Abstract.
March 30, 2003
Publications in Medical Imaging and Biomedical Engineering:
April 2-4, 2003
DIMACS Center, Rutgers University, Piscataway, NJ
Organizers:
Danny Chen, University of Notre Dame, dchen@cse.nd.edu
Jean-Claude Latombe, Stanford University, latombe@cs.stanford.edu
BibTeX references.
SUNY Fredonia
Abstract.
Eva K. Lee¹²³, Tim Fox² and Ian Crocker²
1: School of
Industrial and Systems Engineering, Georgia Institute of
Technology, Atlanta, Georgia.
2: Department
of Radiation Oncology, Emory University School of Medicine,
Atlanta, Georgia.
3: Corresponding author, evakylee@isye.gatech.edu
In this talk, we describe the use of mixed integer programming for simultaneously determining optimal beamlet fluence weights and beam angles in intensity-modulated-radiation-therapy treatment planning. In particular, binary variables are used to capture use/non-use of each beam, and continuous variables capture intensity values for each beamlet. For the tumor, explicit constraints include coverage with tumor underdose specified, conformity, and homogeneity; while DVH re-strictions for critical structures and normal tissues are imposed. Our algorithmic design thus fax has been motivated by clinical cases. To set up the model, 16-24 coplanar candidate fields of size 20x2O cm2 are generated, each consisting of 1600 5mm-beamlets. The maximum number of fields in the final plan will be constrained. Numerical tests are performed and analyzed on three tumor sites: 1)head-and-neck tonsil, 2)posterior-fossa-boost, and 3)prostate. In all cases, homogeneity is kept below 1.25, while requiring 95% tumor-volume coverage by the 100% isodose-curve and 100% tumor-coverage by the 95% isodose-curve. Optimal plans are obtained when maximum number of fields are restricted from 6-24 fields. Sensitivity of quality of plans versus various objective functions are compared and analyzed. For a practical 8-field practical plan, optimal results in the quality of plans for the head-and-neck tonsil case and the posterior fossa boost occur when the conformity is minimized in the objective function. In the prostate cancer case, there are overlapping boundary points in the rectal wall and the prostate, and the best quality plan (for 8-fields) results from an objective which minimizes the total weighted dose to the rectum and bladder, together with a tight ring of 5mm-thick normal tissue drawn around the prostate-PTV. The deviation in best objective choice use may result from the closeness of the rectal wall (and the bladder) to the prostate with overlapping boundary points.
Edgar Garduno
University of California, San Diego
Abstract.
P. M. Pardalos¹, J. C. Sackellares², D. S. Shiau², and V. A. Yatsenko²
1: Center for Applied Optimization
and ISE Department University of
Florida, Gainesville, FL
2: Departments of Neurology
and Neuroscience,
University of Florida, Gainesville, FL
The existence of complex chaotic, unstable, noisy and nonlinear dynamics in brain electrical activity requires new approaches to the study of brain dynamics. One approach is the combination of combining certain geometric concepts, control approach, and optimization. In this paper we discuss the use of a differential geometric approach to the control of Lyapunov exponents and the characterization of statistical information to predict and ``correct'' brain dynamics. This approach assumes that information about the physiological state (in this case the electrocencephalogram) comes in the form of a nonlinear time series with noise. The approach involves a geometric description of Lyapunov exponents for the purpose of correcting of the nonlinear process that provides adaptive dynamic control. We separate the Lyapunov exponents into tangent space (fiber bundle) and its functional space. Control involves signal processing, calculation of an information characteristic, measurement of Lyapunov exponents, and feed-back to the system. With more information, we can reduce uncertainty by a certain degree. We demonstrate the computational aspects of the proposed geometric approach on the base of different mathematical models in the presence of noise of various origins. We review the EEG signal, and outline a typical application of the geometrical representation: three dimensional reconstruction of Lyapunov exponents and correlation dimension obtained from EEG data. The novelty in this paper is in the representation of dynamical and information characteristics in three dimensional vector space in a way that permits practical applications. We discuss an application of this approach to the development novel devices for seizure control though electromagnetic feed-back.
School of Computer Science and Engineering,
The Hebrew University of Jerusalem, Israel
The past ten years have witnessed the development and deployment of a variety of computer-aided system for orthopaedic surgery. The goals of these systems are to enhance the surgeon's dexterity, visual feedback, and information integration before and during surgery. Computer-Aided Orthopaedic Surgery (CAOS) systems aim at improving the accuracy and steadiness of the surgical gestures, reduce complications due to tool and implant misplacement, reduce cumulative exposure to X-ray radiation, and reduce surgery time and patient morbidity. CAOS surgeries, such as joint, trauma and ligament surgeries, are beginning to have an impact on routine clinical practice.
In this talk, we survey the clinical results and technical state of the art in computer-aided orthopaedic surgery (CAOS). After a brief overview of the most common orthopaedic surgeries, we identify their common characteristics and the desirable system goals in terms of accuracy, resolution, robustness, repeatability, and computation and reaction time. We then describe the technical characteristics of existing imaging (X-ray, ultrasound, CT, and MRI), tracking (optical, electromagnetic, and gyroscopic), and robotic systems. We classify navigation systems into four categories: 1. CT-based; 2. Fluoroscopic X-ray based; 3. CT+Fluoroscopic X-ray based; and 4. CT-less spatial navigation, and robotic systems into two categories: 1. positioning and guiding aids, and; 2. active execution systems. For each, we describe the state of the art, its current clinical applications, and its pros and cons. We identify key technical challenges and propose possible solutions. We conclude with perspectives for the emerging field of medical CAD/CAM and its expected impact on orthopaedic surgery.
Neil Molino
Stanford University
Abstract.
Courant Institute, New York University
The modeling of biological structures and the model-based interpretation of medical images present many challenging problems. I will describe a powerful modeling paradigm, known as deformable models, which combines computational geometry, computational physics, and estimation theory. Deformable models evolve in response to simulated forces as dictated by the continuum mechanical principles of flexible materials expressed via variational principles and PDEs. The talk will review several biomedical applications currently under development, including image segmentation using dynamic finite element and topologically adaptive deformable models, as well as our recent work on ``deformable organisms.'' The latter aims to automate the segmentation process by augmenting deformable models with behavioral and cognitive control mechanisms from our work on artificial life.
Jinhui Xu, Guang Xu, Zhenming Chen and Kenneth R. Hoffmann
Abstract
CUNY Brooklyn
We present a mathematical analysis that we call Digital Morse Theory, that has applications to volume data set image visualization and understanding, with applications to Biomedical imaging. Volume data sets arise in many applications areas including CT data, MRI data, and X-ray crystallography. Isosurfaces and volume rendering are two important techniques for viewing the data and also for segmenting areas of interest in the data, for example for identifying anatomical features and tumors in medical image data.
Using our techniques we develop a method for preprocessing and organizing discrete scalar volume data of any dimension. The preprocessing algorithm constructs a criticality tree (independently developed) that is related to the contour tree of the literature, and computes regions of space called topological zones. The criticality tree and the zones hierarchically organize the data, and have a number of useful properties. We also present a simple algorithm that constructs provably correct isosurfaces in any dimension. We describe the implementation of a visual navigation system using our techniques. Typically researchers applying Morse theoretic ideas to scalar volume data have had to reorganize regularly gridded data onto a tetrahedral mesh and have had to perturb identical values. This is done to insure that the discrete data can be extended by a sufficiently well behaved interpolation function, so that a direct application of Morse theory could then be made.
In our approach, we first identify the properties satisfied by some of the more popular isosurface extraction methods. We then demonstrate that the class of interpolating functions that satisfy these properties produces topologically equivalent isosurfaces. We further show that the topological changes in these isosurfaces (changes to topological type, in the sense of homotopy), as the isovalue is varied, are uniquely defined and can be computed from combinatorial properties of the discrete data. Thus one does not need to consider the analytic properties of the interpolating function and one does not need to make a direct application of Morse theory. Consequently, our DMT analysis shows that our methods work on both irregularly and regularly gridded data, and on data with nonunique values, so that perturbation or resampling of the data set is not required.
We use our DMT analysis to define topological zones and we show how to efficiently compute the zones. We prove some interesting properties of the zones and show how these properties may be used in visualization. We discuss the use of the topological zone organization in fast isosurface construction. We show how the zone organization facillitates identifying all topologically distinct isosurface bounded objects in the dataset. We discuss our use of DMT in surface and data simplification, and managing the level of detail of a 3D rendering of a medical image.
Finally we will discuss future extensions of our DMT organizing technology. In particular we discuss our idea of segmentation surfaces, non-isovalued surfaces for anatomical segmentation for distinguishing features in regions of low image gradient. This provides a nice compromise betweem surface and volume-based techniques.
"Digital Morse Theory for scalar volume data," J. Cox, D.B.
Karron & N. Ferdous, July 2002.
Dept. of Applied Informatics, University of Szeged, Hungary
Consider a set of points in the space where all of the points emit the same intensity of radioactivity. Suppose that the space is filled with some homogeneous material having known absorption coefficient. There are detectors measuring the so-called absorbed projections, that is, the sums of the partially absorbed activities along lines between the point sources and the detectors. The problem is to reconstruct the point set from its absorbed projections. This model is similar to the acquisition technique used, for example, in the nuclear medicine, where the radioactivity emitted from the administered radiopharmacon is partially absorbed in the human body before detection.
This new direction of discrete tomography has a few interesting uniqueness results and reconstruction algorithms in special cases. For example, the reconstruction of discrete sets from two absorbed projections is studied if the absorption coefficient is log((1+sqrt(5))/2). A necessary and sufficient condition is given for the uniqueness and a reconstruction algorithm is suggested in this special case. It is proved that the reconstruction of so-called hv-convex discrete sets from such projections can be done in polynomial time. In this presentation we are going to generalize these results for different absorption values and for more than two projections.
D. B. Karron
Extracting precise and accurate objective measurements from image data has been an enduring problem in biomedical engineering. We are in the midst of an explosion of mega pixels from imaging technology. Imaging, from reflected, structured, and transmitted visible light, infrared, ionizing radiation, ultrasound, planar images, tomography, and the admixture of these, has vast application for in image guidance in many applications from image guided surgery to surgical simulation and education. The explosion of pixels from these technologies has motivated our efforts to understand these images in terms of well-defined objects. Further we wish to understand the interrelationship amongst these objects. We are meeting this challenge by the development by the extension of basic mathematical theory from Morse Theory into the pixilated digital domain wit Digital Morse Theory.
Digital Morse Theory is an extension of traditional Morse Theory developed by Marsdon Morse in the 1930's and 1940's at the Princeton Institute for Advanced Studies. The essential idea of Morse Theory is to describe the topology of a function as opposed to the geometry of a function in terms of Morse Criticalities. Traditionally, Morse theory required an analytic description of a function that was twice differentiable and continuous. Digital Morse Theory, as being developed by Cox and Karron with recent work by Jack Snoyink et al, attempts to adapt Morse?s concept of criticalities to discrete image data measurements. This has many practical applications in multidimensional image segmentation in that we can describe all of the objectively segmentable objects in terms of their parent, progeny, and peer criticalities in what we are terming a Morsian organization. A Reeb graph describes the relationship of objects to one another; we are programming a demonstration of the application of Digital Morse Theory for medical image segmentation that shows a catalog of all of the segmentable objects and permits the user to organize component objects into desired ensembles of anatomic objects quickly, precisely, and accurately.
Sample medical images from the CT and MRI images of the Visible Human Project and their associated DMT graph will be presented. We will demonstrate the relationship between the DMT graph and image components and how we can build a 3d scene of segmented objects with important projects for anatomic surgical modeling. Specifically, we desire anatomic dimensional fidelity and accuracy, as we will use these models to make critical care decision. Since models derived from DMT based segmentation have the Jordan and Manifold surfaces, this means that disparate segmented objects do not have interpenetrating surfaces or other geometric pathologies.
Our goal is to have semi-automatic to fully automatic segmentation and fusion of disparate medical images.
"Digital Morse Theory for scalar volume data," J. Cox, D.B.
Karron & N. Ferdous, July 2002.
EPIDAURE Research Project, INRIA Sophia Antipolis, France
I will present the latest developments of a ``beating heart model'', developed at INRIA, which is suitable for the quantitative analysis of cardiac images (gated MRI, echocardiography,..). This model describes three distinct physical phenomena: the propagation of the electrical wave from the apex to the base, the application of contractile stress depending on the action potential level and the mechanical deformation of the ventricles. More information regarding the three models axe provided below.
The current implementation is based on a generic finite element mesh consisting of linear tetrahedra. Our ultimate goal is to adapt this generic model to the anatomy and physiology of a given patient and to recover key physical parameters (stress, aortic pressure, ....) that cannot be measured directly from the images. I will illustrate the presentation with results concerning the generic model and its adaptation to medical images. Readers may refer to [1, 4, 3) for more details.
There exists several mathematical model of the propagation of electric wave across the heart ventricles. We have chosen to implement the macroscopic model of Fitzhugh Nagumo [2] which links the evolution of the transmembrane potential (directly related to the alpha parameter) and an auxiliary variable z.
We solve this paxtial differential equation using a standard P1 Lagrange finite element procedure (with mass lumping and first order numerical integration at vertices), on the given tetrahedral anatomical mesh. The Euler explicit time integration is performed to advance computations. We have introduced an anisotropic behavior by making the diffusion matrix D depending on the local fiber orientation.
One residual and crucial problem (of great issues) is that some boundary conditions must be imposed at the junctions between the special conduction system (Purkinje fibers) and the myocardium, which is not well known so far. We have followed an approach which consists in assuming that the junctions region is located near the apex, below a plane that cuts the main heart axis.
During one heart beat cycle, depolaxization wave propagates along the my-ocardium during about 10% of the total cardiac cycle. It induces a contraction of the whole myocardium, which produces a stress tensor sigma_f = alpha_f × f, where alpha is the activation rate, and f is the fiber direction. Therefore, when a fiber is activated, its contraction is modeled as a pressure applied to the surface of the tetrahedron in the fiber direction.
Furthermore, in addition to the stress along the fiber direction, the electrical activation of the myocaxdium modifies its elastic properties.
The heart myocaxdium is a nonlineax viscoelastic anisotropic material. It is composed of fiber bundles spiraling around the two ventricles. Obviously, the physical model has to be simple enough for computational purposes. To simplify the model, we propose to approximate the non-linear behavior of the myocardium by a series of linear models that are only valid during a small part of the cardiac cycle.
University of Texas at Austin
University of Pennsylvania
Image segmentation - the process of defining objects in images - remains the most challenging problem in image processing despite decades of research. Many general methodologies have been proposed to date to tackle this problem. An emerging framework that has shown considerable promise recently is that of fuzzy connectedness. Images are by nature fuzzy. Object regions manifest themselves in images with a heterogeneity of image intensities owing to the inherent object material heterogeneity, and artifacts such as blurring, noise and background variation introduced by the imaging device. In spite of this gradation of intensities, knowledgeable observers can perceive object regions as a gestalt. The fuzzy connectedness framework aims at capturing this notion via a fuzzy topological notion called fuzzy connectedness which defines how the image elements hang together spatially in spite of their gradation of intensities. In defining objects in a given image, the strength of connectedness between every pair of image elements is considered, which in turn is determined by considering all possible connecting paths between the pair. In spite of a high combinatorial complexity, theoretical advances in fuzzy connectedness have made it possible to delineate objects via dynamic programming at close to interactive speeds on modern PCs. This paper gives a tutorial review of the fuzzy connectedness framework delineating the various advances that have been made. These are illustrated with several medical applications in the areas of Multiple Sclerosis of the brain, MR and CT angiography, brain tumor, mammography, upper airway disorders in children, and colonography.
Danny Z. Chen, Xiaobo S.
Hu, Shuang Luan and Chao Wang,
Department of Computer Science and
Engineering, University of Notre Dame
Charles E. Lee, Shahid A. Naqvi and Cedric X. Yu,
Department of Radiation Oncology, University of Maryland Medical School
Xiaodong Wu,
Department of Computer Science, The University of Texas - Pan American
Radiation therapy uses radiation beams to eradicate tumors while sparing surrounding vital organs and healthy tissues. A good shape matching between the tumor volume and radiation beams is critical to the quality of treatments. In general, tumors have various sizes and shapes, while the radiation beams generated by the linear accelerators (LINAC) (machines that are used to generate and deliver radiation) are cylindrical. One way to overcome the poor shape matching problem with LINAC is to use a device called multileaf collimator (MLC), which uses a set of opposing tungsten leaves to vary the cylindrical beam shape in order to obtain the needed discrete dose distributions. However, one of the challenging issues of using a collimator is its potentially prolonged treatment time. The static leaf sequencing (SLS) problem axises in radiation therapy for cancer treatments, aiming to accomplish the delivery of a radiation prescription to a taxget tumor using an MLC in the minimum amount of delivery time.
Geometrically, the SLS problem can be formulated as a 3-dimensional partition problem for which the 2-dimensional problem of partitioning a polygonal domain (possibly with holes) into a minimum set of monotone polygons is a special case. In this talk, we present new geometric algorithms for a basic case of the 3-D SLS problem (which is also of clinical value) and for the general 3-D SLS problem. Our basic 3-D SLS algorithm, based on new geometric observations, produces guaranteed optimal quality solutions using Steiner points in polynomial time; the previously best known basic 3-D SLS algorithm gives optimal outputs only for the case without using Steiner points, and its time bound involves a multiplicative factor of a factorial function of the input. Our general 3-D SLS algorithm is based on our basic 3-D SLS algorithm and a polynomial time algorithm for partitioning a polygonal domain (possibly with holes) into a minimum set of x-monotone polygons, and has a fast running time.
Experiments and comparisons using real medical data and on a real radiotherapy machine have shown that our 3-D SLS algorithms and software produce treatment plans that use significantly shorter delivery time and give better treatment quality than the current most popular commercial treatment planning system and the most well-known SLS algorithm. Some of our techniques and geometric procedures (e.g., for the problem of partitioning a polygonal domain into a minimum set of x-monotone polygons) axe interesting in their own right.
University of North Carolina at Chapel Hill
Modeling deformation is a key component for many technology assisted medical procedures, such as image-guided surgery, multi-modal image registration, surgical planning and training, instructional medical illustration. In this talk, we discuss geometric issues involved in simulating interaction between flexible bodies and present efficient algorithms for a subset of these problems.
Some of the challenging research issues include rapid detection of contacts, efficient enforcement of non-penetration constraints, and fast contact resolution. We describe two classes of geometric algorithms for computing "deformed distance fields" dynamically using fast-marching level-set methods and modern graphics hardware. We introduce the concept of "multi-level simulation acceleration" techniques for enhancing the performance of simulators used in modeling deformation. We demonstrate the effectiveness of our approaches on moderately complex animated scenes and medical simulations.
Rutgers University
Ultrasound imaging is widely used in medicine because it is safe and interactive. However, raw ultrasound images have significant noise and artifacts relative to other imaging techniques such as CT and MRI, and present interesting algorithmic challenges for extracting useful geometric information. In this talk I will describe the fundamentals of ultrasound imaging, and some of our recent work in extracting 3D geometry from free-hand ultrasound images and for improving ultrasound image quality. (Joint work with Rohling, Salcudean, Wei, and Zhang).
Cedric X. Yu
University of Maryland School of Medicine
In the talk, I will review the current topics and discuss what we are doing at the University of Maryland School of Medicine. These include IMRT optimization, brachytherapy optimization, deformable image registration using an optimization setting and a clinical trial related to it, and IGS/CAS/CAD (computer aided diagnosis). I will introduce the problems and the current status. I will also talk about dose calculation, which has been always a compromise between accuracy and computational speed. There could also be an algorithmic solution for the dose calculation problem. In the second part of my talk, I would like to touch some new areas that are not on our current screen. These include genomics and proteomics, biological modeling, and simulation of radiation induced DNA damage and repair. I will illustrate that the new possibilities are much broader than in the traditional areas.
Freie Universitat Berlin
Gel-electrophoresis is a technique to produce from a probe of human or animal tissue a two-dimensional diagram consisting of numerous spots each of which corresponds to a protein present in the tissue. We developed techniques based on computational geometry to compare several diagrams of this kind and to match corresponding spot patterns. In fact, our technique is based on constructing the Delaunay triangulation of the spots and compare Delaunay edges of approximately the same length and orientation.
In a second project we developed techniques used in brain surgery. Here, before surgery metal markers are fixed to the patient's head. The brain together with the markers is scanned by CT or MR and these data are stored. During surgery the surgical instrument is equipped with a radio transmitter and from its position among the current position of the markers our program, by calculating the correct matching (rigid motion) between the markers and their images is able to show on screen the current location of the instrument within the brain.
SUNY Fredonia
Abstract.
University of Wisconsin
In many cases a radiotherapy treatment is delivered as a series of smaller dosages over a period of time. Currently, it is difficult to determine the actual dose that has been delivered at each stage, precluding the use of adaptive treatment plans. However, new generations of machines will give more accurate information of actual dose delivered, allowing a planner to compensate for errors in delivery. We formulate a model of the day-to-day planning problem as a stochastic linear program and exhibit the gains that can be achieved by incorporating uncertainty about errors during treatment into the planning process. Due to size and time restrictions, the model becomes intractable for realistic instances. We show how neuro-dynamic programming can be used to approximate the stochastic solution, and derive results from our models for realistic time periods. These results allow us to generate practical rules of thumb that can be immediately implemented in current planning technologies.
M. C. Ferris and M. M. Voelker. Neuro-dynamic programming for
radiation treatment planning. Numerical Analysis Group Research Report
NA-02/06, Oxford University Computing Laboratory, Oxford University,
2002.
Achim Schweikard
Informatik, Universtat Luebeck, Germany
Web links:
Stereotactic radiosurgery uses focused beams of radiation from multiple spatial directions to ablate brain tumors. To plan the treatment, spatial directions of the beams must be determined in a first step. The set of such beam directions should be chosen such that the dose to surrounding healthy tissue remains minimal, and the tumor absorbs a homogeneous dose throughout its volume. In addition to choosing the beam directions, one must find the weight for each such beam direction, i.e., the activation duration for the beam.
Earlier systems for stereotactic radiosurgery are inherently limited to treat lesions in the brain. The first goal of robotic radiosurgery is to improve conventional stereotactic radiosurgery with respect to accuracy. A second, more important goal is to allow for treating lesions anywhere in the body.
In the first part of this talk, we present planning methods for robotic radiosurgery. Our methods are based on linear programming. Linear programming gives a fast and globally convergent optimisation strategy. It has been proposed as a planning technique in radiation therapy applications by several authors. In robotic radiosurgery, up to 1500 beam directions are used for a single treatment, and the beam can be moved in space with full kinematic flexibility. We show that the general approach of linear programming is practical for robotic radiosurgery, where a very large number of beam directions is used. An implementation of our methods has been incorporated into the Cyberknife robotic radiosurgery system. Several thousand patients with tumors of the brain, the spine and the lung have been treated with the described linear programming based planning system.
The second part of this talk addresses the problem of navigation for robotic radiosurgery. Conventional stereotactic radiosurgery is limited to the brain. Tumors in the chest and the abdomen move during respiration. The ability of conventional radiation therapy systems to compensate for respiratory motion by moving the radiation source is inherently limited. Since safety margins currently used in radiation therapy increase the radiation dose by a very large amount, an accurate tracking method for following the motion of the tumor is of utmost clinical relevance. We investigate methods to compensate for respiratory motion using robotic radiosurgery. Thus, the therapeutic beam is moved by a robotic arm, and follows the moving target tumor. To determine the precise position of the moving target we combine infrared tracking with synchronized X-ray imaging. Infrared emitters are used to record the motion of the patient's skin surface. A stereo X-ray imaging system provides information about the location of internal markers. During an initialisation phase (prior to treatment), the correlation between the motions observed by the two sensors (X-ray imaging and infrared tracking) is computed. This model is also continuously updated during treatment to compensate for other, non-respiratory motion. Experiments and clinical trials suggest that robot-based methods can substantially reduce the safety margins currently needed in radiation therapy. Our correlation based navigation method has since been incorporated into the Cyberknife robotic radiosurgery system. Our new module is in routine clinical use at a growing number of sites worldwide.
References:
Schweikard, A., Glosser, G., Bodduluri, M., Adler, J. R., Robotic
Motion Compensation for Respiratory Motion during Radiosurgery, Journal
of Computer-Aided Surgery, vol. 5, no 4., 263-277, Sept. 2000.
Schweikard, A., Bodduluri, M., Adler, J. R.: Planning for
Camera-Guided Radiosurgery, IEEE Transactions on Robotics and
Automation, 13, 6, 951-962, 1998.
Marcelo Siqueira¹, Tessa Sundaram², Suneeta Ramaswami³, Jean Gallier¹, James Gee^4
1: Department of Computer and Information Science, University of
Pennsylvania
2: School of Medicine and Department of Bioengineering, University of
Pennsylvania
3: Department of Computer Science, Rutgers University
4: Department of Radiology, University of Pennsylvania
Abstract
Rutgers University
3D segmentation and modeling for improved medical diagnosis of disease promises to revolutionize the way medical practice is done currently which is inherently 2D. In this talk we will first present our 3D hybrid segmentation framework which consists of integrating low level image processing and higher level deformable models. We will show semgentations of a variety of organs including brains, the heart and the colon. In the second part we will present our 3D stress-strain estimation framework of the heart's shape, motion and material properties from MRI-SPAMM.
A. Frank van der Stappen, Han-Wen Nienhuys
Universiteit Utrecht, The Netherlands
Abstract.
Brown University
Abstract.
Computer Science Department, Stanford University
We present several algorithms to deform objects realistically, and interact with these deformable objects, in the context of a general purpose surgery simulator. Our system renders a 3-D graphical virtual environment, which consists of an arbitrary number of tools, such as forceps and laparoscopes, and anatomical structures, such as organs and vessels. These tools are controlled by a user of the system via a variety of tracking devices, and used to manipulate the structures (e.g., grasp, move, cut and suture). A mass-spring model is used to describe 3-D soft tissues, and external forces (usually exerted by the tools) drive the deformation of these tissues. An assumption of quasi-static behavior increases the efficiency of the deformation algorithm, and a method of localizing the defor-mation allows us to scale up to large objects. Our simulation also introduces a novel method of suture motion, which looks natural, and obeys the physical constraints caused by both self-collisions in the suture, and collisions between the suture and other objects. We are able to tie knots in the suture, and identify them. Interactions between the tools, suture, soft tissues, and other objects rely heavily on collision detection and collision management. Focusing our processing power on certain of these interactions allows for the creation of specialized applications (e.g., a microsurgery application involves forceps, blood vessels, and a suture, each interacting with the others). A description of the soft-tissue simulation and collision detection can be found in [1, 2], a more detailed description of the system architecture is discussed in [3], while much of the suture simulation is unpublished.
Simon K. Warfield, Kelly H. Zou and William M. Wells
Harvard University
The extraction of geometry from medical images depends upon accurate image segmentation to appropriately identify the structures of interest. Characterizing the performance of image segmentation approaches has been a persistent challenge. Performance analysis is important since segmentation algorithms often have limited accuracy and precision.
Interactive drawing of the desired segmentation by domain experts has often been the only acceptable approach, and yet suffers from intra-expert and inter-expert variability. Automated algorithms have been sought in order to remove the variability introduced by experts, but automated algorithms must be assessed to ensure they are suitable for the task.
The accuracy of segmentations of medical images has been difficult to quantify in the absence of a known true segmentation for clinical data. Although physical and digital phantoms can be constructed for which ``ground truth'' is known or readily estimated, such phantoms don't fully reflect clinical images in that it is difficult to construct phantoms which reproduce the full range of imaging characteristics and normal and pathological anatomical variability observed in clinical data. Comparison to a collection of segmentations by experts is an attractive alternative since it can be carried out directly on the clinical imaging data for which the validation of the segmentation is desired, but the most appropriate measure or measures with which to compare such segmentations has not been clarified and several measures are used in practice.
We present here an Expectation-Maximization algorithm we call STAPLE (Simultaneous Truth and Performance Level Estimation). The algorithm takes a collection of segmentations and computes a probabilistic estimate of the true segmentation and a measure of the performance level represented by each segmentation. The source of each segmentation in the collection may be an appropriately trained human expert or experts, or it may be an automated segmentation algorithm. STAPLE is straightforward to apply to clinical imaging data, it readily enables the validation of an automated image segmentation algorithm, and allows direct comparison of expert and algorithm performance.
Illustrative results are presented on digital phantoms for which the true segmentation is known, and for several clinical applications that require validation of image segmentations.
Page created & maintained by Frederic Leymarie, 2003.
Comments, suggestions, etc., mail to: leymarie@lems.brown.edu