The majority of the implant design work will be an automated process that draws from patient CT data and a
database of related healthy bone data. The resulting implant model will be presented to the surgeon for
verification and minor modification, and then sent to an off-site facility for fabrication.

Figure 1. A computer-aided design of missing bone tissue (left) is used to specify the fabrication of a
biocompatible osteoconductive hydroxyapatite scaffold (center) used to support bone regrowth, demonstrated
by the bone cell (right) which attached to microscopic pores after three days in simulated body fluid.
The goal of the initial two-year R21 Phase of this proposal is to focus on the computer science objectives
necessary to support the computational science of automated implant design.

              (a)                                      (b)                                 (c)    (d)
Figure 2. (a) Injured mandible reconstructed from CT. (b) Healthy version of this mandible predicted from
database. (c) Manual merge resulting from non-rigid registration, regularized differencing and morphological
operations. (d) Resulting implant shape design for scaffold.
The purely computational aspects of our process for implant design are illustrated in figure 2. A geometric
representation of a (section of) mandible that has experienced resorption due to chronic periodontal disease is
shown in (a), extracted, segmented, and polygonized from a CT scan. A healthy (green) ―template‖ of a
matching region is constructed by interpolation and deformation of similar mandibles from a database, shown
in (b). The difference (ideally) between these two shapes is computed (c), and this difference is exported as a
shape description for fabrication of a bone implant scaffold (d).
While conceptually simple, the implementation of this idealized procedure is not simple. The construction of a
healthy match for an injured mandible requires a database indexed by efficient shape descriptors resulting
from a lengthy and detailed analysis of the variations of the mandible’s form. Furthermore, the differencing
PHS 398/2590 (Rev. 09/04)                             Page     26 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
between an injured mandible and a healthy reconstruction inevitably results in noise (shown in figure 3) that
must be filtered to produce a clean, usable implant shape specification.

Figure 3. The variation in shape of human mandibles (left), demonstrated by the aligned juxtaposition of seven
different mandibles segmented from CT-scans, and the noise (right) resulting from the difference between an
injured (green) mandible and a healthy (red) version predicted from a database match.
The R21 Phase of this proposal thus focuses on the following two fundamental contributions to shape
modeling for medical applications.

Specific Aim #1: Prediction of the healthy version of an injured mandible.
Our goal of providing computational assistance in the design of custom bone implant scaffolds is based on the
difference between the injured mandible and the healthy mandible. The injured mandible will almost always be
available whereas a CT-scan of the healthy mandible will likely not be available as such injuries are rarely
predictable. We will instead construct a database of healthy mandibles whose variations are indexed by
descriptors derived from carefully chosen landmark positions. The proposed work toward this aim thus seeks
to answer the questions:

    1. Along which principle axes does the shape of the human mandible vary?

    1. What is the minimum number of feature points (landmarks) it takes 1) to effectively describe a
       mandible’s shape; 2) to adequately deform a healthy mandible to fit a damaged one; and 3) where are
       these landmarks located?
    2. How many mandibles do we need in our database to produce effective predictions?

    3. How accurately can (and need) we predict a healthy mandible from a damaged mandible?

Specific Aim #2: Prediction of the healthy version of an injured mandible.
The second fundamental contribution focuses on the analysis and removal of the noise resulting from the
shape differencing needed to automate the custom design of an implant scaffold for the injury. This aim seeks
to answer the questions:

    1. How well can we discriminate between missing bone and alignment/prediction error between the
       injured CT scan and the healthy reconstruction?

    2. How much user interaction (if any) is needed to specify and refine the injured region?

    3. What are the statistical properties of the noise incurred by alignment and prediction error, and how well
       can this noise be filtered?

The conclusion of the R21 Phase of this proposal then seeks to deliver robust and efficient algorithms for bone
database construction and query, and for regularized shape differencing for implant specification. These
deliverables will feed into the R33 Phase of this proposal as the computational backbone for a turnkey system
PHS 398/2590 (Rev. 09/04)                             Page     27 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
engineered to increase the availability and quality of custom fabricated bone regrowth implants, while reducing
the costs in dollars and pain to the patient.

PHS 398/2590 (Rev. 09/04)                             Page     28 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

                                B. Background and Significance (R21 & R33 Phases)

The replacement of bone and restoration of structure and function in severely injured patients can present
formidable challenges to even the most skillful surgeons. An example of such an injury is shown in figure 4.
This is a battlefield injury incurred several months ago by an American soldier serving in Iraq and currently
being treated by one of us (Goldwasser). This type of injury is unfortunately not unusual in the current conflict
and is representative of the extent and complexity of bone loss in such patients. Rapid evacuation of injured
soldiers to mobile surgical hospitals and improvements in body and head armor, have reduced the lethality of
battlefield injuries. In all previous conflicts dating to World War I, the lethality of combat wounds experienced by
U.S. soldiers was approximately 25 percent. In the current Iraq war, lethality of battlefield injuries has been
dramatically reduced to 9 percent (Gawande A. 2004). Approximately half of these non-lethal injuries involve
injuries to the head, neck, and extremities. Complex injuries of this type, involving both hard and soft tissues,
often require extensive reconstruction procedures to replace the missing tissue.

Figure 4. Photograph (a) and CT image (b) of a patient with extensive bone and soft tissue loss resulting from
blunt force trauma.

It is worth to mention the RP4Baghdad project of the rapid prototyping industry to support severely injured
people in Iraq ( The mission of this project, to provide custom made implants, would be
immensely helped by an automatic tool as the one we propose here.

These non-emergent procedures typically occur weeks or months after the patient has been stabilized and
frequently occur in stages. The current "gold standard" for surgical repair of bone loss involves the use of
autograft bone, i.e. bone taken from another site in the patient’s body. The surgical procedures for harvesting
such bone can result in complications that are "minor" (hematoma, temporary sensory loss, acute pain); or
"major" (permanent sensory loss, chronic pain, infection, and gait disturbances). Complication rates exceeding
30% have been reported for autograft harvesting from the iliac crest of the pelvis, a common source for
autograft bone (Younger, E. et al. 1989).

An alternative to autografting is implantation of cadaver bone. From a clinical perspective this is an even less
attractive option that increases the complication rates associated with autografting and adds the risk of disease
transmission. It is estimated that 500,000 bone graft surgeries are performed annually in the United States.
Approximately 90% of the procedures utilize either autograft or allograft. The estimated cost of these
procedures approaches $2.5 billion per year (
substitute-materials.htm ). A conservative estimate of the healthcare cost savings resulting from use of custom
implants made from a bone graft substitutes of the type we propose exceeds $600 million per year. When the

PHS 398/2590 (Rev. 09/04)                             Page     29 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
indirect costs of complications and recuperation are included, this number exceeds $1 billion per year in the
U.S. alone.

There are a large number of papers that address scaffold engineering methods, but very few, to our
knowledge, discuss the technological issue of building a software tool that can be used by a surgeon for the
scaffold design. For instance, in Hollister S.J. et al. 2002, and Hollister S.J. et al. 2005, the authors aim at
developing an integrated approach for engineering craniofacial scaffolds and at demonstrating that these
engineered scaffolds have mechanical properties in the range of craniofacial tissue and support bone
regeneration for craniofacial reconstruction. However, they do not address the software issue, using
unspecialized commercially available software (Analyze, IDL) for (non-automated) implant design.

Dean D. et al. present a system for computer-aided design and direct computer-aided manufacture of large-
format cranioplasties (Dean D. et al. 2003). Similarly to our aim, their system uses CT data and requires no
specialized training. On the other hand, their thin-plate spline modeling system is suitable for cranial plates
(large low curvature surfaces) but cannot model more complex shapes as our modeling system can, being
based on mesh morphing and regularized volumetric difference operations. While their system uses a right-to-
left mirrored or average skull surface template for the reconstruction, we intend to build and use a population
atlas since better matching and more complex bony defects reconstructions are thus possible and necessary
for the more complex geometries found in lower parts of the skull.

Although there exist numerous software tools that are related to our work, in the sense that they help the
custom design and fabrication of bone implants, none of them, to our knowledge, addresses the key goal of
our proposal: to develop an automated implant design tool that requires little user input and no 3-D modeling
skills (i.e. surgeon-friendly) so that surgeons can easily and quickly design and have fabricated custom-fitting
implants. Further, our tool is distinguished by its use of a population bone atlas as the source material for
generating an implant that fits the missing part as closely as possible.

For exemplification, it is important to mention several reputed companies and their relevant products for the
medical community:

Anatomics ( of Queensland, Australia produces stereolithography models of a patient's
anatomy to assist in the preoperative planning and the actual surgery. They use their own BioModelling
technology to build custom titanium implants (J. F. Arvier et al. 1994). Their tool, called BioBuild, is not an
automated design tool, but instead supports manual design and technological expertise, experience with 3-D
modelling, and training (i.e. it is not surgeon-friendly).

Medical Modeling ( of Golden, Colorado produces patient-specific
stereolithography models to aid in complex reconstructive surgeries. They produce software for medical data
visualization, 3D printing and surgical simulation and some advanced tools for model editing, but the implants
are manually fabricated.

Materialise ( produces Mimics, that it is described on their web site as ―a fully integrated,
user-friendly 3D image processing and editing software based on scanner data‖. Their relevant application,
"Cranioplasty: Digital Shaping Of the Prosthesis and Automatic Building of Wax Model of Prosthesis", uses
mirroring techniques for the implant design. The mirroring techniques are useless in bilateral injuries. Only the
statistical atlas based shape modeling can handle well these cases. Moreover, one can infer from the Mimics
software specifications that it is not a turnkey program. It is similar to the Amira software, on which Jerez and
Stroila have a very good working knowledge. One should not expect a surgeon to go through the steep
learning curve required for a practical use of these software packages, given the limited available time and
rapid pace of his profession.

In conclusion, many medical visualization tools exist nowadays, but very few are specifically for medical
modeling, and none are automated. If this concept is to become a viable clinical tool in common practice, it is
essential that busy surgeons are not required to spend months learning a general-purpose software package
mostly intended for other tasks. Instead, we must almost entirely automate the design process, leaving only

PHS 398/2590 (Rev. 09/04)                             Page     30 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
the subtle refinements for the surgeon to tackle. This is where the computer graphics expertise (especially in
shape modeling and mesh processing/painting) of one of our members (Hart) comes to play a very important
role. Using Hart’s expertise to automate those tasks not already part of a surgeon’s toolkit will enable those
surgeons to concentrate on their specialty—surgery.

PHS 398/2590 (Rev. 09/04)                             Page     31 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

                                C. Preliminary Studies/Progress Report (R21 Phase)

Our work on the automatic detection of mandible injury regions and the automatic construction of a
replacement implant shape has received bootstrap funding for one year by the University of Illinois’ Vice
Chancellor for Research as a Critical Research Initiative.

In this first year of funded research, we have created a small collection of CT scans containing healthy
mandibles through Carle Foundation Hospital. We have also extracted and created polygonization of the
mandibles from the CT data using the image segmentation and surface reconstruction modules of Amira

In Fall, 2005, Hart, Stroila and Grosser met with maxillofacial surgeons at Goldwasser’s practice for a
demonstration of current software tools for planning surgery that constructed anatomical segmentation curves
based on the placement of landmarks on x-rays. As a result, Stroila developed a similar method in 3-D for
placing landmarks on the reconstructed polygonal models of mandibles in the current database. These
landmarks serve as feature points used to measure the statistical variation in mandible shape and for non-
linear transformation (morphing) between two different mandibles.

Based on our progress we have successfully prototyped a proof-of-concept demonstration of the two main
components of automatic implant design:

    1. The morphing algorithm (Figure 5), which takes the closest-match healthy mandible and subtly deforms
       it to match the damaged mandible. This is programmed in C++, using the VTK library ( for
       visualization, the GSL library ( for numeric computations, and the QT library
       ( for graphical user interface.

    2. The subtraction algorithm (Figure 2), which subtracts the damaged mandible from the morphed healthy
       mandible, producing the implant shape. This is programmed as a network of data sources and data
       processing filters using the visualization and image processing software Amira.

Figure 5. Superposition of a damaged jaw (red) and a morphed template (green). The missing bone is detected
as the large green mass found by differencing the injured mandible with the predicted healthy mandible, but
without regularization, this differencing yields a significant amount of geometric noise that must be filtered
without altering the desired shape of the implant.

PHS 398/2590 (Rev. 09/04)                             Page     32 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

                                     D. Research Design and Methods (R21 Phase)

Database Construction
We will collect and maintain a shape database consisting of a range of healthy skull bones. We represent the
morphology of these mandibles by the selection of a small number of currently five ―landmark‖ feature points,
as shown in the figure below (Figure 6), placed at the left/right mandibular condyle, the left/right gonial angle
(also known as the angle of the mandible) and the menton.

Figure 6. Five landmarks are currently used to indicate mandible morphology, shown chosen manually (left)
but can also be chosen automatically based on shape statistics (center). These landmarks, once measured
and aligned across a number of different mandibles, indicate the range of variation in mandible morphology

We then build the ―shape space‖ of individual variations of the mandible using the principal component analysis
(PCA) method. In this case, we align the mandibles by translating their center of mass to the origin and
orienting them to minimize their variation. We then plot each of these mandibles as a single point in the 15-
dimensional space corresponding to the three coordinates (x,y,z) of the five landmark point positions. Though
the dimensionality of this space is much higher than the 3-D spaces we are used to visualizing, the distribution
of the points in this case only varies significantly in three dimensions, scattering across a 3-D box embedded in
this higher dimensional space.

Figure 7. First four principle component axes (in order of decreasing magnitude, plotted by extremes and
averages of landmark position) indicating three major parameters affecting mandible shape, whereas the fourth
and further axes indicate minor/subtle differences in mandibles.

The figure above demonstrates these ―axes of variation‖ corresponding to the three directions over which the
points vary significantly in this 15-D space. PCA extracts these axes as the three 15-D eigenvectors
corresponding to the three largest eigenvalues of the cross-correlation matrix of the landmark points. The
fourth eigenvalue is much smaller so it and subsequent axes represent less important variations and in this

PHS 398/2590 (Rev. 09/04)                             Page     33 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
case can be ignored. Thus the location of five ideally bilaterally symmetric landmark positions defines three
PCA coordinates used to organize, index and navigate the space of human mandible variations.

Based on the shape ―fingerprint‖ of the scanned injured bone (the coordinates in the PCA shape space), the
program will search the database for the closest healthy bone. Then it aligns and rescales the healthy bone to
best match the proportions and shape of the injured bone, using the ―Procrustes‖ alignment method available
in the Visualization Toolkit (VTK), or through a level-set evolution on the volume data. (D.E. Breen et al. 2001)

Mandible Interpolation
Given the unique qualities of every individual, we will not find an exact match in the database of mandibles, but
will instead find one or several close mandibles. We will then need to interpolate and deform these mandibles
to construct a healthy match to the injured query.

Styner, M.A. et al. present a comparative study of diverse landmark-correspondence methods for the statistical
shape analysis of anatomical regions of interest and model building, in the framework of the statistical shape
models (SSMs) (Styner, M.A. et al. 2003). According to this study, the minimum descriptor length method
(MDL) (Davies, Rh.H. et al. 2002) and the covariance determinant (DetCov) (Kotcheff, A.C.W. et al. 1998) are
the best correspondence methods. The major drawback of these methods is that they use spherical
parameterizations of the meshes in the training population. This type of parameterization is suitable for round
shapes like the femoral head presented in that paper, but no robust implementations exist in the case of more
complex shapes.

Instead, we plan to design and implement an automatic segmentation and atlas-building method based on
deformable models (Michael Leventon et al. 2000). A template shape is obtained from a deformable model
using forces derived from the image data. These forces are applied to an initial mesh, simulating the elastic
deformation of a model until it fits the anatomical structure of interest.

We will also investigate the possibility of using statistical deformation models. There are several advantages in
using non-landmark statistical atlas construction, although most software libraries use landmark-based
methods at the moment. Rueckert D. et al. introduce a new shape analysis method that eliminates the need of
landmarks, based on a new concept called the statistical deformation model (SDM) (Rueckert D. et al. 2003).
The concept of SDM is similar to statistical shape models, but offers several advantages: it can be constructed
directly from images without the need for segmentation, by using a non-rigid registration algorithm based on
free-form deformations and normalized mutual information to compute the deformations required to establish
dense correspondences between the samples; secondly, SDMs allow the construction of an atlas of the
average anatomy as well as its variability across a population of subjects; and finally, SDMs take the three
dimensional nature of the underlying anatomy into account by analyzing dense three dimensional deformation
fields rather than only information about the surface shape of anatomical structures. The modularity of our
software design will permit the use of several alternative methods and easy updates.

An aligned comparison between the healthy and damaged bones will allow us to automatically detect and
indicate the damaged portion of the bone. We can then perform a volumetric subtraction of the injured bone
from the healthy bone to yield a geometric representation and specification of the desired implant shape.

The subtraction is performed as follows. The healthy model is morphed into the damaged model through an
energy minimization process that ignores the previously outlined damaged areas, using a similar method to
one implemented in Allen, B. et al. 2003. Then, binary volumetric images for both the healthy and the damaged
models are constructed using a three-dimensional rasterization algorithm. The image of the damaged model is
subtracted from that of the healthy model and then automatic morphologic operations are applied to smooth
out the noise from the final model of the implant.

This simple differencing inevitably yields a large amount of material due to small errors between the actual
injured mandible and the database-assisted prediction of its healthy version. These errors, when sufficiently
small are removed through a process called regularization, which in the case of volumetric data, can occur
PHS 398/2590 (Rev. 09/04)                             Page     34 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
through the 3-D voxel morphological operations of opening and closing. Opening is a one-voxel contraction
followed by a one-voxel dilation whereas closing is its opposite. If the error is sufficiently small and scattered, it
will not survive an opening operation. Larger errors will require more significant morphological operations such
as openings that use dilations and contractions larger than one voxel. The opening operation will also round off
sharp corners in the detection of the injured region revealed by differencing, but these features can be restored
by extraction of the original connected component resulting from the raw, un-regularized difference.

The computational kernel of our software will be based on scientifically ascertained open source software
libraries: the Insight Segmentation and Registration Toolkit (ITK, (Terry Yoo (ed.) 2004) for image
processing and the GNU Scientific Library (GSL) for numerical computations. We will also use the Visualization
Toolkit (VTK) for visualization and modeling and the QT library for graphical user interface components.


At the end of the first year of support we expect to have at least prototype implementations of all four of our
main objectives:

    1. Construction of an anthropomorphic human skull database consisting of about 100 samples.

    2. Parameterization of the space of mandibles via principle component analysis of landmark positions.

    3. Interpolation and deformation of the mandibles most similar to an injured ―query‖ mandible to form a
       healthy template.

    4. Implementation of an automatic regularized differencing algorithm.

This will allow us to focus the second year on evaluating and analyzing their accuracy and robustness, to

    1. Statistical analysis of the mandible database to determine its representational accuracy to verify its
       population and distribution is sufficiently dense.

    2. Measurement of the accuracy of the predicted template to verify the landmarks are of sufficient number
       and placement, the PCA parameterization contains a sufficient number of axes, the deformations are
       sufficiently expressive and the approach itself is sound.

    3. Experimentation to find the ―breaking point‖ of the automatic regularized differencing to measure its

The second year furthermore provides a buffer in the event that unforeseen complications impede our progress
on the initial four objectives.

Metrics to assess project success

Do we have enough mandibles? We plan to answer this question by randomly separating it into a 90%
portion for ―training‖ and a 10% portion for ―verification,‖ and constructing template matches of each
―verification‖ mandible using only mandibles in the ―training‖ database. We can then use the resulting accuracy
to gauge the representational accuracy of the database to determine if it contains enough mandibles and the
distribution of mandibles is sufficient.

Do we have enough landmarks and are they in the right places? The criteria for good landmark position
choice is a combination of how pronounced the geometric features are so that they are sufficiently easy to find,
and how well the landmark positions correlate with the axes of PCA analysis. If the number of landmarks

PHS 398/2590 (Rev. 09/04)                             Page     35 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
differs from the number of axes, we will need to find a new collection of landmarks whose number is consistent
with the number of significant axes of variation revealed by PCA. We will find the position of these new
landmarks based on the correlation of feature position variation with PCA axis variation.

Is the interpolation and deformation physically plausible? The laws of structural mechanics and
physiology demand that we not deform a diminutive mandible into a Herculean mandible, which prompts the
need for a database of mandibles to capture the structural and physiological differences across a population.
Some interpolation and deformation will be necessary as the nearest neighbor to a given injured mandible in
our database of healthy mandibles will not be exact. We will measure the error incurred by this deformation by
morphing mandibles into their nearest neighbors in the database, and peaks in this error will indicate gaps in
the distribution of mandibles in our database.

Does regularization of the differencing damage the implant shape? Regularization removes the noise
resulting from the approximation of the injured mandible and the database-generated template fit to it. The raw
difference with a good template fit will yield uniformly distributed uncorrelated noise occurring at a dominant
spatial frequency near the sampling frequency which can be reliably removed through morphological operators
without damage to the main differenced shape. Errors with higher levels of correlation will require more
significant and drastic operations to filter out the resulting differencing noise, and these operations can
potentially damage the main difference shape. Analysis of the distribution of noise resulting from template fit
errors will better indicate the level of filtering required and the potential impact on the main difference shape.

What is the actual and necessary accuracy of the process? Our goal of implant precision will be 1mm
which is consistent with the resolution of the CT scan and our initial experiments. This goal is well within the
precision of existing autograph techniques which implant bone harvested from other locations on the body,
often in several pieces as a specific shape and size is not usually available from an individual’s sufficiently
redundant bone.

Figure 8. Example of the current procedure that our work on bone implant design will replace. Bone is
harvested (left) elsewhere in the body, and sculpted into an implant shape relying on the skill of the surgeon.
The center image shows how scaffolds, at the time not approved for implant, were used as guidelines for this
sculpting process. The harvested bone pieces are then fastened into the injured mandible (right) indicating that
our goal accuracy of 1mm falls well within (and in fact greatly improves) currently acceptable practices.

PHS 398/2590 (Rev. 09/04)                             Page     36 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

                                                    A. Specific Aims (R33 Phase)

Building on the computational tools of the initial R21 Phase, specifically the ability to synthesize a clean bone
replacement design given only an injured mandible, the R33 Phase of this proposal seeks to reintegrate these
tools with progress in scaffold materials composition, structural simulation and anatomical accommodation,
and to transfer this technology into a complete turnkey system suitable for medical deployment suitable for on-
site hospital and clinic use.

Figure 9. The figure at left demonstrates the presentation of an injured mandible with controls that allow the
surgeon to “paint” the desired implant/bone interface region. The figure at the right demonstrates the inclusion
of anatomical accommodation in implant design, in this case a ridge in the implant to provide a pathway for the
inferior alveolar nerve.

Specific Aim #1: Software and usability engineering of a turnkey interactive platform.
What platforms for software, hardware, I/O and the user-interface will best support the oversight and input of
the surgeon in a medical environment? For example, how best can we provide a 3-D interface to allow the
surgeon to carefully delineate the bone scaffold placement in areas requiring delicate anatomical positioning?
What user interface conventions for data manipulation and examination are best suited for the medical
environment, and more specifically for a surgeon without 3-D modeling skills? Surgeons bring unique
knowledge and understanding about the anatomical characteristics of the subjects they will be working with.
How can we build an interface that leverages that knowledge in an intuitive manner for the intended user? How
will data be effectively transferred between CT scans, medical databanks and scaffold construction devices
within given medical informatics infrastructure? The proposal includes a team of computer scientists led by the
PI consisting of a research scientist, two research assistants and a research programmer devoted to the
engineering of such a system.

Specific Aim #2: Accommodation of anatomical, structural and material feasibility.
Our goal is to provide automated tools that allow less reliance on the non-medical (e.g. sculpting) skills of the
surgeon for implant design. The R21 Phase of the proposal focuses only on the geometry of the implant,
whereas the successful design of a functioning implant must also take into account anatomical features (e.g.
nerve pathways), structural support (e.g. for proper attachment with enough strength to support chewing) and
materials consideration (e.g. to avoid impractical configurations during fabrication).

PHS 398/2590 (Rev. 09/04)                             Page     37 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

                                C. Preliminary Studies/Progress Report (R33 Phase)

Members of our team have previously used available general-purpose interactive medical visualization and
computer-aided design tools (e.g. Amira, Rhino) to design a custom implant for a patient who had experienced
resorption of mandibular bone as a consequence of chronic periodontal disease (Grosser, B. et al. 2004). This
section describes the steps of that process as evidence of the feasibility of computational design towards a
working bone implant scaffold, and as motivation to automate this process to deliver this level of care more
rapidly to a broader segment of patients in need of such care.

From this experience we learned that implant design is an expensive, team effort requiring extensive
knowledge and skill in computer-aided design tools, medical imaging, anatomy and physiology, static structure
and even material science and cell biology. The goal of the R33 Phase of this proposal is to encapsulate this
multi-disciplinary expertise into a software system designed to allow a single surgeon in the field to expertly
specify a custom fabricated bone implant given a single CT-scan of the injured region.

Using the case of a 73 year old female who has experienced severe bilateral bone loss in the mandible,
materials scientists, engineers, medical 3D artists, computer-aided designers, and the patient's attending
physician created a workflow by which a synthetic ceramic scaffold was designed, and fabricated specifically
for this patient. This workflow involved true collaboration between all parties involved, as the surgeon sought to
transfer his intuitive knowledge of the precise structure of the implant to the 3D modelers at the Imaging
Technology Group (ITG) of the University of Illinois Beckman Institute, who then in turn transferred their work
to the fabricators of the implant at Sandia.

ITG constructed a 3-D polygonized model of the damaged mandible from a CT scan, in consultation with CT
technicians at Carle to ensure the accuracy of the CT data measurements and to establish that the extracted
model was true to the original data. Dr. Goldwasser then worked with ITG in defining the boundaries for the
implant and making accommodations for an existing nerve.

Figure 10. Sketches by Sinn-Hanlon and Goldwasser defining the path of the inferior alveolar nerve (left) and
the boundaries of the implant (right).

ITG then created a computer-generated 3D model whose bottom surface precisely fit the eroded mandibular
surface on which it would rest. A canal was built into the ventral surface of the implant that was large enough to
accommodate the exposed nerve, but would leave an adequate amount of contiguous surface on either side
for jaw strength and the insertion of screws to anchor the implant into the mandible. The top surface of the
implant was modeled with the intent of restoring the natural shape of the jaw and providing a surface that
would support dentures. The two surfaces were welded together to complete the model for the implant.

PHS 398/2590 (Rev. 09/04)                             Page     38 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

Once the implant was created, the model was 'printed' as a 3-D plaster sculpture using ITG's rapid prototyping
stereolithography deposition printer so that the researchers could evaluate the fit of the implant with a physical
model. Evaluation of these models concluded that the fit was very precise and well within the tolerances
required. The 3D computer model was then e-mailed to Sandia, accompanied by shipped physical prints of the
jaw and implant for reference.

Figure 11. The final implant model (left) and a physical 'print' of the jaw and implant model (right).
Once received at Sandia, they proceeded to investigate methods for fabrication of the object using their
patented process called 'robocasting'--a technology similar to the rapid prototyping machine used by ITG, but
unique in its ability to work with various special materials. In this project, the device is used to create scaffolds
of a substance primarily made up of hydroxyapatite--a substance chemically identical to those found in human
bone. These scaffold structures, developed by Dr. Jamison, are unique in their ability to withstand the extreme
forces that a bone implant would undergo.

Figure 12.The final implant scaffold was test fit in the plaster reconstruction of the jaw (left two images). The
underside of the implant scaffold shows the canal modeled to accommodate the nerve path (right).
Finally, during the patients' previously scheduled autograft procedure, the implant was sterilized in an
autoclave and inserted into place for fit testing. The surgeon proclaimed the implant to ―fit like a glove‖.

PHS 398/2590 (Rev. 09/04)                             Page     39 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

Figure 13.The final implant scaffold fit tested in the patients jaw.

The task of developing a device for bone replacement in the mandible carries with it the additional requirement
of high strength. The pressure transmitted through the teeth to the mandible during chewing can exceed 400
pounds per square inch. Consequently we have designed the scaffold for this application to exceed the
strength of the natural bone it replaces. Current research is directed toward understanding the change in
strength that occurs as the scaffold is broken down over time in the body and is built up by ingrowth of new
bone tissue. The purpose is to ensure that adequate strength is maintained throughout the period of bone

This initial proof-of-concept provided the team with several pertinent points that motivate this proposal. First,
the final fit test showed that a virtual model could be developed from CT scans that truly fit fine anatomical
structures. Second, it illuminated the cross-disciplinary activity required to produce such an object. The 3-D
modeler could not have built the implant without the knowledge of the surgeon, just as the surgeon could not
have built the implant without 3-D modeling skills. Third, the tangibility of a custom implant not only illustrated
the feasibility of using implants in this manner, but also showed that such implants could exceed current
autografting methods (see Figure 8). Each of these points motivates the team to produce a simple tool that will
enable surgeons to design and use custom-implants in the field without learning all the other skills necessary
to fabricate the device.

PHS 398/2590 (Rev. 09/04)                             Page     40 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.

                                     D. Research Design and Methods (R33 Phase)

Proposed Software System
The proposed software system begins with an automated comparison and matching of the patient’s healthy
skull bones to a set of normal skull bones. This database is built in the earlier R21 Phase of the project and
represents a morphometric anatomical atlas of variations in the human skull. The software then builds a
template of the injured bone by a semi-automated process that involves several steps. The software then
guides the surgeon through a mostly-automated implant design process, pausing to let the surgeon make
minor adjustments as necessary. This process involves several steps:

    1. The first step consists in retrieving the patient data from the hospital’s picture archiving and
       communication system (PACS). The patient’s injury is scanned using a commonly available method
       including computed-tomography (CT) or magnetic resonance imaging (MRI) to construct a 3-D
       volumetric array of the affected region, represented in DICOM data format. This scan is transferred
       from PACS through a secure network to the surgeon’s treatment planning computer.

    2. The second step is skull classification and extraction. The volumetric data is a 3-D array of values that
       reflect either the x-ray absorption (CT) or the water content (MRI) of the scanned skin, bone, and
       tissue. We prefer and will focus on CT data as the classification and segmentation of bone is easier
       from such data. From this we will determine which portions of the volumetric dataset correspond to the
       skull and will use an atlas-based segmentation of the bone(s) of interest. Then, we will reconstruct a
       CAD solid model representation of the appropriate bones of the skull (e.g. the mandible) using a
       combination of our own custom software (John C. Hart et al. 2002) and a model-based segmentation
       algorithm that uses the database to identify the desired bone through a template matching and fitting

    3. We construct a template model of the bone of interest using the closest match from our database and
       appropriate morphing to match the damaged bone. This healthy bone model is used, which was also
       used to guide segmentation of the damaged bone, is used to perform the regularized (noise-free)
       differencing algorithm developed in the R21 Phase to construct an initial prototype implant design.

    4. The surgeon is asked to evaluate the initial implant design. If the automated differencing approach fails
       to produce an adequate implant prototype, the surgeon will ―paint‖ the damaged area onto the injured
       bone model to more precisely indicate the damaged region. This area can be projected back to the
       template mandible model, and a new design can be constructed from the union of the two.

    5. If necessary, the surgeon paints regions of the prototype that should be eroded for anatomical
       accommodations of nerves and other elements. We have recently completed the implementation of a
       system that supports the ―level-set‖ erosion of a meshed surface as part of an unrelated project that we
       will apply to this application. The interface for this erosion method is intuitive, and allows the user to
       manipulate 3-D objects in a clay-like manner, pushing and pulling sections as needed.

    6. The implant design result is decomposed into a finite element model for computation of its structural
       strength. Structurally weak regions are indicated and a prototype modification is provided, as is a
       painting interface that supports dilation of structurally weak regions.

    7. The implant is likewise analyzed for fabrication, identifying areas that would confound the fabrication
       process, again offering an automatic solution accompanied by a painting interface for interactive

    8. The final step is the output and electronic transmission of an approved, satisfactory implant design for
       offsite fabrication and shipment.

PHS 398/2590 (Rev. 09/04)                             Page     41 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
The visualization and manipulation of a three dimensional environment during this ―painting‖ process
constitutes a challenge that is very well addressed with the aid of the Computer Graphics (CG) tools. Painting
on a three dimensional model is challenging even for the experienced computer graphics artists. Because of
this, well-established CG methods simplify this task by decomposing the three dimensional model’s surface
into patches that can be projected with minimal distortion onto a plane where painting is a much easier task;
then, the painted areas are assembled and projected back to the three dimensional surface (Nathan A. Carr et
al. 2004).

Software and Usability Engineering
Development of such production software systems commonly requires that more than half the total effort be
devoted to the user interface. The software and usability engineering lifecycles follow the same waterfall
models composed of the steps of Requirement Specification, Architectural Design, Detailed Design,
Implementation, Evaluation and Maintenance. We have already completed the first two steps of this process.
The first was forming the requirements specification summarized by this proposal and its articulation of the
surgeon’s needs and the measures of a successful scaffold design process. The second was by forming an
initial architectural design embodied by the process flowchart. This leads to the third step, the detailed design
of the interface, which is a main objective of the proposed research. This detailed design will require a detailed
human- computer interaction (HCI) modeling of the task (including both cognitive and motor models of the
surgeon-user) and a model for the interaction (including the surgeon-computer dialog). We will design the
interface paying special attention to the HCI principles of familiarity (using terms and tools familiar to the
surgeon); recoverability (so the surgeon can ―undo‖ a mistake multiple times as necessary, to facilitate the
experimentation of different solutions to a difficult bone-reconstruction problem); and task conformance (so the
surgeon understands precisely the limits of what bone reconstruction problems the software is capable of
solving). Once complete, we can then implement this detailed design using existing graphical user interface
libraries, and test it on a variety of scenarios constructed from data (e.g. MRI/CT scans) available from existing
bone repair case studies.

Figure 14. Process flowchart indicating the main components of our software system.

We will measure the effectiveness of the interface through user studies, initially through dialog and feedback
from the surgeon on our team, and eventually through a formal evaluation with multiple surgeons. These
evaluations will be used at each stage of the development to verify and validate our design choices in the
construction of the interface. We will also investigate the effectiveness of a variety of input devices to provide
the surgeon more effective control over the design process, ranging from the ubiquitous mouse to the precision
of a graphics tablet to the multi-modality of a 3-D force-feedback stylus, while maintaining the level of cost,
familiarity and utility needed for clinical implementation.

The development of a clinical system for scanning, matching, processing, and synthesizing bone structures will
provide our research groups with a driving application for further scientific and engineering development of the
state of the art.
PHS 398/2590 (Rev. 09/04)                             Page     42 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
During the R33 phase of support we will test the toolkit in hospitals to evaluate its effectiveness on its target
user population, and use the feedback for further interface refinement. No actual patients will be treated in this
alpha stage. We will perform tests for accuracy using error analysis and incorporate validation principles
according to the FDA guidelines.

Following the FDA Guidelines for software development outlined in "General Principles of Software Validation;
Final Guidance for Industry and FDA Staff" ( ), we will
introduce from the beginning of the R33 phase the following essential components:

1.   Continuous Testing via Dashboard (Ctest-Dart) ( )
2.   Version control (Subversion) ( )
3.   Configuration standardization (CMake) ( )
4.   Bug tracking (phpBugTracker) ( )
5.   Continuous Documentation (Doxygen) ( )

Software quality assurance should consider preventing the introduction of defects into the software
development process. It does not suffice to perform quality tests of the software code after it is written. Due to
the high complexity of the automatic tool we propose, an exhaustive testing of the final product is very difficult.
Therefore, we will begin early preparation for software validation, during the first phase of R33, the one
consisting in the design and development planning in the process of technological transfer, from concepts to
the actual implementation.

The project will follow the usual waterfall software life cycle model and will contain the software engineering
tasks and documentation necessary to support the software validation effort. It will also contain specific
verification and validation tasks that are appropriate for the intended use of the software.

We will outline a validation plan, which specifies the scope, approach, resources, schedules and the types and
extent of activities, tasks, and work items. We will also outline the validation procedures, which identify the
specific actions or sequence of actions that must be taken to complete individual validation activities, tasks,
and work items.

Complete regression testing (CTest-Dart) will be designed and performed after each software change, in order
to validate not only the individual change, but also to evaluate the extent and impact of that change on the
entire software system. At the end of the R33 phase, we will provide complete validation plans and procedures
documentation. This will allow independent reviewers to perform thorough validation procedures after the beta
testing phase. This beta testing phase, the delivery of a final product and plans for software maintenance fall
beyond the scope of this proposal and will need to be the focus of subsequent work.

Teamwork Plan
The R33 phase depends intrinsically on collaboration between experts in computer aided design, maxilofacial
surgery, medical imaging, (bio-)materials science and (bio-)mechanical engineering. It is the diversity of this
team that has led to the unique results achieved in the two proof-of-concepts completed in preparation for this
proposal. Having an investigator with intimate knowledge of each component of the final system—from the
development of the computational tools, to the production and validation of biocompatible bone scaffold
materials, to the use of these scaffolds in clinical applications—is what will make the final result a viable clinical
tool that will be adopted for use in the field. We propose to support this collaboration through the employment
of a key multidisciplinary research scientist and shared multidisciplinary research assistants and research

        One research scientist, (Stroila) will serve as the main project manager, overseen by lead-PI Hart.
         Stroila is currently supported as a post-doc on projects that will expire this summer (the one-year
         campus funding of this project and an unrelated NSF project) which makes the timing of this proposals
         funding quite critical. We have thus established an aggressive start-date of September 1, 2006 to
         provide maximum benefit from our current momentum, as his unique combination of expertise in
PHS 398/2590 (Rev. 09/04)                             Page     43 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
         mathematical modeling, computer science and medical imaging would be quite difficult to replace if his
         position expires and he is forced to seek other employment possibly elsewhere.

        Two CS research assistants (one as a software engineer and one as a usability engineer) will perform a
         majority of the software and user interface development, overseen by Hart and managed by Stroila.

The computational PI’s (Hart, Stroila and Grosser) currently meet weekly, with the entire team (+ Jamison,
Goldwasser and Wagoner Johnson) currently meet as necessary, usually monthly. The following research
assistants and programmers will serve to allow this team to oversee their individual components of this project
while collaborating as a team through these assistants and programmers, who will interact with each other, the
software/usability engineers and Stroila daily.

        A research assistant in computational anatomy will derive and implement the details of anatomical
         accommodation, overseen by Goldwasser.

        A research assistant in computational mechanics will derive and implement the details of structural
         accommodation, overseen by Wagoner Johnson.

        A research assistant in bioengineering will derive and implement the details of material fabrication
         formats and constraints, overseen by Jamison.

        A research programmer at the University of Illinois Beckman Imaging Technology Group will work on
         the details of medical image processing, overseen by Grosser.

In addition, each of the Co-PI’s is budgeted for one month of summer support to provide them sufficient time to
commit to the collaboration and meeting the project’s objectives.

Year 3: Alpha System Prototype. We anticipate that at the end of the third year we will have an ―alpha‖
version of the proposed software system generated under the FDA validation principles for medical software
development such as continuous testing, version control, configuration standardization and bug tracking.

Year 4: Software and Usability Evaluation. In the fourth year, we plan to evaluate the alpha version of the
system, testing the program within a small set of clinical collaborators using real patient data. Such evaluations
provide valuable feedback into the iterative design of large software systems, but also require a large
investment in effort on measurement and establishing the statistical significance of design choices.

Year 5: Beta System Delivery. Given the insights from the fourth year’s evaluation we plan to focus our fifth
year on delivery of a revised beta version of the software, ready for FDA validation and clinical trial. In lieu of
an actual patient trial which would be premature at this stage, we plan to perform a test run on a cadaver skull,
creating one or more mandible injuries designed to explore the limits of the capabilities of our beta system.
Furthermore, our team surgeon will use the tool for an emergent (or past, if nothing is emergent during year 5)
pre-operative planning for an autograft surgery of the mandible. This will parallel previous proof-of-concept
work and allow direct comparisons between manual and automated methods for implant design.

Metrics to assess project success
At the end of five years, success of the project and return on NIH funding investment can be measured using
the following metrics:

1. Product delivery. The beta version of the automated design tool will be complete and functional on a
   standard personal computer. An NIH reviewer, for example, will be able to use it. A maxillofacial surgeon,
   unfamiliar with the development of the tool, will be able to use and critique it from practioner’s viewpoint;

PHS 398/2590 (Rev. 09/04)                             Page     44 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
2. Scientific Impact. Has the work been published in high impact journals in diverse fields of practice?
   Descriptions of the results of the project will appear in refereed journals in the fields of computer science,
   bioengineering and maxillofacial surgery;

3. Medical Impact. Has the project been disseminated to a large medical audience? Descriptions of the
   research will appear in the lay press, both traditional and electronic, to communicate the importance of the
   work in reducing the cost and improving the quality of treatment for patients who have experience bone
   loss due to trauma or disease.


This project presents a new and significantly different approach to the surgical treatment of patients who have
experienced bone loss through disease or trauma. This approach will dramatically improve the efficiency of
pre-operative planning, improve the quality and congruency of the defect-filling implant, eliminate the necessity
for bone harvesting surgery, and reduce the cost and morbidity of such procedures. Our diverse team has the
right expertise (materials science, biology, computer-aided design, computer-aided fabrication, and
maxillofacial surgery) and the right experience (two recently completed proofs-of-concept), to tackle this
complex problem. It is our shared recognition of the importance of the human dimension of our research that
has cemented our collaboration and motivates the current proposal.

1. Allen, B., Curless, B., and Popović, Z. 2003. The space of human body shapes: reconstruction and
   parameterization from range scans. ACM Trans. Graph. 22, 3 (Jul. 2003), 587-594.
2. J. F. Arvier , T. M. Barker, Y. Y. Yau, P. S. D'Urso, R. L. Atkinson, G. R. McDermant. 1994. Maxillofacial
   biomodelling, in British Journal of Oral and Maxillofacial Surgery (1994) 32: 276-83.
3. D.E. Breen and R.T. Whitaker, 2001, A Level-Set Approach for the Metamorphosis of Solid Models, IEEE
   Transactions on Visualization and Computer Graphics, Vol. 7, No. 2, pp. 173-192, April-June 2001.
4. Nathan A. Carr, John C. Hart. Painting Detail, Proc. SIGGRAPH 2004
5. Chang B-S, Lee C-K, Hong K-S, Youn H-J, Try H-S, Chung S-S, Park K-W. 2000. Osteoconduction at
   porous hydroxyapatite with various pore configurations. Biomaterials 2000;21:1291- 1298.
6. Davies, Rh.H, Twining, C.J., Cootes, T.F., Waterton, J. C., Taylor, C.J. 2002. A Minimum Description
   Length Approach to Statistical Shape Model. IEEE TMI 21 (2002)
7. Dean D, Min KJ, Bond A. 2003. Computer aided design of large-format prefabricated cranial plates. J
   Craniofac Surg. 2003 Nov;14(6):819-32.
8. Eggli PS, Muller W, Schenk RK. 1988. Porous hydroxyapatite and tricalcium phosphate cylinders
   with two different pore size ranges implanted in the cancellous bone of rabbits - A comparative
   histomorphometric and histologic study of boney ingrowth and implant substitution. Clinical Orthopaedics
   and Related Research 1988;232:127-138.
9. P. Golland, W.E.L. Grimson, M.E. Shenton, R. Kikinis. 2005. Detection and Analysis of Statistical
   Differences in Anatomical Shape. Medical Image Analysis, 9(1):69-86, 2005.
10. Gawande A. 2004. Casualties of war--military care for the wounded from Iraq and Afghanistan. N Engl J
   Med. 2004 Dec 9;351(24):2471-5.
11. Grosser, B., Sinn-Halon, J., Burton, C., Jamison, R., Goldwasser, M., Cesarano, J. 2004. Mandible
   Reconstruction Project (Animation). Siggraph Video Review, Issue 149, ACM, NY. August, 2004.
12. John C. Hart, Ed Bachta, Wojciech Jarosz, Terry Fleury. 2002. Using particles to sample and control
   more complex implicit surfaces. Proc. Shape Modeling International, May 2002.
13. Hollinger JO, Schmitz JP, Mizgala JW, Hassler C. 1989. An evaluation of two configurations of tricalcium
   phosphate for treating craniotomies. Journal of Biomedical Materials Research 1989;23:17-29.
14. Hollister SJ, Lin CY, Saito E, Lin CY, Schek RD, Taboas JM, Williams JM, Partee B, Flanagan CL,
   Diggs A, Wilke EN, Van Lenthe GH, Muller R, Wirtz T, Das S, Feinberg SE, Krebsbach PH. 2005.
   Engineering craniofacial scaffolds. Orthod Craniofac Res. 2005 Aug;8(3):162-73.
15. Hollister SJ, Maddox RD, Taboas JM. 2002. Optimal design and fabrication of scaffolds to mimic tissue
   properties and satisfy biological constraints. Biomaterials. 2002 Oct;23(20):4095-103.

PHS 398/2590 (Rev. 09/04)                             Page     45 Continuation Format
          Principal Investigator/Program Director (Last, First, Middle):   Hart, John C.
16. Michael Leventon, Eric Grimson, Olivier Faugeras. 2000. Statistical Shape Influence in Geodesic Active
   Contours, Comp. Vision and Patt. Recon. (CVPR), 2000.
17. Holmes R, Mooney V, Bucholz R, Tencer A. A. 1984. Coralline Hydroxyapatite Bone Graft Substitute:
   Preliminary Report. Clinical Orthopaedics and Related Research 1984;188:252-262.
18. Kotcheff, A.C.W., Taylor, C.J., 1998. Automatic Construction of Eigenshape Models by Direct
   Optimization. Med. Image Analysis 2 4 (1998) 303–314
19. Liu YL, Schoenars J, de Groot K, de Wijn JR, Schepers E. 2000. Bone healing in porous implants: a
   histological and histometrical comparative study on sheep. Journal of Materials Science: Materials in
   Medicine 2000;11:711-717.
20. Magan A, Ripamonti U. 1996. Geometry of Porous Hydroxyapatite Implants Influences Osteogenesis in
   Baboons (Papio ursinus). The Journal of Craniofacial Surgery 1996;2(7):71-78.
21. Rueckert D, Frangi AF, Schnabel JA. 2003. Automatic construction of 3-D statistical deformation models
   of the brain using nonrigid registration. IEEE Trans Med Imaging. 2003 Aug;22(8):1014-25.
22. Styner MA, Rajamani KT, Nolte LP, Zsemlye G, Szekely G, Taylor CJ, Davies RH. 2003. Evaluation of
   3D correspondence methods for model building. Inf Process Med Imaging. 2003 Jul;18:63-75.
23. Uchida A, Nade SM, McCartney ER, Ching W. 1984. The Use of Ceramics For Bone Replacement A
   Comparative Study of Three Different Porous Ceramics. Journal of Bone and Joint Surgery Br 1984;66:269-
24. van Blitterswijk CA, Grote JJ, Kuijpers W, Daems WT, de Groot K. 1986. Macropore tissue in-growth: a
   quantitative and qualitative study on hydroxyapatite ceramic. Biomaterials 1986;7:137-143.
25. Terry Yoo (editor), 2004, Insight into Images, A.K. Peters 2004, ISBN: 1-56881-217-5
26. Younger, E. and Chapman, M. 1989. Morbidity of bone graft donor sites, J. Orthop. Trauma, 1989, 3, 192-
27. Wen Yu Su and John C. Hart. 2005. A Programmable Particle System Framework for Shape Modeling.
   Proc. Shape Modeling International, June 2005.

PHS 398/2590 (Rev. 09/04)                             Page     46 Continuation Format

To top