## GIST: An interactive GPU-based level-set segmentation tool for 3d medical images (2004)

### Cached

### Download Links

Venue: | JOURNAL ON MEDICAL IMAGE ANALYSIS |

Citations: | 13 - 3 self |

### BibTeX

@ARTICLE{Cates04gist:an,

author = {Joshua E. Cates and Aaron E. Lefohn and Ross T. Whitaker},

title = { GIST: An interactive GPU-based level-set segmentation tool for 3d medical images},

journal = {JOURNAL ON MEDICAL IMAGE ANALYSIS},

year = {2004},

volume = {8},

pages = {217--231}

}

### OpenURL

### Abstract

While level sets have demonstrated a great potential for 3D medical image segmentation, their usefulness has been limited by two problems. First, 3D level sets are relatively slow to compute. Second, their formulation usually entails several free parameters which can be very difficult to correctly tune for specific applications. The second problem is compounded by the first. This paper describes a new tool for 3D segmentation that addresses these problems by computing level-set surface models at interactive rates. This tool employs two important, novel technologies. First is the mapping of a 3D level-set solver onto a commodity graphics card (GPU). This mapping relies on a novel mechanism for GPU memory management. The interactive rates level-set PDE solver give the user immediate feedback on the parameter settings, and thus users can tune free parameters and control the shape of the model in real time. The second technology is the use of region-based speed functions, which allow a user to quickly and intuitively specify the behavior of the deformable model. We have found that the combination of these interactive tools enables users to produce good, reliable segmentations. To support this observation, this paper presents qualitative results from several different datasets as well as a quantitative evaluation from a study of brain tumor segmentations.

### Citations

3080 | Snakes: Active contour models
- Kass, Witkin, et al.
- 1988
(Show Context)
Citation Context ...e underlying parameterization. The level-set method is a general framework that must be tuned to specific applications. As with the original work on image segmentation by parametric deformable models =-=[8]-=-, the level-set approach to segmentation typically combines a data-fitting term with a smoothing term. However, there are alternatives. For instance, Whitaker [9] proposes a formulation that mimics pa... |

1078 | M.: Fast geodesic active contours
- Goldenberg, Kimmel, et al.
- 2001
(Show Context)
Citation Context ...hat slows at image edges. Because of the monotonic expansion/contraction, convergence to local minima is less of a problem, but the results tend to be biased either inward or outward. Caselles et al. =-=[11]-=- propose an alternative that minimizes an edge-weighted area metric. In that case the data term is weighted more heavily as the model approaches its target. These methods (and many others) focus on im... |

671 |
Level Set Method and Dynamic Implicit Surfaces
- Osher, Fedkiw
- 2002
(Show Context)
Citation Context ...n a volume. The use of level sets has been widely documented in the medical imaging literature, and several works give more comprehensive reviews of the method and the associated numerical techniques =-=[5, 6]-=-. For certain classes of applications level sets have several advantages over parametric models. Because they are implicit, level sets can change topology. This means that during a deformation a user ... |

482 | Comparing images using the Hausdorff distance
- Huttenlocher, Rucklidge
(Show Context)
Citation Context ... usually given through some measure of a region’s overlap and its distance from the ground truth. Common distancesUnder review at Medical Image Analysis, 2004 7 metrics include the Hausdorff distanc=-=e [46]-=- and the root mean squared distance between selected boundary points [39, 40]. Often overlap is characterized by a similarity measure between experimental and ground truth volumes. One common similari... |

466 |
Accelerated volume rendering and tomographic reconstruction using texture mapping hardware
- Cabral, Cam, et al.
- 1994
(Show Context)
Citation Context ...nsional transfer functions to directly render spectral data. For rendering the evolving level-set model, we use a modification of the conventional 2D sliced approach to texture-based volume rendering =-=[51]-=-. The modification to the conventional approach is the rendering of the level-set solution directly from the packed tiles, which are stored as a single 2D texture. The level-set data and tile configur... |

428 | Level Set Methods and Fast Marching Methods: Evolving
- SETHIAN
- 1996
(Show Context)
Citation Context ...n a volume. The use of level sets has been widely documented in the medical imaging literature, and several works give more comprehensive reviews of the method and the associated numerical techniques =-=[5, 6]-=-. For certain classes of applications level sets have several advantages over parametric models. Because they are implicit, level sets can change topology. This means that during a deformation a user ... |

237 | Linear Algebra Operators for GPU Implementation of Numerical Algorithms
- Kruger, Westermann
- 2003
(Show Context)
Citation Context ...e level set moves with each time step, and thus the packed representation must readily adapt to the changing position of the model. This requirement is in contrast to the recent sparse matrix solvers =-=[31, 32]-=- and previous work on rendering with compressed data [33, 34]. In the two sparse-matrix solvers[31, 32], a packed texture scheme is used to efficiently compute sparse matrix-vector multiplications as ... |

236 |
Pattern Recognition Principles
- Tou, Gonzalez
- 1974
(Show Context)
Citation Context ...s the accuracy of the expert segmentations. A second, less conservative measurement that produces a more unbiased estimate of the manual segmentation accuracy is a round-robin leave-one-out strategy, =-=[56], -=-where p, q, and c values for each Di j are computed using Wk generated by all segmentations k �= j. Accuracy metrics must be interpreted carefully. Note that where a segmentation technique shows hig... |

218 | Sparse matrix solvers on the gpu: conjugate gradients and multigrid
- Bolz, Farmer, et al.
(Show Context)
Citation Context ...e level set moves with each time step, and thus the packed representation must readily adapt to the changing position of the model. This requirement is in contrast to the recent sparse matrix solvers =-=[31, 32]-=- and previous work on rendering with compressed data [33, 34]. In the two sparse-matrix solvers[31, 32], a packed texture scheme is used to efficiently compute sparse matrix-vector multiplications as ... |

163 | Multidimensional transfer functions for interactive volume rendering
- Kniss, Kindlmann, et al.
- 2002
(Show Context)
Citation Context ...ginal volume, the input data and its gradient vectors are kept on the GPU as 3D textures. This GPU-based volume rendering incorporates multidimensional transfer functions as described in Kniss et al. =-=[50]-=-. The current implementation of GIST renders only scalar volume data, and thus for spectral data it renders only a derived scalar quantity (e.g. one component or magnitude). Future work will include t... |

150 | A Level-Set Approach to 3DReconstruction From Range Data
- Whitaker
- 1998
(Show Context)
Citation Context ...Sethian [22] have proposed the narrow band method, which updates the embedding on a band of 10-20 pixels around the model, and reinitializes that band whenever the model approaches the edge. Whitaker =-=[23]-=- proposed the sparse-field method, which introduces a scheme in which updates are calculated only on the wavefront, and several layers around that wavefront are updated via a distance transform at eac... |

143 | P.: Automated model-based tissue classification of mr images of the brain
- Leemput, Maes, et al.
- 1999
(Show Context)
Citation Context ...h volumes. One common similarity measure is the cardinality of the intersection (in pixels or voxels) of positive classifications in two volumes volumes over the union of the positive classifications =-=[41, 47]-=-, denoted s. Another overlap metric is the total correct fraction, c, which is simply the percentage of correctly classified pixels in the image volume (negative and positive) [1]. Another strategy fo... |

108 |
M.: Geometrically deformed models: A method for extracting closed geometric models from volume data
- Miller, Breen, et al.
- 1991
(Show Context)
Citation Context ... level sets do not require reparameterization as they deform far from their initial conditions—e.g. deformable meshes typically require the insertion or deletion of triangles under such circumstance=-=s [7]-=-. Finally, level sets allow for geometric surface deformations, which means that the results of a deformation process depend on the shape of the surface and the input data and not on some underlying p... |

100 | Fast matrix multiplies using graphics hardware
- Larsen, McAllister
(Show Context)
Citation Context ...el has been used by a number of researchers to map a wide variety of computationally demanding problems to GPUs. Examples include matrix multiplication, finite element methods, and multi-grid solvers =-=[25, 26, 27]-=-. All of these examples demonstrate a homogeneous sequence of operations over a densely populated grid structure. Strzodka et al. [28] were the first to show that the level-set equations could be solv... |

90 | A multigrid solver for boundary value problems using programmable graphics hardware
- Goodnight, Woolley, et al.
- 2003
(Show Context)
Citation Context ...el has been used by a number of researchers to map a wide variety of computationally demanding problems to GPUs. Examples include matrix multiplication, finite element methods, and multi-grid solvers =-=[25, 26, 27]-=-. All of these examples demonstrate a homogeneous sequence of operations over a densely populated grid structure. Strzodka et al. [28] were the first to show that the level-set equations could be solv... |

83 |
Morphometric analysis of white matter lesions in MR images: method and validation
- Zijdenbos, Dawant, et al.
- 1994
(Show Context)
Citation Context ... a user-assisted segmentation technique, which requires a user-based evaluation to capture variations in the individual decision-making process. Experimental trials across a number of users and images=-=[41, 42]-=- can generate data appropriate for statistical analysis that account for user variability. A combination of different factors determines the effectiveness of a segmentation. For instance Udupa et. al[... |

81 | A level-set algorithm for minimizing the Mumford–Shah functional in image processing - Chan, Vese |

70 |
A methodology for evaluation of boundary detection algorithms on medical images
- Chalana, Kim
- 1997
(Show Context)
Citation Context ...e from the ground truth. Common distancesUnder review at Medical Image Analysis, 2004 7 metrics include the Hausdorff distance [46] and the root mean squared distance between selected boundary points =-=[39, 40]-=-. Often overlap is characterized by a similarity measure between experimental and ground truth volumes. One common similarity measure is the cardinality of the intersection (in pixels or voxels) of po... |

70 | Valmet: A new validation tool for assessing and improving 3d object segmentation
- Gerig, Jomier, et al.
- 1994
(Show Context)
Citation Context ...e from the ground truth. Common distancesUnder review at Medical Image Analysis, 2004 7 metrics include the Hausdorff distance [46] and the root mean squared distance between selected boundary points =-=[39, 40]-=-. Often overlap is characterized by a similarity measure between experimental and ground truth volumes. One common similarity measure is the cardinality of the intersection (in pixels or voxels) of po... |

69 |
A.: Deformable meshes with automated topology changes for coarse-to-fine 3D surface extraction. Medical Image Analysis 3
- Lachaud, Montanvert
- 1999
(Show Context)
Citation Context ...rmation a user need not worry about surfaces colliding or pinching off. Deformable and topologically adaptable meshes typically require the insertion or deletion of triangles under such circumstances =-=[7, 8, 9]-=-. Finally, level sets allow for geometric surface deformations, which means that the results of a deformation process depend on the shape of the surface and the input data and not on some underlying p... |

68 |
Statistical region snake-based segmentation adapted to different physical noise models
- Chesnaud, Refregier, et al.
- 1999
(Show Context)
Citation Context ...e documents several other strategies for fitting level sets to image data. For instance, several authors have propose using the statistics of the greyscale interior of the model to control the motion =-=[12, 13]. -=-Alternatively, the motion of the level set can depend on a variational formulation that positions the interface to create discontinuities that best model the discontinuities in the input data [14–16... |

54 | Level set segmentation in graphics hardware
- Rumpf, Strzodka
- 2001
(Show Context)
Citation Context ...multiplication, finite element methods, and multi-grid solvers [26–28]. All of these examples demonstrate a homogeneous sequence of operations over a densely populated grid structure. Strzodka et al=-=. [29]-=- were the first to show that the level-set equations could be solved using a graphics processor. Their solver implements the two-dimensional level-set method using a time-invariant speed function for ... |

49 |
Volumetric deformable models: Active blobs
- Whitaker
- 1994
(Show Context)
Citation Context ...ation by parametric deformable models [8], the level-set approach to segmentation typically combines a data-fitting term with a smoothing term. However, there are alternatives. For instance, Whitaker =-=[9]-=- proposes a formulation that mimics parametric deformable models, in which level surfaces move toward edges (high gradient magnitude) in volumes. In that formulation the model must be within a somewha... |

49 | N.: Rendering from compressed textures
- Beers, Agrawala, et al.
- 1996
(Show Context)
Citation Context ...presentation must readily adapt to the changing position of the model. This requirement is in contrast to the recent sparse matrix solvers [31, 32] and previous work on rendering with compressed data =-=[33, 34]-=-. In the two sparse-matrix solvers[31, 32], a packed texture scheme is used to efficiently compute sparse matrix-vector multiplications as well as compute values of the sparse matrix elements on the G... |

45 |
A pde based fast local level set method
- Peng, Merriman, et al.
- 1999
(Show Context)
Citation Context ...d method, which introduces a scheme in which updates are calculated only on the wavefront, and several layers around that wavefront are updated via a distance transform at each iteration. Peng et al. =-=[24]-=- present a similar local method. Even with this very narrow band of computation, update rates using conventional processors on typical medical data sets (e.g. 256 3 voxels) are not interactive. This i... |

39 |
T.: Adaptive texture maps
- Kraus, Ertl
- 2002
(Show Context)
Citation Context ...presentation must readily adapt to the changing position of the model. This requirement is in contrast to the recent sparse matrix solvers [31, 32] and previous work on rendering with compressed data =-=[33, 34]-=-. In the two sparse-matrix solvers[31, 32], a packed texture scheme is used to efficiently compute sparse matrix-vector multiplications as well as compute values of the sparse matrix elements on the G... |

34 | S.: Fast volume segmentation with simultaneous visualization using programmable graphics hardware
- Sherbondy, Houston, et al.
- 2003
(Show Context)
Citation Context ...neralization of [29], which includes a second-order curvature computation, and a significantly improved GPU solver that implements a narrow-band strategy. Also related is the work of Sherbondy et al. =-=[30]-=-, in which they identify regions of interest to solve a diffusion equation for volume segmentation. This paper describes a GPU computational model that supports time-dependent, sparse grid problems. T... |

33 | R.: A streaming narrow-band algorithm: Interactive deformation and visualization of level sets
- Lefohn, Kniss, et al.
- 2004
(Show Context)
Citation Context ...ST, an interactive level-set segmentation tool, and the GPU implementation that makes it possible. It begins with a brief review of the GPU-based level-set solver (for a more complete description see =-=[49]-=-), describes the visualization of the volume data and surface models, and then describes the user interface to GIST. 4.1 GPU Level-Set Solver The efficient solution of the level-set PDEs relies on onl... |

30 | Geometric Surface Processing via Normal Maps
- Tasdizen, Whitaker, et al.
- 2003
(Show Context)
Citation Context ...lood vessels; this is equivalent to a space-curvature shortening for very thin objects. Recently, Tasdizen et al. propose the diffusion of normals in order to approximate higher-order geometric flows =-=[18, 19]-=-. Solving level-set PDEs on a volume requires proper numerical schemes [20] and entails a significant computational burden. Stability requires that the surface can progress at most a distance of one v... |

29 | W.: Validation of image segmentation and expert quality with an expectation-maximization algorithm
- Warfield, Zou, et al.
- 2002
(Show Context)
Citation Context ...Averaging of multiple nonparametric shapes, however, is itself a difficult problem. One technique for combining multiple segmentations is Simultaneous Truth and Performance Level Estimation (STAPLE), =-=[45]-=-. This treats segmentation as a pixelwise classification, which leads to an averaging scheme that accounts for systematic biases in the behavior of experts The accuracy of an individual experimental s... |

28 |
F.: Applying watershed algorithms to the segmentation of clustered nuclei. Cytometry 28
- Malpica, Solórzano, et al.
- 1997
(Show Context)
Citation Context ...gmentation algorithm. There are two strategies for evaluating segmentation algorithms. One strategy is to study segmentation performance in the context of a particular clinical or scientific question =-=[35, 36]-=-. For instance, the effectiveness of the algorithm within a study that monitors the volumes or sizes of tumors. The second approach is to study to evaluate segmentation in the absence of a specific cl... |

28 |
Geometry-Limited Diffusion
- Whitaker
- 1993
(Show Context)
Citation Context ...ity and specificity, researchers have proposed using receiver operator characterizations (ROC), which monitor the behavior of this tradeoff for different segmentation algorithms or parameter settings =-=[48, 38]-=-. The precision of a segmentation method is an indicator of how repeatable the results are using that technique. Alternatively, precision is an indicator or the degree of randomness inherent to the me... |

26 | R.T.: Interactive, gpu-based level sets for 3d segmentation
- Lefohn, Cates, et al.
- 2003
(Show Context)
Citation Context ...spectral) data. – Quantitative and qualitative evidence that this interactive level-set approach is effective for brain tumor segmentation. The remainder of the paper, which is an extended version o=-=f [3]-=-, is organized as follows. The next section gives some technical background and related work on level sets, GPUs, and segmentation evaluation methods. Section 3 describes the formulation of the level-... |

25 | M.: Using graphics cards for quantized FEM computations
- Strzodka, Rumpf
- 2001
(Show Context)
Citation Context ...el has been used by a number of researchers to map a wide variety of computationally demanding problems to GPUs. Examples include matrix multiplication, finite element methods, and multi-grid solvers =-=[25, 26, 27]-=-. All of these examples demonstrate a homogeneous sequence of operations over a densely populated grid structure. Strzodka et al. [28] were the first to show that the level-set equations could be solv... |

24 | M.: Validation of medical image processing in image-guided therapy
- Jannin, Fitzpatrick, et al.
(Show Context)
Citation Context ...ntation method (time). Other researchers have studied the sensitivity of the technique to various disruptive factors such as data artifacts, pathology, or individual anatomical variation (robustness) =-=[43]. -=-Accuracy metrics typically rely on a ground truth segmentation—segmentations that are somehow close to this ground truth are considered better than those that are not. Studies with digital or physic... |

23 | E.: Watershed-based segmentation of 3d mr data for volume quantization. Magnetic Resonance Imaging 15
- Sijbers, Verhoye, et al.
- 1997
(Show Context)
Citation Context ...gmentation algorithm. There are two strategies for evaluating segmentation algorithms. One strategy is to study segmentation performance in the context of a particular clinical or scientific question =-=[35, 36]-=-. For instance, the effectiveness of the algorithm within a study that monitors the volumes or sizes of tumors. The second approach is to study to evaluate segmentation in the absence of a specific cl... |

22 | C.: An adaptive level set method for medical image segmentation
- Droske, Meyer, et al.
- 2001
(Show Context)
Citation Context ...ecial case of the level-set PDEs in which the surface motion is strictly inward or outward. Such equations can be solved somewhat efficiently using the fast marching method [5] and variations thereof =-=[21]-=-. However, this case covers only a very small subset of interesting speed functions, and such speed functions are inconsistent with interactive parameter tuning. In general we are concerned with probl... |

21 | A.: A curve evolution approach to smoothing and segmentation using the mumford-shah functional
- Tsai, Yezzi, et al.
- 2001
(Show Context)
Citation Context ...1, 12]. Alternatively, the motion of the level set can depend on a variational formulation that positions the interface to create discontinuities that best model the discontinuities in the input data =-=[13, 14, 15]-=-. In this paper we use a supervised, statistical classifier to drive the motion of the level-set model. Virtually all of these methods include a form of mean curvature to keep the level-set smooth 4a... |

20 |
Y.: A Methodology for Evaluating Image Segmentation Algorithms
- Udupa, Leblanc, et al.
- 2002
(Show Context)
Citation Context ...] can generate data appropriate for statistical analysis that account for user variability. A combination of different factors determines the effectiveness of a segmentation. For instance Udupa et. al=-=[38]-=- propose a quantification of performance based on validity of the results (accuracy), reproducibility of the results (precision), and efficiency of the segmentation method (time). Other researchers ha... |

17 | Computer Graphics on a Stream Architecture
- Owens
- 2002
(Show Context)
Citation Context ...ications.sUnder review at Medical Image Analysis, 2004 5 Graphics processors outperform central processing units (CPUs)—often by more than an order of magnitude—because of their streaming architec=-=ture[25]-=- and dedicated high-speed memory. In the streaming model of computation, arrays of input data are processed identically by the same computation kernel to produce output data streams. The GPU takes adv... |

17 |
level-set curvature for image denoising
- Whitaker, Xue, et al.
- 2001
(Show Context)
Citation Context ...t and left optic nerves. The texture information in this data set posed a significant challenge, and therefore we preprocessed the data by smoothing with 10 iterations of modified-curvature diffusion =-=[53]-=-. This diffusion step blurs the more homogeneous regions of the data while preserving object boundaries. This nonlinear diffusion is relatively computationally expensive, especially on spectral data, ... |

16 |
Fronts Propogating with Curvature-Dependent Speed: Algorithms Based on Hamilton-Jacobi Formulations
- Osher, Sethian
- 1988
(Show Context)
Citation Context ...1 Level Sets Under review at Medical Image Analysis, 2004 3 This paper relies on an implicit representation of deformable surface models called the method of level sets, proposed by Osher and Sethian =-=[4]-=-. The level-set method (See also Sect. 3) computes the motion of a moving interface by solving a partial differential equation (PDE) on a volume. The use of level sets has been widely documented in th... |

15 |
J.A.: A fast level set method for propogating interfaces
- Adalsteinson, Sethian
- 1995
(Show Context)
Citation Context ...everal authors have proposed numerical schemes that compute solutions for only those voxels that lie in a small number of layers adjacent to the surface as shown in Fig. 1b. Adalsteinsson and Sethian =-=[22]-=- have proposed the narrow band method, which updates the embedding on a band of 10-20 pixels around the model, and reinitializes that band whenever the model approaches the edge. Whitaker [23] propose... |

15 | R.: Interactive, gpu-based level sets for 3d brain tumor segmentation
- Lefohn, Cates, et al.
- 2003
(Show Context)
Citation Context ... a user-assisted segmentation technique, which requires a user-based evaluation to capture variations in the individual decision-making process. Experimental trials across a number of users and images=-=[41, 42]-=- can generate data appropriate for statistical analysis that account for user variability. A combination of different factors determines the effectiveness of a segmentation. For instance Udupa et. al[... |

12 |
B.C.: Shape modeling with front propogation: A level set approach
- Malladi, Sethian, et al.
- 1995
(Show Context)
Citation Context ...rrow band of attraction (defined by the second derivative) in order to lock onto such edges, and therefore the author proposes a multiscale computational method to improve convergence. Malladi et al. =-=[10]-=- describe a formulation in which level curves/surfaces expand (or contract) with a motion that slows at image edges. Because of the monotonic expansion/contraction, convergence to local minima is less... |

12 | M.: Toward a common validation methodology for segmentation and registration algorithms
- Yoo, Ackerman, et al.
- 2000
(Show Context)
Citation Context ...ntation evaluation is difficult because of the lack of standard metrics and the difficulty of establishing ground truth in clinical data. Our evaluation methodology is derived from ideas developed by =-=[37], -=-and others [38–40], who emphasize the importance of quantitative evaluation and statistical metrics. The study in this paper concerns a user-assisted segmentation technique, which requires a user-ba... |

11 | G.: Robust estimation for brain tumor segmentation
- Prastawa, Bullitt, et al.
- 2003
(Show Context)
Citation Context ...definition of ground truth. However, for biological or clinical data sets, ground truth is usually unknown. In such a case, researchers typically rely on experts to delineate the ground truth by hand =-=[43, 44]-=-. Experts seldom all agree, but a statistical combination (averaging) of several expert segmentations can account for expert variability. Averaging of multiple nonparametric shapes, however, is itself... |

9 |
V.: A direct approach to estimating surfaces in tomographic data
- Whitaker, Elangovan
- 2002
(Show Context)
Citation Context ...1, 12]. Alternatively, the motion of the level set can depend on a variational formulation that positions the interface to create discontinuities that best model the discontinuities in the input data =-=[13, 14, 15]-=-. In this paper we use a supervised, statistical classifier to drive the motion of the level-set model. Virtually all of these methods include a form of mean curvature to keep the level-set smooth 4a... |

8 |
R.: Automated segmentation of mri of brain tumors. Radiology 218
- Kaus, Warfield, et al.
- 2001
(Show Context)
Citation Context ...spite its general nature, we demonstrate the effectiveness of GIST by a quantitative comparison to a specialized tool and the associated gold standard for a specific problem: brain tumor segmentation =-=[1, 2]-=-. This paper make the following contributions: • A 3D segmentation tool that uses a new level-set deformation solver to achieve interactive rates (approximately 10-15 times faster than previous soluti... |

7 | Automated segmentation of MRI of brain tumors
- Kaus, Warfield, et al.
- 2001
(Show Context)
Citation Context ...spite its general nature, we demonstrate the effectiveness of GIST by a quantitative comparison to a specialized tool and the associated gold standard for a specific problem: brain tumor segmentation =-=[1, 2]. -=-This paper make the following contributions: – A 3D segmentation tool that uses a new level-set deformation solver to achieve interactive rates (approximately 10-15 times faster than previous soluti... |