A discrete-continuous Bayesian model for Emission Tomography

Please download to get full document.

View again

of 4
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Document Description
A discrete-continuous Bayesian model for Emission Tomography
Document Share
Document Transcript
  A DISCRETE-CONTINUOUS BAYESIAN MODEL FOR EMISSION TOMOGRAPHY  Mame Diarra Fall  † ⋆  , ´  Eric Barat   ⋆  , Claude Comtat   ‡  , Thomas Dautremer   ⋆  , Thierry Montagu  ⋆ and Ali Mohammad-Djafari  †† Laboratoire des Signaux et Syst`emes, (UMR8506: CNRS–SUPELEC–Univ Paris-Sud), 91192 Gif-sur-Yvette, FRANCE. ⋆ Laboratoire de Mod´elisation, Simulation et Syst`emes (LM2S), CEA-Saclay, 91191 Gif-sur-Yvette, FRANCE ‡ Service Hospitalier Fr´ed´eric Joliot, CEA/DSV/I2BM, 91401 Orsay, FRANCE ABSTRACT In this contribution, we propose a discrete-continuous recon-struction method for Positron Emission Tomography (PET).The goal is to reconstruct a continuous radiotracer activitydistribution from a finite set of measurements (namely, thediscrete projections of detected random emissions). Our ap-proach can be viewed as an indirect density estimation prob-lem,  i.e , the problem of recovering a probability density func-tion based on indirect observations. We cast the reconstruc-tion problem in a  Bayesian nonparametric estimation  frame-work where regularization of the ill-posed inverse problem isachieved by putting a prior on the investigated radiotracer ac-tivity distribution.We proposea hierarchicalmodeland useitfor theMCMCschemes to generate samples from the posterior activity distri-bution and compute its functionals (mean, standard deviationetc.). Results will illustrate the performances of the proposedmethod and we compare our approach to another Bayesianmethod, the maximum a posteriori estimation (MAP), whichis based on a fully discrete-discrete problem formulation.  Index Terms — Bayesian nonparametrics, indirect densityestimation, MCMC sampling, Positron Emission Tomogra-phy 1. INTRODUCTION An interesting feature of Emission computed Tomographyis in producing 3D images that give information about themetabolic activity of an organ (the brain, for example). InPositron Emission Tomography (PET), the subject is injectedwith a molecule labelled with a positron emitter radionuclide.The emitted positrons travel a small distance before anni-hilating with an electron. This annihilation generates twogamma photons that fly-off at the speed of light in oppositedirections. These photons are detected by rings of detectorsaround the subject. If two photons are detected within a shorttime interval, a coincidence event is recorded revealing thata positron emission occurred in the virtual line joining bothdetectors (called line-of-response (LOR)). Tomographic algo-rithms are used to reconstruct 3D images of the radioactivitydistribution from the collected data sets.Conventional image reconstruction algorithms in Emis-sion Tomography are roughly divided into two categories: an-alytical and statistical methods. Analytical methods like fil-tered back-projection algorithmsare based on the directinver-sion of the X-ray transform [1] and address the reconstructionproblem (at least theoretically) in continuous data and imagespaces. They provide fast reconstructions, however they arebasedonover-simplifiedmodelsofthephysicalprocessesanddo not account for the stochastic nature of positron emissionand photon detection. As a consequence, reconstructed im-ages suffer from a lot of artifacts and significant noise. Toovercome these deficiencies, statistical methods have beenproposed to model more realistically the process of acquir-ing data by taking into account the random nature of the phe-nomenon. They are based on optimization methods, typi-cally maximum likelihood approaches (ML) ([2], [3]). They usually provide reconstructed images with a higher signal-to-noise ratio than analytical methods but nevertheless have highnoise due to ill-conditioned nature of the problem. This canbe solved by adding a regularization term in the form of a pri-ori density for the image using Bayesian approaches.WhenBayesian methods are used in Emission Tomography, theyare based on the maximum a posteriori approach (MAP) ([4],[5]). However, all these statistical methods require a discreteformulation of the radiotracer distribution in the space overwhich the reconstruction is needed, which is not natural be-cause emission distribution is continuous.The approach presented in this paper is a statistical onebut is fundamentally different from those mentioned above.The continuous radioactivity distribution is directly recon-structed from the discrete data. Thus, this can be viewed asa third category in image reconstruction algorithms for Emis-sion Tomography. We consider the radiotracer activity distri-bution as being a probability density on  R 3 and infer on it.This approach is called nonparametric because the parame-ter of interest, which is a probability distribution, is infinitedimensional and it is Bayesian since we put a prior on thisdistribution and infer on its posterior. So, the advantage of this approach is in offering a Bayesian regularization contextfor nonparametric estimation. In addition, we have accessto the entire distribution of the posterior activity distribution,not only a point estimate like in ML and MAP approaches.Therefore, any posterior uncertainty (like variance, highestposterior density regions etc.) can be estimated, which is wellsuited for quantitative imaging in “small samples” (referredto as “low doses” in PET).The Bayesian nonparametric approach was first used forthe two-dimensional (2D) PET reconstruction in [6]. In thispaper, we adopt the same approach for the three dimensionalcase. The 3D case is, however, more complicated since itrequires to deal with truncated data due to the limited field-of-view of the system (FOV). The sampling scheme used inthis work is different from that in [6] and is not based on anytruncation of the infinite dimensional distribution.The rest of this paper is structured as follows. In section 2,we describe the general problem formulation in the Bayesiannonparametric context. In section 3, the nonparametric priormodel for the unknown parameter of interest is given. Our hi- 2011 18th IEEE International Conference on Image Processing 978-1-4577-1303-3/11/$26.00 ©2011 IEEE1401  erarchical model for emission tomographic data is presentedin section 4. Afterwards, the derived MCMC sampling is out-lined in section 5. Results illustrating the quality of the pro-posed algorithm compared to MAP-EM are shown in section6 and we conclude the paper in section 7. 2. PROBLEM STATEMENT IN THE BAYESIANNONPARAMETRIC FRAMEWORK In this section, we develop the statiscal framework for theBayesian nonparametric reconstruction applied to PET data.We are interested in reconstructing the 3D image of theradiotracer activity distribution from measured data. We as-sume that data are stored in the so-called  list-mode  formatsuch that the coordinates of the detectors receiving two coin-cident photons are observed (the approach is also applicableto  sinogram mode , see [6]). We denote by the term “LOR”the virtual line that connects two detectors in coincidence.A LOR  l  is parametrized by  ( s l ,̟ l , r l ,ϕ l )  where  s l  is thedistance between the  z -axis (ring center) and the LOR in atransaxial plane  ( xy ) ; the LOR and the  x -axis define the az-imuthal angle  ̟ l ;  r l  is the axial coordinate of the LOR (if the coordinates of the two detector crystals in coincidence are ( z 1 ,z 2 ) ,  r l  = ( z 1  +  z 2 ) / 2) . Finally, the LOR forms the an-gle  ϕ l  with the transaxial plane. We map LOR parameters tothe index  l  such that the random variable  y i  is the index of the LOR corresponding to the  i th observed coincidence. Theinverse problem can be formulated as follows in the Bayesiannonparametric context [7] G ∼G  F   ( · ) =   X  P  ( ·| x )  G (d x ) y i iid ∼ F,  for  i  = 1 ,...,n (1)where X ⊂  R 3 is the object space ( i.e , the space over whichreconstructioniscarriedout);  G standsforthespatialdistribu-tion of recorded events we want to estimate from the observed F  -distributed data sets Y  = ( y 1 ,...,y n ) . In Bayesian non-parametrics, we must put a prior over the unknown distribu-tion  G . This is given by the prior law G  of   G . The relation be-tween the data  y i  and the emissions locations x i  is embodiedin a probability distribution P  ( ·| x )  called projection distribu-tion,  i.e , the probability for detecting photons pair in a LOR  l given that the corresponding positron emission occurred in x .In the formulation (1) of the inverse problem,  G ( · )  is thespatial distribution for  recorded events . Nevertheless, dataare truncated in the 3D case due to the finite axial extent of the PET scanner and detector efficiencies. Let introduce therandom variable  y ∗ :=  y >  0  to denote a recorded data. So, G (d x ) =  G ∗ (d x | y ∗ ) , where  G ∗ (d x | y ∗ )  is the distribution of  x conditionedonthefactthattheeventhasbeendetected. Thedistribution  G † ( · )  for whole events (not only the recorded)can be expressed using a derivation of Bayes’ rule, G † (d x ) =  G ∗ (d x | y ∗ ) Pr ( y ∗ ) Pr ( y ∗ | x )  (2)where for all  x  ∈ X  , Pr ( y ∗ | x ) =   Ll =1  Pr ( y  =  l | x )  isthe probability for recording an emission located in  x  bythe system accounting for system’s geometry and physi-cal properties. Since  G † (d x )  is a normalized version of  G ∗ (d x | y ∗ ) / Pr ( y ∗ | x ) , posterior inference on  G ∗ (d x | y ∗ )  issufficient to estimate the posterior of   G † (d x )  as soon asPr ( y ∗ | x )  >  0 ,  for all  x  ∈ X  . For sake of simplicity in thenotations, in the rest of this paper we omit the condition-ing since we deal with the recorded events in the inference.We use  G (d x )  to stand for  G ∗ (d x | y ∗ ) , the distribution of arecorded event.We model the detection probability using Bayes’ theorem:Pr ( y  =  l | x ) =  f  geom ( x | y  =  l ) Pr ( y  =  l )  Ll =0  f  geom ( x | y  =  l ) Pr ( y  =  l ) (3)where  y  = 0  stands for an unobserved event,  f  geom ( x | y  =  l ) is the geometric probability density that a photons pair de-tected in a LOR  l  srcinated from a spatial location x . Thisdistribution accounts for the fact that the probability to de-tect in coincidence emissions from inside the system’s FOVis position-dependent. Indeed, if we consider perfect detec-tors, the probability to record true events is merely geometric.It is determined by computing for each location x , the solidangle subtended from this point to the faces of the detectors.Other effects such as non-collinearity of photons and positronrange due to the small distance the emitted positron travelsbefore annihilating with an electron can also be included in f  geom .The probability for observing an event in a LOR  l  is modeledas follows, for any  l >  0 Pr ( y  =  l ) ∝ n ( l )  a ( l ) .  (4)For each detector pair  l ,  n ( l )  is a normalizing factor standingfor its detection efficiency and is obtained from calibrationprocedures;  a ( l )  is the attenuation factor and can be estimatedby a preliminary transmission scan of the object. In absenceof attenuation and imperfections in detectors, Pr ( y  =  l ) =1 /L , where  L  is the total number of possible LORs. 3. NONPARAMETRIC PRIOR FOR THE RANDOMACTIVITY DISTRIBUTION The formulation (1) requires to put a prior distribution G  overthe random activity distribution  G  of recorded events. Wechoose the prior on the density of   G  (denoted  f  G ), as being aDirichlet process mixture (DPM). We refer to [7] for more de-tails and [6] for a short introduction to this process and its ap-plication to 2D PET reconstruction. We briefly recall that thisis obtained by first generating a random discrete distributionfrom a Dirichlet process (DP) with parameters  α >  0  whichtunes the prior number of components and  G 0  a probabilitydistribution of components locations. Second, the realization H   is smoothed out with a parametric kernel. We choose thiskernel as being a multivariate normal distribution (with den-sity  f   N  , whose parameters are the mean  m  and the covari-ance matrix  Σ ), and  G 0  the Normal-Inverse Wishart model(  NIW  ρ,n 0 , µ 0 , Σ 0 ). H   ∼ DP   ( α,G 0 ) ⇒ H  ( · ) = ∞  k =1 w k δ  θ k ( · ) f  G  ( x ) =    f   N  ( x | θ ) H   ( d θ ) = ∞  k =1 w k  f   N   ( x | m k ,  Σ k ) (5)where 2011 18th IEEE International Conference on Image Processing 1402  •  the sequence of weights  w  = ( w 1 ,w 2 ,... )  is con-structed as follows: for all  i ,  V  i  ∼  Beta (1 ,α )  and w 1  =  V  1  and for all  k  ≥  2 ,  w k  =  V  k  k − 1 j =1  (1 − V  j ) .ThisiscalledGEMdistribution, written w ∼ GEM ( α ) , •  θ k  = ( m k ,  Σ k ) ∼ G 0 In the following section, we introduce our hierarchicalmodel for PET data. 4. THE BAYESIAN HIERARCHICAL MODEL Let  X  = ( x 1 ,..., x n )  be  n  emissions locations in the do-main of interest. In emission imaging systems, these loca-tions are never observed directly, only their projections y  =( y 1 ,...,y n )  are seen. Then our problem is an indirect den-sity estimation since observations are not distributed accord-ing to the distribution of interest  G  but instead according to F  , which is related to  G  via the integral in (1). We introduceemissions locations as hidden variables to complete the data.We also introduce allocation variables c  = ( c 1 , c 2 ,..., c n ) such that  c i  =  k  if   θ i  = ( m ,  Σ ) i  =  θ k . Then the jointdistribution of all variables in the model can be written as f  ( y , X , c , Θ , w ) = P  ( y | X ) × f  ( X | c , Θ ) × P  ( c | w ) × f  ( w ) × f  ( Θ )  (6)with distributions given by the following generative hierarchi-cal model y i | x i ind ∼ P  ( y i | x i ) x i | c i , Θ  ind ∼ N   ( x i | θ c i ) c i | w  iid ∼ ∞  k =1 w k δ  k ( · ) w ∼ GEM ( α ) θ k iid ∼NIW  ρ,n 0 , µ 0 , Σ 0 (7) 5. MCMC SAMPLING Having our hierarchical model (7), we can obtain samplesfrom the posterior distribution of the random activity  G ( · ) | y .For this purpose, we need to get draws from the posterior lawsof model variables. Since these posteriors have intractableforms, we opted for a MCMC sampling to generate thesedraws. The algorithm will update each block of variables inturn according to its conditional posterior as followsProposal of annihilations locations:  X | w , Θ , y Allocation of locations to components:  c | w , Θ , X Updating of the components parameters:  Θ | c , X Updating of the components weights:  w | c . (8)By choosing conjugate priors, these conditional posteriors aredirectlygeneratedbyaGibbssampler. However, thesamplingof  X | w , Θ , y  is more complex since the distribution has noknown form. We use the Metropolis-Hastings algorithm witha proposal distribution being the product of the Dirichlet pro-cess mixture and a normal distribution extended in each ob-served LOR direction. The result is a re-weighted mixture of Gaussian distributions in the LOR.The major problem in the sampling is to tackle the infi-nite number of dimensions that appears in (7), without anyhard truncation. We developed an algorithm belonging to theclass of slice sampling methods [8]. The general idea of theslice sampling strategy is to introduce latent auxiliary vari-ables that allow to work with a conditionally finite number of components at each iteration of the sampler. 6. SIMULATION RESULTS We applied our algorithm to PET simulated data and com-pared its performances to those obtained with a Bayesianvoxels-based algorithm using the maximum a posteriori ap-proach (MAP) with a Gibbs distribution as prior.We randomly generated decays from a 3D voxelized brainphantom by a Monte Carlo technique such that  n  = 10 7 events were recorded. We have not included scattered andrandom coincidences in the model and detection probabilitieswere assumed purely geometrical.We give now the implementation details for the testedmethods. Proposed method We choose the parameters of the DPM as follows:  α  = 500 , Σ 0  = 6 . 25  × I 3 , n 0  = 4 . We first computed  5000  iterationsfor the burn-in period followed by  10000  for computing func-tionals of the posterior spatial emission distribution. After theburn-in, each draw  ( X ( t ) , c ( t ) , w ( t ) , Θ ( t ) )  obtained at itera-tion  t  is used to construct a sample of the random probabilitydensity of recorded events, f  G ( t )  ( x ) ≈ K  ∗  k =1 w ( t ) k  f   N   x | θ ( t ) k   (9)where  K  ∗ is the number of components required by the slicesampler. Using equations (2), (3) and (4), the probability den-sity of all events is given by f  G † ( t )  ( x ) ∝  f  G ( t )  ( x )  Ll =1  n ( l ) a ( l ) f  geom ( x | y  =  l ) .  (10)We choose as estimator for the density the conditional expec-tation of   G † , E ( f  G † | y ) ≈  1 N  N   t =1 f  G † ( t ) with  N   being the number of iterations after the burn in. MAP method The Gibbs prior is used to express spatial constraints suchthat neighboring voxels must have the same intensity valuewhile allowing curt changes at tissues boundaries. The po-tential function used was the  log cosh  function suggested by[4]. The parameters of the Gibbs prior were chosen such thatto minimize the mean squared error (MSE) w.r.t to the brainphantom. We used a 5-voxels neighboorhood, the weights are 2011 18th IEEE International Conference on Image Processing 1403  a) b)c) d) Fig. 1 . Reconstruction results: a) 3D brain fantom; b) pro-posed estimator; c) MAP-EM estimator; d) proposed estima-tor standard deviation.the inverse of the L2-distance between the two voxels consid-ered. The algorithm used to maximize the posterior distribu-tion is a modified version of the Expectation Maximizationalgorithm. Results Results obtained by the two algorithms are depicted in[Fig.1]. Image a) is the 3D brain phantom used to generatethe data, image b) the conditional mean of our method andimage c) the MAP-EM reconstruction. Since any posteriorfunctional is available, we display in image d) the conditionalstandard deviation w.r.t. the phantom. Credible intervalsare also available for each region of interest (figures notshown). It is substantially notable on 3D curves [Fig.1] thatour method provides smooth isosurfaces while not removingboundaries and edges in the image. In contrast, the MAP-EMapproach exhibits very noisy images. This underlines that theproposed method is suitable to improve signal-to-noise ratioand image resolution. It is worth noting that in our results,the discretization of the spatial distribution is only for visu-alization and was chosen  a posteriori  to  256  ×  256  ×  128 corresponding to the phantom size. The same discretizationwas used for MAP-EM. 7. CONCLUSION We have presented an indirect-density estimation model forimage reconstruction in emission tomography. The inferencescheme was done through a specific MCMC method to dealwith the infiniteness of the distribution. The structure of thesampler allows its easy parallelization. Simulation results onPET data have provided good quality images in the context of a relatively low number of events ( 10  millions) and our recon-structions are by far better than those obtained by the MAP-EM method. However, we did not account for background ef-fects (randoms and scattered for example) in our simulations.Future works will be devoted to include these effects.The major drawback of our approach is its computationaldemands, although it was implemented on a parallel computer( ≈  one day on a high performance computer). To circum-vent the time due to the MCMC sampling, an alternative isthe variational Bayesian method which approximates the pos-terior distributions analytically [9]. Another interesting wayto speed up our algorithm can be achieved by implementing itonGraphicalprocessor units(GPU)hardware. Sincewework directly with functionals belonging to R 3 , this is appealing forthese architectures. 8. REFERENCES [1] F. Natterer,  The Mathematics of Computerized Tomog-raphy , Society for Industrial and Applied Mathematics,1986.[2] L.A. Shepp and Y. Vardi, “Maximum likelihood recon-struction in positron emission tomography,”  IEEE Trans. Med. Imag. , vol. 1, pp. 113–122, 1982.[3] K. Lange and R. Carson, “EM reconstruction algorithmfor emission and transmission tomography,”  J. Comp. As-sist. Tomo. , vol. 8, pp. 306–316, 1984.[4] P. J. Green, “Bayesian reconstructions from emission to-mography data using a modified EM algorithm,”  IEEE Trans. Med. Imaging , vol. 9, no. 1, pp. 84–93, 1990.[5] S. Geman and D. E. McClure, “Statistical methods fortomographic image reconstruction,” in  Proceedings of the 46th Session of the International Statiscal Institute ,Bulletin of the ISI, Ed., 1987, vol. 52, pp. 5–21.[6] ´E. Barat, C. Comtat, T. Dautremer, T. Montagu, andR. Tr´ebossen, “A Bayesian nonparametric approach forPET reconstruction,” in  IEEE Med. Imaging Conf. Rec., NSS/MIC ’07. , 2007, vol. 6, pp. 4155–4162.[7] N.L. Hjort, C. Holmes, P. M¨uller, and S.G. Walker,  Bayesian Nonparametrics , Cambridge University Press,April 2010.[8] S.G. Walker, “Sampling the Dirichlet mixture model withslices,”  Comm. Statist. , vol. 36, pp. 45–54, 2007.[9] D.M. Blei and M.I. Jordan, “Variational inference fordirichlet process mixtures,”  Bayesian Analysis , vol. 1,no. 1, pp. 121–144, 2006. 2011 18th IEEE International Conference on Image Processing 1404
Search Related
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks