G. E. Christensen, M. I. Miller and M. W. Vannier, "Individualizing neuroanatomical atlases using a massively parallel computer," Computer, pp. 32-38, January 1996.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... basis [19, 20] Finally, true displacement fields, sometimes also called optical flow, result from the use of a (continuous) functional optimization scheme, in which an independent displacement is computed for each point in the image, with constraints arising from some a priori regularization [17, 21, 22, 23]. 2.3 Search strategy Given a set of features and a parametric deformation, both the criterion to optimize and the optimization algorithm itself define the search strategy. The use of a least squares criterion jointly with geometric primitives is popular [24] although it is sometimes replaced ....

G.E. Christensen, M.I. Miller, M.W. Vannier and U. Grenander, "Individualizing Neuroanatomical Atlases Using a Massively Parallel Computer," Computer, vol. 29, no. 1, pp. 32--38, January 1996.

....present more precisely the minimization problem studied here, as well as the assumptions to impose on the model. The rst attempts dealing with the issue of 2 medical image matching can be traced back to Bajcsy and Kovacic [2] who used the Jacobi method. These ideas were further developed in [3] using successive overrelaxation and [1] using Fourier and wavelet representations of a nonlinear functional. We will describe a linearized method for the minimization of a nonlinear so called output least square functional. The corresponding Euler equations are a coupled system of linear partial ....

....we use a nite di erence method and assign nodes in the elastic model to each picture element by a (pixel or voxel ) centered discretisation. In particular in the case of 3D digital images we face a huge problem which, in practice, is unsolvable only with relaxation methods. Christensen et al. [3]use successive overrelaxation (SOR) with checkerboard update to compute viscous uid deformations of a template. Because of h dependence of the SOR, the algorithm is rather slow. For instance, they have an execution time of 7 days with a MIPS R440 processor and 2 6 hours using a massively parallel ....

G.E. Christensen, M.I. Miller, M.Vannier and U.Grenander, Individualizing neuroanatomical atlases using a massively parallel computer, IEEE Computer, 29(1):32-38, (1996).

....by the registration algorithms. This different subject registration problem is not specifically addressed in this thesis. 2. 3 Surface based Registration Related three dimensional registration approaches have been based on features (points [58] or curves [31] surfaces [49] 56] and volumes [16]: Feature based approaches extract points or curves that are likely to be stable and descriptive and then search for consistent correspondences between the features extracted from both datasets. Sample features are points with maximum Gaussian curvature along isosurfaces of volumetric data. ....

G.E. Christensen, M.I. Miller, M.W. Vannier, U. Grenander. "Individualizing Neuroanatomical Atlases Using a Massively Parallel Computer". IEEE Computer, 29(1): 32-38, January 1996.

....work devoted to the warping of 3D brain data sets is also closely related to the approach explained in this dissertation. The brain warping algorithms make use of continuum mechanics and present material models to regularize the motion field estimation which are similar to those used in this work [7,14,92]. 1.3. Summary of Results The take home message from this dissertation is that by more accurately modeling the elastic properties of the materials being imaged in cardiac PET, the motion estimation algorithm can be made more accurate. The estimated motion field can be used to compensate for the ....

....a material model of continuous media gave acceptable results. The linear elastic model assumed infinitesimal displacements, and since the deformations required to match different brain data sets can be quite large, it was a simplifying approximation to the actual material being imaged. Christensen [15,14] tried to overcome this problem by introducing a viscous fluid model capable of tracking large deformations. He used this technique to match largely differing brain data sets from different patients. Other authors have used different similarity measures than voxel matching while retaining the ....

[Article contains additional citation context not shown here]

G E Christensen, M I Miller, and MW Vannier. "Individualizing neuroanatomical atlases using a massively parallel computer." Computer, 29(1):32--38, 1996.

.... may be advantageous, as large deformations are made more probable than in the elastic theory [7] On the other hand, their implementation requires extraordinary computational resources because of the nonlinearity introduced through the kinematic variables and the constitutive relations [6]. A more fundamental difficulty with large displacements is that the likelihood of false matches increase. A standard approach is to solve the problem at different spatial scales, as in the multi resolution version of elastic matching described in [2] large scale displacements are first ....

G. E. Christensen, M. I. Miller, M. W. Vannier, and U. Grenander. Individualizing neuroanatomical atlases using a massively parallel computer. IEEE Computer, 29(1):32--38, 1996.

....their atlas can have difficulty matching complicated object boundaries. Their method is computationally expensive and requires interactive and time consuming preprocessing. Christensen et al. presented a method very close to ours, except that they used a fluid dynamic model for the deformation [1], 2] It takes 1.8 hours to match two 128x128x100 volumes on a 16384 processor MasPar computer. In [10] Thirion takes a similar approach to ours, except that he assumes the volumes are already globally aligned, and applies optical flow from the beginning. To reduce computation time, he used the ....

Christensen et al., "Individualizing Neuroanatomical Atlases Using A Massively Parallel Computer", IEEE Computer, pp. 32-38, January 1996.

....shape, size, and location in these two brains. For registration algorithms that use only intensity or shape templates to achieve correspondence, results are typically poor due to these inherent variations. Currently there exist many intensity correspondence based registration algorithms [1], 3] 5] 13] Figure 2 shows a registration result using method [5] The right image volume is deformed to register with the left image volume, and outlines of its anatomical structures are overlaid on the left image to illustrate the alignment. Note that there is significant misalignment ....

Christensen et al., "Individualizing Neuroanatomical Atlases Using A Massively Parallel Computer", IEEE Computer, pp. 32-38, January 1996.

....atlas can have difficulty converging to complicated object boundaries. Also their method is more computationally expensive and requires interactive and timeconsuming preprocessing. Christensen et al. presented a method very close to ours, but they used a fluid dynamic model for the deformation [1], 2] It constrains neighboring voxels to have similar deformations, while allowing large deformation for small sub volumes. It takes 1.8 hours to match 128x128x100 volumes on a 16384 processor MasPar, while our algorithm takes 12 minutes to match 256x256x124 volumes on an SGI with four 194 MHz ....

Christensen et al., "Individualizing Neuroanatomical Atlases Using A Massively Parallel Computer", IEEE Computer, pp. 32-38, January 1996.

....directly to the volume data. Some algorithms in the recent past have used the direct approach, and The technique in [6] uses the concept of maximizing mutual information and reported results are quite impressive however, there is no provision for handling local motions in this method. In [3], a novel approach is described wherein the registration is modeled by a viscous fluid flow model expressed as a PDE. The reported cpu times for registration are however slow. Thirion [10] introduced an interesting demon based registration that can be viewed as being similar to the fluid based ....

G.E.Christensen et. al. "Individualizing neuroanatomical atlases using a massively parallel computer". IEEE Computer, 29(1):32--38, January 1996.

....and multiple motions. Bajscy and Kovacic (Bajscy and Kovacic, 1989) studied registration under nonrigid deformations using volumetric deformations based on elasticity theory of solids. Other direct approaches have used the so called fluid registration approach introduced by Christensen et al. (Christensen el al. 1996). In this approach the registration transformation is modeled by a viscous fluid flow model expressed as a partial differential equation (PDE) The model primarily describes the registration as a local fluid deformation expressed by a nonlinear PDE and more recently by a linear version ....

Christensen, G. E., Miller, M. I. and Vannier, M. (1996) Individualizing Neuroanatomical Atlases Using a Massively Parallel Computer. IEEE Computer, 1(29):32-38.

....an atlas image to each individual, or study, image, in order to have a common coordinate system for comparison. Shape differences between the atlas and study s anatomy are contained in the non rigid transformation. There have been many approaches to non rigid registration in recent years [2] 3] [4] [6] 8] 9] 10] 15] Usually, the transformation is constrained in some way because of the ill posedness (i.e. in this case, the existence of many possible solutions) of the problem. Physical models, for example, linear elastic models, are widely used to enforce topological properties on the ....

....in some way because of the ill posedness (i.e. in this case, the existence of many possible solutions) of the problem. Physical models, for example, linear elastic models, are widely used to enforce topological properties on the deformation and then constrain the enormous solution space [2] [4] [8] 9] 10] Here, we are particularly interested in intensity based deformation using elastic models. Our goal is to incorporate statistical shape information into this type of elastic model based registration and to develop a more accurate and robust algorithm. Christensen et al. 4] present ....

[Article contains additional citation context not shown here]

G. E. Christensen, M. I. Miller and M. W. Vannier, "Individualizing neuroanatomical atlases using a massively parallel computer," Computer, pp. 32-38, January 1996.

....Unfortunately, the algorithm proposed by Christensen et al. is rather slow. They originally implemented the algorithm using a massively parallel DECmpp 128x64 MasPar computer on which the algorithm used on the order of 5 10 minutes for 2D and 2 6 hours for 3D registrations. In a recent paper [6] they show estimates of the execution time on a MIPS R4400 processor on the order of 2 hours for 2D and 7 days for 3D. In practice this means that the algorithm is not feasible unless a massively parallel computer is available. The contribution of this paper is a new fast algorithm based entirely ....

....But our timings are quite different. We have achieved stable timings on a single processor workstation similar to those stated by Christensen et al. for computations on a 128x64 DECmpp 12000 Sx Model 200 massively parallel computer. When compared to estimates of timings for a MIPS R4400 processor [6] we can conclude that we achieve a speed up of at least an order of magnitude. We hope to be able to share the data used by Christensen et al. for more elaborate comparison. 3.1 Comparison with demon based registration In [13] Thirion proposed a demon based registration method. This is an ....

G.E. Christensen, M.I. Miller, M. Vannier and U. Grenander, Individualizing neuroanatomical atlases using a massively parallel computer, IEEE Computer, 29(1):32-38, January 1996

....matching an atlas image to each individual, or study, image, in order to have a common coordinate system for comparison. Shape differences between the atlas and study s anatomy are contained in the nonrigid transformation. There have been many approaches to non rigid registration in recent years [2, 3, 6, 7, 8, 11]. Usually, the transformation is constrained in some way because of the ill posedness (i.e. in this case, the existence of many possible solutions) of the problem. Physical models, for example, linear elastic and viscous fluid models, are widely used to enforce topological properties on the ....

....some way because of the ill posedness (i.e. in this case, the existence of many possible solutions) of the problem. Physical models, for example, linear elastic and viscous fluid models, are widely used to enforce topological properties on the deformation and constrain the enormous solution space [3, 4, 6, 7, 8]. Here, we are particularly interested in intensity based deformation using elastic or fluid models. Our goal is to incorporate statistical shape information into this type of physical model based registration and to develop a more accurate and robust algorithm. Christensen et al. 3] present two ....

[Article contains additional citation context not shown here]

G. E. Christensen, M. I. Miller and M. W. Vannier, "Individualizing neuroanatomical atlases using a massively parallel computer," Computer, pp. 32-38, January 1996.

....and are solved in sequence starting at and These PDE s are solved numerically for the instantaneous velocity using successive overrelaxation (SOR) 46] with checkerboard update at each fixed time step. The discrete version of (10) is given by (12) Automatic regridding is performed as in [7] and [47], by propagating templates as the nonlinear transformations evaluated on the finite spatial lattice become singular. New templates are propagated when the Jacobian of the transformation of the current template drops below 0.5. C. Small Versus Large Deformation Models Due to the complex shape of ....

.... 200 (MasPar) a 128 128 meshconnected single instruction multiple data (SIMD) architecture which is well suited for solving partial differential equations such as (11) The 3 D fluid transformation takes roughly 2 h for a 128 128 100 voxel data set, 100 SOR iterations, and 250 time steps [47]. The solution of the nonlinear fluid PDE (11) is iterative. Fig. 2 shows a plot of the squared difference of the intensities of the deformed template and the target versus number of iterations for the transformation of two 128 128 100 voxel data sets. The number of iterations used to compute the ....

G. E. Christensen, M. I. Miller, U. Grenander, and M. W. Vannier, "Individualizing neuroanatomical atlases using a massively parallel computer," IEEE Comput., Mag., pp. 32--38, Jan. 1996.

No context found.

G. E. Christensen, M. I. Miller and M. W. Vannier, "Individualizing neuroanatomical atlases using a massively parallel computer," Computer, pp. 32-38, January 1996.

No context found.

G. E. Christensen, M. I. Miller, M. W. Vannier, and U. Grenander, "Individualizing neuroanatomical atlases using a massively parallel computer, " IEEE Comput., vol. 29, no. 1, pp. 32--38, Jan. 1996.

No context found.

G. E. Christensen, M. I. Miller, M. W. Vannier, and U. Grenander, "Individualizing neuroanatomical atlases using a massively parallel computer, " IEEE Comput., vol. 29, no. 1, pp. 32--38, Jan. 1996.

No context found.

G.E. Christensen, M.I. Miller, M.W. Vannier and U. Grenander. Individualizing Neuroanatomical Atlases Using a Massively Parallel Computer. IEEE Computer, January, 1996.

No context found.

G.E. Christensen, M.I. Miller, M.W. Vannier and U. Grenander. Individualizing Neuroanatomical Atlases Using a Massively Parallel Computer. IEEE Computer, January, 1996.

No context found.

Christensen et al., "Individualizing Neuroanatomical Atlases Using A Massively Parallel Computer", IEEE Computer, pp. 32-38, January 1996.

No context found.

Christensen et al., "Individualizing Neuroanatomical Atlases Using A Massively Parallel Computer", IEEE Computer, pp. 32-38, January 1996.

No context found.

G. E. Christensen, M. I. Miller, M. W. Vannier, and U. Grenander, "Individualizing neuroanatomical atlases using a massively parallel computer," IEEE Computer 29(1), pp. 32--38, 1996.

No context found.

G. E. Christensen, M. I. Miller, M. W. Vannier, and U. Grenander, "Individualizing neuroanatomical atlases using a massively parallel computer," IEEE Computer 29(1), pp. 32--38, 1996.

No context found.

G.E. Christensen, et. al., "Individualizing Neuroanatomical Atlases Using a Massively Parallel Computer," IEEE Computer, pp. 32-38, Jan. 11986.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC