M.I. Miller and B. Roysam. Bayesian image reconstruction for emission tomography: Implementation of the EM algorithm and Good's roughness prior on massively parallel processors. Proc. of the Natl. Acad. of Sci., 88:3223--3227, April 1991.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....the problem of coupled equations by plugging in values from the previous iteration. Unfortunately, such an approach can diverge, unless modified to include a line search [32] Similar strategies include the BIP algorithm [33] the methods in [34,35] and nested gradient or Jacobi iterations [36,37,21]. Most such strategies include a user specified step size parameter, and one user has noted that finding good values for [the step size] and the number of times to iterate requires painful experimentation [38] Other OSL like methods are given in [38,39] which have been reported to occasionally ....

M I Miller and B Roysam, "Bayesian image reconstruction for emission tomography incoporating Good's roughness prior on massively parallel processors", Proc. Natl. Acad. Sci., vol. 88, pp. 3223--3227, Apr. 1991.

.... in nonparametric probability density estimation; a thorough analysis in this context is given by Tapia and Thompson [TT78] Following the suggestion of Snyder and Miller ( SM85] Section II.1) Good s roughness was later applied to closely related problems of Poisson intensity estimation in PET [MR91], SPECT [MM91, BM93, MB93, BMMW94] and optical sectioning microscopy [JM93] In the next few subsections, in order to conveniently express the penalties, we will index # 2 using two spatial coordinates, as in # 2 i 1 ,i 2 , i 1 = 1, I 1 , i 2 = 1, I 2 , where I = I 1 I 2 . ....

....d 2 dx 2 ln # 2 (x)dx. 1.20) The last equality in (1.20) which was established independently in [O S95] and [Fri91] follows from integration by parts. Consider the rightmost expression. The penalty is straightforwardly extended to two dimensions (see pp. 155 156 of [SM91] or Section 3 of [MR91]) # # # 2 (x, y) # # 2 #x 2 # 2 #y 2 # ln # 2 (x, y)dxdy. 1.21) Discretizing Equation (1.21) yields #G (f) # i 1 ,i 2 # 2 i 1 ,i 2 [ln # 2 i 1 1,i2 ln # 2 i 1 1,i2 ln # 2 i,i 2 1 ln # 2 i 1 ,i 2 1 4 ln # 2 i 1 ,i 2 ] 1.22) O Sullivan ....

M.I. Miller and B. Roysam. Bayesian image reconstruction for emission tomography incorporating Good's roughness prior on massively parallel processors. Proc. of the National Academy of Sciences, 88:3223--3227, April 1991.

.... e.g. regularization by the method of kernel of sieves [9, 10, 11] smoothing [12] and MAP estimation which finds the that maximizes the posterior probability P (jn ) P (n j) P ( using Poisson and Gaussian priors [13, 14, 15, 16, 17] penalized likelihoods [18] Good s measure of roughness [19], and Gibbs priors that induce Markov random fields [20, 21, 22, 23, 24] With respect to the computational cost, various attempts have been made to reduce the number of iterations by accelerating convergence, e.g. overrelaxation [25] line search [7, 8] multigrids [26, 5] incremental ....

.... computing has been studied and many implementations have been reported, e.g. for the iPSC 2 hypercube and the BBN Butterfly GP 1000 shared memory computer [29] the Cray 1 vector computer [30] the Alliant FX 8 shared memory multiprocessor [31] the MasPar 4096 and AMT DAP 4096 SIMD computers [19, 32, 33], the Thinking 3 Machines CM 5 MIMD computer and a network of loosely coupled SUN workstations [34] We present a focus of attention preprocessing scheme for reducing the time and space requirements of the EM ML algorithm. Simply stated, our approach is to first determine which of the equations ....

M. I. Miller and B. Roysam, "Bayesian image reconstruction for emission tomography incorporating Good's roughness prior on massively parallel processors," Proc. Natl. Acad. Sci., vol. 88, pp. 3223--3227, 1991.

.... of the EM reconstruction technique have been proposed as early as 1985 (Llacer and Meng, 1985) and preliminary results on the application of the EM algorithm on parallel machines, especially in attempts to perform 3 D image reconstruction in PET, have been published since the late 80s (Miller, et al., 1988; Hebert and Leahy, 1989; Herman, et al., 1990) The main disadvantage of these implementations is that parallel programming for dedicated architectures is highly platform dependent. Recently, the idea of distributed processing on a cluster of workstations seems to become more and more popular. ....

M.I. MILLER, B. ROYSAM AND A.W. McCARTHY, Bayesian image reconstruction for Emission Tomography: implementation of the EM algorithm and Good's roughness prior on massively parallel processors, Intl. Conf. on Acoustics, Speech and Signal Processing, March 1988.

....that exactly match the original kernel results. A variation of this recursive blurring approach, based on the diffusion equation, has been used extensively by a group at the Washington University of St. Louis in massively parallel implementations of iterative PET and SPECT reconstructions [48, 85, 86, 87, 88]. They use a single three point Gaussian kernel to generate the entire family of depth dependent Gaussians. The desired fwhm is obtained by repeated convolutions with the single base kernel. This implementation maps efficiently to mesh connected arrays of bit serial processing elements. Each depth ....

....base kernel. This implementation maps efficiently to mesh connected arrays of bit serial processing elements. Each depth requires m iterations to produce the result of a convolution with a kernel of the desired variance of m 2 . That is, they approximate convolution with a 1 D Gaussian by [85]: y m 1 [n] 1 4 y m [n 1] 1 2 y m [n] 1 4 y m [n Gamma 1] 5.8) This evokes mention of other multiplier free approaches such as used in [89] which appeal to the central limit theorem and begin with simple box filters to generate Gaussians. A more efficient approach for most computed ....

[Article contains additional citation context not shown here]

M. I. Miller and B. Roysam, "Bayesian image reconstruction for emission tomography incorporating Good's roughness prior on massively parallel processors," in Proc Natl Acad Sci USA, vol. 88, pp. 3223--3227, 1991.

....the problem of coupled equations by plugging in values from the previous iteration. Unfortunately, such an approach can diverge, unless modified to include a line search [32] Similar strategies include the BIP algorithm [33] the methods in [34,35] and nested gradient or Jacobi iterations [36,37,21]. Most such strategies include a user specified step size parameter, and one user has noted that finding good values for [the step size] and the number of times to iterate requires painful experimentation [38] Other OSLlike methods are given in [38,39] which have been reported to occasionally ....

M I Miller and B Roysam, "Bayesian image reconstruction for emission tomography incoporating Good's roughness prior on massively parallel processors", Proc. Natl. Acad. Sci., vol. 88, pp. 3223--3227, Apr. 1991.

....the problem of coupled equations by plugging in values from the previous iteration. Unfortunately, such an approach can diverge, unless modified to include a line search [46] Similar strategies include the BIP algorithm [47, 48] the methods in [49, 50] and nested gradient or Jacobi iterations [51, 52, 29]. Most such strategies include a user specified step size parameter, and one user has noted that finding good values for [the step size] and the number of times to iterate requires painful experimentation [53] Other OSLlike methods are given in [53, 54] which have been reported to occasionally ....

M I Miller and B Roysam. Bayesian image reconstruction for emission tomography incoporating Good's roughness prior on massively parallel processors. Proc. Natl. Acad. Sci., 88:3223--3227, April 1991.

.... reported for hypercube iPSC 2 and shared memory multiprocessor BBN Butterfly GP1000 [5] Alliant FX 8 shared memory multiprocessor [13] vector machine Cray 1 [17] special parallel architecture using two arrays of VLSI chips [15] and massively parallel SIMD computers (AMT DAP 4096, MasPar 4096) [4, 27, 28]. 1.3 Objectives and Overview This research extends the convergence results for the EM ML algorithm and presents new results in parallel implementations using both simulated and real clinical data. The dissertation presents the following results: 1. Aspects of sequence and series convergence of ....

Miller, M. I. and Roysam, B. Bayesian image reconstruction for emission tomography incorporating Good's roughness prior on massively parallel processors. Proc. Natl. Acad. Sci. USA., 88:3223--3227, 1991.

....for maximum likelihood estimation of parameters tend to be iterations with long computation times involving large arrays. One way to address these problems of time and memory is to implement the iteration on multiple processors in parallel, for example, using a massively parallel computer [3, 4] or taking advantage of specific interconnection topologies [5, 6] Parallel iteration can be distributed over a network of workstations. Distributed iteration generally reduces the memory required at any one processor by allowing each processor to use only portions of large, sparse matrices. ....

M I Miller and B Roysam. Bayesian image reconstruction for emission tomography incorporating Good's roughness prior on massively parallel processors. Proc. Natl. Acad. Sci., 88:3223--3227, 1991.

No context found.

M.I. Miller and B. Roysam. Bayesian image reconstruction for emission tomography: Implementation of the EM algorithm and Good's roughness prior on massively parallel processors. Proc. of the Natl. Acad. of Sci., 88:3223--3227, April 1991.

No context found.

M. I. Miller and B. Roysam, "Bayesian image reconstruction for emission tomography: Implementation of the EM algorithm and Good's roughness prior on massively parallel processors," in Proc. Nat. Acad. Sci., 1991, vol. 88, pp. 3223--3227.

No context found.

M. I. Miller and B. Roysam. "Bayesian Image Reconstruction for Emission Tomography Incorporating Good's Roughness Prior on Massively Parallel Processors". Proc. Natl. Acad. Sci. USA, 88:3223--3227, April 1991.

....developed. The algorithms developed are very computationally intensive. In problems such as these the class of massively parallel single instruction multiple data stream computers (SIMD) is promising implementation speeds which are orders of magnitude faster than conventional RISC workstations [11, 12, 13]. Currently we use a DECmpp 12000Sx and an AMT DAP 610 SIMD parallel processors to implement our algorithms in clinical time frames. Many of the above problems are inherently 3 D and some 4 D (3 spatial and 1 temporal) in nature and require complex visualization tools for analysis. Although the ....

M.I. Miller and B. Roysam. Bayesian image reconstruction for emission tomography: Implementation of the EM algorithm and Good's roughness prior on massively parallel processors. Proc. of the Natl. Acad. of Sci., 88:3223--3227, April 1991.

No context found.

M. Miller and B. Roysam, "Bayesian image reconstruction for emission tomography incorporating Good's roughness prior on massively parallel processors," Proc. of the Natl. Acad. of Sci. 88, pp. 3223--3227, April 1991.

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