## A data-parallel algorithm for iterative tomographic image reconstruction (1999)

Venue: | In Proceedings of the 7th Symposium on the Frontiers of Massively Parallel Computation |

Citations: | 4 - 0 self |

### BibTeX

@INPROCEEDINGS{Johnson99adata-parallel,

author = {Calvin A. Johnson},

title = {A data-parallel algorithm for iterative tomographic image reconstruction},

booktitle = {In Proceedings of the 7th Symposium on the Frontiers of Massively Parallel Computation},

year = {1999},

pages = {126--137},

publisher = {IEEE Computer Society Press}

}

### OpenURL

### Abstract

In the tomographic imaging problem, images are reconstructed from a set of measured projections. Iterative reconstruction methods are computationally intensive alternatives to the more traditional Fourier-based methods. Despite their high cost, the popularity of these methods is increasing because of the advantages they pose. Although numerous iterative methods have been proposed over the years, all of these methods can be shown to have a similar computational structure. This paper presents a parallel algorithm that we originally developed for performing the expectation maximization algorithm in emission tomography. This algorithm is capable of exploiting the sparsity and symmetries of the model in a computationally efficient manner. Our parallelization scheme is based upon decomposition of the measurement-space vectors. We demonstrate that such a parallelization scheme is applicable to the vast majority of iterative reconstruction algorithms proposed to date. 1.

### Citations

687 | A new polynomial-time algorithm for linear programming
- Karmarkar
- 1984
(Show Context)
Citation Context ... constrained optimization methods such as activeset methods. Although interior-point methods were originally developed in the 1960’s by Fiacco and McCormick, recent developments in linear programming =-=[16, 23]-=- and nonlinear programming [25] have revived interest in them within the optimization community. In [15], we proposed two interior-point methods for the regularized ML reconstruction problem: a specia... |

351 |
Nonlinear Programming: Sequential Unconstrained Minimization Techniques
- Fiacco, McCormick
- 1968
(Show Context)
Citation Context ...y. These subproblems are solved for a descending sequence of barrier parameters 0 <µ k →0. Each subproblem solution satisfies a perturbed version of the Karush-Kuhn-Tucker (KKT) optimality conditions =-=[9, 26]-=-. Under mild conditions, it can be shown that as µ → 0, the subproblem solutions approach the optimal solution to the constrained problem (9) [9]. Theslogarithmic barrier method updates only the prima... |

305 |
Maximum likelihood reconstruction for emission tomography
- Shepp, Vardi
(Show Context)
Citation Context ...-negativity constraints on the image variables. Thus the constrained optimization problem is max fML (θ) s.t. θ ≥ 0. The ML-EM algorithm for emission tomography was first presented by Shepp and Vardi =-=[28]-=-. It requires a strictly positive initial solution θ 0 > 0 and consists of the following update equation: θ k+1 i = θk i qi N� j=1 (3) Ci,jyj ˆy k , i =1,...,n, (4) j where θ k i is the current (k-th ... |

290 |
Parallel Optimization: Theory, Algorithms, and Applications
- Censor, Zenios
- 1997
(Show Context)
Citation Context ...nfortunately, the parallelization strategy of Section 3 cannot be applied to a purely sequential-update algorithm such as ART. Block-iterative variants of ART such as the variable-block ART algorithm =-=[3]-=-, however, are indeed amenable to our parallelization scheme. 5.2 Generalized Landweber Iterations The generalized Landweber iteration is a scaled steepest descent method, where each iteration is give... |

215 |
Image Reconstruction from Projections: the Fundamentals of Computerized Tomography
- Herman
- 1980
(Show Context)
Citation Context ...s are reconstructed from measured projections collected from the scanning instrument. These projections are essentially a finite set of blurred and noisy line integrals through the object of interest =-=[11]-=-. Applications of the tomographic image reconstruction problem are quite numerous and include x-ray computed tomography (CT), ultrasound computed tomography, emission tomography (positron emission tom... |

197 | Accelerated image reconstruction using ordered subsets of projection data
- Hudson, Larkin
- 1994
(Show Context)
Citation Context ...mage is updated with each pass, but the update is performed only on the current (lth) block of projection space. A popular example of this is Hudson and Larkin’s “ordered subsets EM” (OSEM) algorithm =-=[13]-=-, in which the lth pass of the kth iteration consists of the following operations: ˆy k,l j θ k,l+1 = n� Ci,jθ k,l , j ∈ Sl, i=1 � i = θk,l i ql i j∈Sl Ci,jyj ˆy k,l , i =1,...,n, j where the outer it... |

157 |
Image Reconstruction and Restoration: Overview of Common Estimation Structures and Problems
- Demoment
- 1989
(Show Context)
Citation Context ...-4444 asofer@gmu.edu Reconstruction from projections is an inverse problem where the goal is to invert the forward process T {} from which the projections were generated in continuously defined space =-=[7]-=-. Defining the set of measurements to be the vector y ∈ℜNwith elements yj, j =1,...,N and the image solution estimate to be the vector x ∈ℜnwith elements xi,i =1,...,n, the inverse problem attempts to... |

156 |
reconstruction algorithms for emission and transmission tomography
- Lange, EM
(Show Context)
Citation Context ...ise vector update (6) is comparatively insignificant. A similar Poisson-likelihood problem exists in transmission tomography, and ML reconstruction methods for that problem (including an EM algorithm =-=[20]-=-) roughly parallel the methods presented here for the emission case.s2.1 Exploiting Sparsity in the Measurement Data ML reconstructions have been shown to yield improved image quality (over filtered b... |

115 | A unified approach to statistical tomography using coordinate descent optimization
- Bouman, Sauer
- 1996
(Show Context)
Citation Context ... algorithms are generally posed as solution methods for the regularized ML-EM algorithm. Recent papers have shown that carefully designed coordinate-ascent algorithms can result in faster convergence =-=[1, 8]-=-. Here we consider the single-coordinate update algorithm of Bouman et al. [1]. Defining ei ∈ℜnto be a vector with zeros in all elements except the ith, for which ei i =1, one iteration of the algorit... |

73 |
On use of the EM algorithm for penalized likelihood estimation
- Green
- 1990
(Show Context)
Citation Context ...egularized ML reconstruction under restricted classes of priors. One such method, which requires an upper threshold for γ and a finite bound on �∇R (θ)� ∞ , is Green’s “One Step Late” (OSL) procedure =-=[10]-=-. One iteration of the OSL algorithm is computed as θ k+1 i = θ k i ν k i qi + γ ∂ ∂θi R � θ k�, (10) where νk can be computed in parallel � in the manner of Algorithm 1. The computation of ∇R θ k� do... |

56 |
Alternatives to voxels for image representations in iterative reconstruction algorithms’, Phys
- Lewitt
- 1992
(Show Context)
Citation Context ...Second, the partitioning of projection space into symmetryrelated groups is much more difficult to implement in a coordinate-ascent algorithm. Incorporation of radiallysymmetric voxel basis functions =-=[22]-=- and on-the-fly computation of the system matrix may be more appropriate for this class of algorithm. 4.5 Interior-Point Methods Interior-point methods are constrained optimization methods that mainta... |

53 | Groupedcoordinate ascent algorithms for penalized-likelihood transmission image reconstruction
- Fessler, Ficaro, et al.
- 1997
(Show Context)
Citation Context ...he Poisson-likelihood reconstruction problem in emission and transmission tomography. The use of a Poisson model has been shown to improve image quality significantly in low-count scanning situations =-=[8, 14]-=-. The Poisson ML reconstruction problem is generally solved using a regularized ML functional and non-negativity constraints on the image variables. The emphasis of this paper is on the computational ... |

50 |
Implementing and accelerating the EM algorithm for positron emission tomography
- Kaufman
- 1987
(Show Context)
Citation Context ... (�r =0in the figure.) plexity analysis based only on multiplications and additions in (2) and (5). The three types of symmetries in the system matrix are illustrated in Figure 3. In-plane symmetries =-=[17]-=- involve reflection and rotation about φ and provide an 8-fold storage reduction. The axial symmetry [4] involves a reflection about β and yields a 2-fold symmetry (unless β =0,in which case there is ... |

46 |
Algebraic reconstruction techniques can be made computationally efficient
- Herman, Meyer
- 1993
(Show Context)
Citation Context ... methods. 5.1 ART-Family Methods The algebraic reconstruction technique (ART) is an iterative method that attempts to solve the inconsistent and generally overdetermined system of equations C T x = y =-=[11, 12]-=-. ART is a fully sequential-update algorithm, meaning that the kth iteration of ART updates the entire image estimate but operates on only the jkth measurement. The ART update is given by θ k+1 = θ k ... |

46 |
A theoretical study of some maximum likelihood algorithms for emission and transmission tomography
- Lange, Bahn, et al.
- 1987
(Show Context)
Citation Context ...sson-likelihood reconstruction in 3-D emission tomography, although our notation does not explicitly limit the scope of consideration. Similar ML methods exist for the transmission tomography problem =-=[8, 21]-=-. 4.1 Regularized ML-EM Methods Due to ill-posedness of the reconstruction problem, the objective function often must be regularized in order to obtain useful images [7]. The regularized ML objective ... |

40 |
A Row-action alternative to the EM algorithm for maximizing likelihoods in emission tomography
- Browne, AR
(Show Context)
Citation Context ... solution of (3), but rather to a limit cycle of N distinct solutions that are within a proximity to the optimal solution. The “row action maximum likelihood algorithm” (RAMLA) of Browne and DePierro =-=[2]-=- is computationally similar to OSEM and converges to an optimal solution of (3). To our knowledge, no block-iterative algorithm has been proven to converge to the optimal solution of the regularized p... |

37 |
Preconditioning methods for improved convergence rates in iterative reconstruction
- Clinthorne, Pan, et al.
(Show Context)
Citation Context ...ous update equation θ k+1 = θ k + M −1 k ∇f . � θ k� , (12) � where ∇f θ k� is the gradient of the objective function evaluated at the current point θ k , and Mk is a shaping matrix or preconditioner =-=[6, 30]-=-. We shall discuss preconditioners in Section 5.3. A common objective function in such problems is the weighted least squares (WLS) objective fWLS (θ) = 1 2 � � �D −1� � 2 T y −C θ �2 � � 2 (13) � y −... |

21 |
A barrier method for large-scale constrained optimization
- Nash, Sofer
- 1993
(Show Context)
Citation Context ...uch as activeset methods. Although interior-point methods were originally developed in the 1960’s by Fiacco and McCormick, recent developments in linear programming [16, 23] and nonlinear programming =-=[25]-=- have revived interest in them within the optimization community. In [15], we proposed two interior-point methods for the regularized ML reconstruction problem: a specialized logarithmic barrier metho... |

14 |
Block truncated-Newton methods for parallel optimization
- Nash, Sofer
- 1989
(Show Context)
Citation Context ... nonlinear CG methods is only marginally better than that of steepest descent methods. A faster solution of the unconstrained minimizer of fRW LS (θ) may be obtained via a truncatedNewton method (see =-=[24, 26]-=-.) 5.4 Interior-Point Methods Although the least-squares objective functions fLS (θ), fWLS (θ), and fRW LS (θ) are well defined outside of the non-negative orthant, negative solutions have no physical... |

10 | Evaluation of 3D reconstruction algorithms for a small animal PET camera
- Johnson, Seidel, et al.
- 1997
(Show Context)
Citation Context ...he Poisson-likelihood reconstruction problem in emission and transmission tomography. The use of a Poisson model has been shown to improve image quality significantly in low-count scanning situations =-=[8, 14]-=-. The Poisson ML reconstruction problem is generally solved using a regularized ML functional and non-negativity constraints on the image variables. The emphasis of this paper is on the computational ... |

7 |
A fast and stable maximum a posteriori conjugate gradient reconstruction algorithm, Med. Phys
- Lalush, Tsui
- 1995
(Show Context)
Citation Context ...tion. The regularized, weighted least-squares objective function is given by fRW LS (θ) = fWLS (θ)+γR(θ). The unconstrained problem of minimizing fRW LS (θ) can be solved via a nonlinear CG iteration =-=[19]-=-. Each iteration of the nonlinear CG method requires a gradient calculation ∇fRW LS (θ) =∇fWLS (θ)+γ∇R(θ), which is dominated by the cost of ∇fWLS (θ). Recalling from Section 5.2 that ∇fWLS (θ)is comp... |

6 |
On parallelizing the EM algorithm for PET image reconstruction
- Chen, Lee
(Show Context)
Citation Context ...hree types of symmetries in the system matrix are illustrated in Figure 3. In-plane symmetries [17] involve reflection and rotation about φ and provide an 8-fold storage reduction. The axial symmetry =-=[4]-=- involves a reflection about β and yields a 2-fold symmetry (unless β =0,in which case there is no gain). Axial parallel chord redundancies are the simplest to implement and yield the largest reductio... |

6 |
Iterative algebraic reconstruction algorithms for emission computed tomography: a unified framework and its application to positron emission tomography
- Xu, Liow, et al.
- 1993
(Show Context)
Citation Context ...ous update equation θ k+1 = θ k + M −1 k ∇f . � θ k� , (12) � where ∇f θ k� is the gradient of the objective function evaluated at the current point θ k , and Mk is a shaping matrix or preconditioner =-=[6, 30]-=-. We shall discuss preconditioners in Section 5.3. A common objective function in such problems is the weighted least squares (WLS) objective fWLS (θ) = 1 2 � � �D −1� � 2 T y −C θ �2 � � 2 (13) � y −... |

5 |
A general class of preconditioners for statistical iterative reconstruction of emission computed tomography
- Chinn, SC
- 1997
(Show Context)
Citation Context ...� θ = C T D −1 y. (15) The linear conjugate gradient (CG) method for solving symmetric, positive semidefinite systems of equations such as (15) [26] has been applied to the WLS reconstruction problem =-=[5, 18]-=-. The quality of the estimate for θ improves with each CG iteration. As such, an inexact solution for θ can be obtained from the CG method within (hopefully) a reasonable number of iterations. The cos... |

4 |
Constrained iterative reconstruction by the conjugate gradient method
- Kawata, Nalcioglu
- 1985
(Show Context)
Citation Context ...ion tomography (positron emission tomography or PET and single photon emission computed tomography or SPECT), electron microscopy, geophysical imaging, diffraction-limited optics, and radio astronomy =-=[18, 27]-=-. Closely related applications exist in other fields, e.g. the inverse radiotherapy planning problem [29]. ∗Ariela Sofer is partly supported by National Science Foundation grant DMI-9414355. Ariela So... |

3 |
Iterative methods for inverse treatment planning.” Phys
- Xing, Chen
- 1996
(Show Context)
Citation Context ... electron microscopy, geophysical imaging, diffraction-limited optics, and radio astronomy [18, 27]. Closely related applications exist in other fields, e.g. the inverse radiotherapy planning problem =-=[29]-=-. ∗Ariela Sofer is partly supported by National Science Foundation grant DMI-9414355. Ariela Sofer Operations Research and Engineering Department George Mason University Fairfax, VA 22030-4444 asofer@... |

2 |
Nonlinear optimization for volume PET reconstructions
- Johnson
- 1997
(Show Context)
Citation Context ... developed in the 1960’s by Fiacco and McCormick, recent developments in linear programming [16, 23] and nonlinear programming [25] have revived interest in them within the optimization community. In =-=[15]-=-, we proposed two interior-point methods for the regularized ML reconstruction problem: a specialized logarithmic barrier method and a primal-dual method. The logarithmic barrier method solves the con... |

2 |
Interior point path following primal-dual algorithms
- Monteiro, Adler
- 1989
(Show Context)
Citation Context ... constrained optimization methods such as activeset methods. Although interior-point methods were originally developed in the 1960’s by Fiacco and McCormick, recent developments in linear programming =-=[16, 23]-=- and nonlinear programming [25] have revived interest in them within the optimization community. In [15], we proposed two interior-point methods for the regularized ML reconstruction problem: a specia... |