Parallelized formulation of the maximum likelihood-expectation maximization algorithm for fine-grain message-passing architectures

We investigated the spatial resolution and quantitative properties of SPECT images reconstructed using filteredbackprojection (FB) and maximum likelihood-expectation maximization (ML-EM) algorithms. We studied ML-EM as a function of iteration number and projectorhackprojector (php) model, and FE3 as a function of filter. To quantify the image resolution for a shift-variant process we defined a local point response function (PRF). To justify the use of the local PRF to quantify resolution for nonlinear ML-EM we tested the regime over which the algorithm behaved linearly with respect to input perturbations. Using Fourier domain analogs of the local PRF we demonstrated that ML-EM with an accurate php can lead to images of better resolution than FB with a Butterworth or Metz filter or ML-EM with a php that does not fully model the imaging process. We also showed that the resolution improves with increasing MLEM iteration number, and the extent of the improvement depends on the accuracy of the php model. We conclude that the local PRF can be used to characterize spatial resolution in SPECT images, and that ML-EM offers the possibility of superior resolution, particularly if the algorithm contains a good model of the imaging process.

Three-dimensional (3D) iterative maximum likelihood expectation maximization (ML-EM) algorithms for single-photon emission tomography (SPET) are capable of correcting image-degrading effects of non-uniform attenuation, distance-dependent camera response and patient shape-dependent scatter. However, the resulting improvements in quantitation, resolution and signal-to-noise ratio (SNR) are obtained at the cost of a huge computational burden. This paper presents a new acceleration method for ML-EM: dual matrix ordered subsets (DM-OS). DM-OS combines two acceleration methods: (a) different matrices for projection and back-projection and (b) ordered subsets of projections. DM-OS was compared with ML-EM on simulated data and on physical thorax phantom data, for both 180 degrees and 360 degrees orbits. Contrast, normalized standard deviation and mean squared error were calculated for the digital phantom experiment. DM-OS resulted in similar image quality to ML-EM, even for speed-up factors of 200 compared to ML-EM in the case of 120 projections. The thorax phantom data could be reconstructed 50 times faster (60 projections) using DM-OS with preservation of image quality. ML-EM and DM-OS with scatter compensation showed significant improvement of SNR compared to ML-EM without scatter compensation. Furthermore, inclusion of complex image formation models in the computer code is simplified in the case of DM-OS. It is thus shown that DM-OS is a fast and relatively simple algorithm for 3D iterative scatter compensation, with similar results to conventional ML-EM, for both 180 degrees and 360 degrees acquired data.

The maximum likelihood expectation maximisation (ML-EM) algorithm can be used to reduce reconstruction artefacts produced by filtered backprojection (FBP) methods in pinhole single-photon emission tomography (SPET). However, ML-EM suffers from noise propagation along iterations, which leads to quantitatively unpleasant reconstruction results. To avoid this increase in noise, the median root prior (MRP) algorithm for pinhole SPET was implemented. Projection data of a line source and Picker's thyroid phantom were collected using a single-head gamma camera with a pinhole collimator. MRP was added to existing pinhole ML-EM reconstruction algorithm and the phantom studies were reconstructed using MRP, ML-EM and FBP for comparison. Coefficients of variation, contrasts and full-widths at half-maximum were calculated and showed a clear reduction in noise without significant loss of resolution or decrease in contrast when MRP was applied. MRP also produced visually pleasing images even with high iteration numbers, free of the checkerboard-type noise patterns which are typical of ML-EM images.

Small animal and clinical SPECT cameras based on multi-pinhole detector arrangements typically boast spatial resolutions of 1–2mm and 4mm respectively. However, directional variability of the local imaging capability within the field-of-view (FOV) can impact our ability to detect abnormalities. We aimed to demonstrate the dependence of resolution on object position and orientation within the FOV in a multi-pinhole configuration using a resolution model validated by maximum likelihood expectation maximization (MLEM) reconstructions. The local directional resolution model was based on the smallest angle of view to all pinholes with respect to the normal to the resolution direction. The test configuration was two arcs of 16 pinholes covering 150° in 10° intervals connected by 2 pinholes at each end totaling 36 pinholes. Pinholes were 45mm from the Z-axis, the focal length was 132mm, and detectors were 215mm×215mm leading to a ∼28.5mm radius FOV. For validation, a simulated phantom of 3-disks (radius of 5.75mm, 0.75mm thick, and 1.5mm center-to-center spacing) was shifted to different positions in both the X- and Y-directions where differences in resolution were predicted by the theory. At each position the disks were oriented to align their normals with the X and Y-axes. Simulations using idealized pinholes were performed with noise-free forward projections without attenuation and scatter effects (256×256 pixels of 0.84mm×0.84mm in size) and forty iterations of MLEM reconstructions (128-cubed with 0.375mm-cubed voxel size). As predicted by the model, disks with the normal in the Y-direction were reconstructed with nearly uniform resolution at all positions along the X-axis of the FOV. The same was true for disks with the normal in the X-direction at various Y-positions. However, at many of those locations, the resolution behaved poorly when the disk phantom was rotated 90°. Overall, the analytical model and simulations reconstructed using MLEM demonstrated a strong dependence of resolution on position and orientation of an object within the FOV. The greatest resolution losses were observed for positions outside of the hemisphere enclosed by the pinhole array.

A statistical method called maximum likelihood expectation maximization (MLEM) is quite attractive, especially in PET/SPECT. However, the convergence rate of the iterative scheme of MLEM is quite slow. This study aims to develop and test a new method to speed up the convergence rate of the MLEM algorithm. We introduce a relaxation parameter in the conventional MLEM iterative formula and propose the relaxation strategy on the condition that the spectral radius of the derived iterative matrix from the iterative scheme with the accelerated parameter reaches a minimum value. Experiments with Shepp-Logan phantom and an annual tree image demonstrate that the new computational strategy effectively accelerates computation time while maintains reasonable image quality. The proposed new computational method involving the relaxation strategy has a faster convergence speed than the original method.

The iterative maximum-likelihood expectation-maximization (ML-EM) algorithm is an excellent algorithm for image reconstruction and usually provides better images than the filtered backprojection (FBP) algorithm. However, a windowed FBP algorithm can outperform the ML-EM in certain occasions, when the least-squared difference from the true image, that is, the least-squared error (LSE), is used as the comparison criterion. Computer simulations were carried out for the two algorithms. For a given data set the best reconstruction (compared to the true image) from each algorithm was first obtained, and the two reconstructions are compared. The stopping iteration number of the ML-EM algorithm and the parameters of the windowed FBP algorithm were determined, so that they produced an image that was closest to the true image. However, to use the LSE criterion to compare algorithms, one must know the true image. How to select the optimal parameters when the true image is unknown is a practical open problem. For noisy Poisson projections, computer simulation results indicate that the ML-EM images are better than the regular FBP images, and the windowed FBP algorithm images are better than the ML-EM images. For the noiseless projections, the FBP algorithms outperform the ML-EM algorithm. The computer simulations reveal that the windowed FBP algorithm can provide a reconstruction that is closer to the true image than the ML-EM algorithm.

Reasonable trade-off between the ASIC performance and GPP flexibility is the main objective of reconfigurable computing systems. Dynamic Reconfigurable computing platform using embedded just-in-time (JIT) compilation is the most flexible platform among others. All complex computing kernels can be translated to bitstream to be executed on FPGA using an embedded processor of dedicated specialized hardware. The main challenge in these systems is FPGA design automation time. Executing the CAD algorithms on embedded processor is too time-consuming and normally is not feasible for real applications. Placement is the most computative part of CAD algorithm. Therefore, a new FPGA placement algorithm is proposed in this paper which makes reasonable trade-off between execution time and quality of placement. The proposed algorithm includes two stages: force-directed placement and simulated annealing placement. The proposed algorithm is very low execution time to be useful in JIT compilation without considerable degradation on the quality of placement. Experimental results show $$2.33\times $$2.33A— speedup in execution time in cost of 3Â % overhead in channel tracks numbers.

MLEM reconstruction method applied to partial defect verification using simulated data

The purpose of this study is to compare and evaluate the performance of a multiple-scattering Compton imager (MSCI) to measure prompt gamma-rays emitted during proton therapy. Because prompt gamma-rays are generated simultaneously during the proton beam delivery, the falloff position of the Bragg peak of the proton beam can be determined from the distribution of prompt gamma-rays. The detection system was designed using three CdZnTe detector layers that can track radiation of unknown energy on the basis of effective Compton scattering events. The simple back-projection, filtered back-projection, and maximum likelihood expectation maximization (MLEM) algorithms were applied for the reconstructed Compton images. The falloff positions of the Bragg peaks determined from individual MSCIs were compared with the theoretical values calculated using the Monte Carlo N-Particle eXtended code. Moreover, the performance of the MSCI was compared with that of a previously developed system based on a slit collimator gamma camera. In summary, the MSCI with the MLEM reconstruction algorithm was better than the other reconstruction methods in terms of the falloff position of the Bragg peak, the angular resolution, and the signal-to-noise ratio.

Emission tomography, including single-photon emission computed tomography (SPECT), requires image reconstruction from noisy and incomplete projection data. The maximum-likelihood expectation maximization (MLEM) algorithm is widely used due to its statistical foundation and non-negativity preservation, but it is highly sensitive to noise, particularly in low-count conditions. Although total variation (TV) regularization can reduce noise, it often oversmooths structural details and requires careful parameter tuning. We propose a Graph-Enhanced Expectation Maximization (GREM) algorithm that incorporates graph-based neighborhood information into an MLEM-type multiplicative reconstruction scheme. The method is motivated by a penalized formulation combining a Kullback–Leibler divergence term with a graph Laplacian regularization term, promoting local structural consistency while preserving edges. The resulting update retains the multiplicative structure of MLEM and preserves the non-negativity of the image estimates. Numerical experiments using synthetic phantoms under multiple noise levels, as well as clinical 99mTc-GSA liver SPECT data, demonstrate that GREM consistently outperforms conventional MLEM and TV-regularized MLEM in terms of PSNR and MS-SSIM. These results indicate that GREM provides an effective and practical approach for edge-preserving noise suppression in emission tomography without relying on external training data.

<para xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> Compton imaging is a gamma ray imaging technique that has many possible applications including homeland security and medical imaging. Using the Compton scattering formula the origin of a scattered <formula formulatype="inline"> <tex Notation="TeX">$\gamma$</tex></formula>-ray can be localized to a point on the surface of a cone using a minimum of two position and energy measurements. List-mode Maximum likelihood expectation maximization (MLEM) is an iterative statistical algorithm that makes successive approximations to the most probable source distribution that would have led to the observed data. Conventional MLEM is often performed in two dimensions, which still requires knowledge of the source-to-detector distance. We propose iteration over all three spatial dimensions in order to maximize the probability in three-dimensional space. This paper reports on the development of a three-dimensional MLEM (3DMLEM) algorithm to reconstruct images from a Compton imager in three dimensions for single, multiple and line sources. Several techniques for reducing convergence times are also discussed. </para>

The energy-imaging integrated deconvolution (EIID) algorithm is capable of deconvolving the source image at any specific energy, as well as the incident spectrum from any direction. This spectrum-image reconstruction method takes place in integrated spatial and energy domain using the maximum likelihood expectation maximization (MLEM) algorithm. This technique has been demonstrated for combined two-, three- and four-interaction Compton events from an array of 2 cm × 2 cm × 1.5 cm detectors in our previous work. By including Compton continuum events in the system model and using a sensitivity image to correct efficiency effects, the MLEM solution estimates the true incident gamma-ray spectrum with correct branching ratio. For each direction, we can then identify the source isotope based on the peak energy and relative peak area. After identifying the source isotope, the source activity can be estimated. From the reconstructed image and change of the relative peak areas from the branching ratio, the presence of shielding and the material and thickness of shielding can be estimated.

Tomographic absorption spectroscopy (TAS) is experiencing a surge of interest due to recent progress in laser technology and advanced imaging concepts such as nonlinear tomography and compressed sensing. Nevertheless, even though numerous algorithms have been adapted from other tomographic areas such as medical imaging and engineering process control, and applied in TAS applications, systematic comparison between those methods has not been investigated. In this work, we aim to test major inversion algorithms on both the so-called rank-deficient (RD) and discrete ill-posed (DIP) problems. Comparative studies were performed on extensive cases, and the results for three representative phantoms are reported here. Other important topics such as determination of control parameters and the semiconvergent behavior of iterative methods have also been studied. According to our study, Landweber outperformed other methods for two RD cases and is slightly inferior to the maximum likelihood expectation maximization (MLEM) algorithm for the third one. It has also been found that Landweber and MLEM feature better immunity to semiconvergence than the others; however, since MLEM is sensitive to an initial guess, it is less robust than Landweber. On the other hand, the truncated singular value decomposition (TSVD) method is recommended for DIP problems due to its superior performance as it can effectively suppress detrimental effects from noise, which will be amplified during the inversion procedure. It should be emphasized that even though this study was conducted under the context of TAS, we expect it to provide useful insights for other tomographic modalities.

Bayesian spiking neurons (BSNs) provide a probablisitic and intuitive interpretation of how spiking neurons could work and have been shown to be equivalent to leaky integrate-and-fire neurons under certain conditions [1]. The study of BSNs has been restricted mainly to small networks because online learning, which currently involves a maximum-likelihood-expectation-maximisation (ML-EM) approach [2,3], is quite slow. Here a new approach to estimating the parameters of Bayesian spiking neurons, referred to as fast learning (FL), is presented and compared to online ML-EM learning. Learning in a BSN is local to the neuron and involves estimation of the transition rate and observation rate parameters of an underlying implicit hidden Markov model (HMM), the hidden state of which the BSN output encodes [1]. Rather than estimating the parameters by maximizing the log-likelihood of the hidden states and the synaptic observations given the parameters as is done in ML-EM [2,3], the FL algorithm directly calculates statistics upon which the parameters depend. This is achieved by taking advantage of the relationship between the log-odds ratio of the hidden state computed by the BSN and the probability that the hidden state is ‘on’ given the past synaptic observations, . Online learning in a two BSN neuron hierarchy is explored, where the first neuron receives N=20 synapses driven by Poisson processes and the second neuron receives input from only the first neuron. Simulations were performed for a fixed set of transition rates and observation rates for 10 different perturbations of the initial transition and observation rate estimates: ±0-20%, ±20-40%,..., ±180-200%. Initial rates were not allowed to go below 10-6. Each perturbation condition was simulated 100 times by randomly selecting the initial parameter values. Although the FL algorithm is not as exact as the ML-EM at estimating the true parameter values for small perturbations of the initial rate estimates relative to the true rates, the FL algorithm is able to reliably estimate the parameters for initial perturbations of up to 200%, whereas the ML-EM algorithm estimates begin to deviate after perturbations of about 40-60%. Moreover, the simplicity of the FL algorithm means that it runs on the order of 25 times faster than the ML-EM implementation considered. These results hold true for both the first and second neurons in the two BSN neuron hierarchy. For the first neuron in the hierarchy, the RMS difference in the time series of the probability calculated for the estimated and the true parameter values, follows a similar pattern to the parameter estimates when the FL and ML-EM algorithms are compared, with average RMS errors of 0.2% obtained for the FL algorithm across the range of perturbations studied. Although we do not have a formal proof of convergence of the FL algorithm, we conclude that the FL algorithm can stably estimate the parameters over a large range of initial perturbations and it can do this very quickly. Thus the FL algorithm makes online learning in networks of BSNs much more tractable.

This work explores the feasibility and impact of including both the motion correction and the image registration transformation parameters from positron emission tomography (PET) image space to magnetic resonance (MR), or stereotaxic, image space within the system matrix of PET image reconstruction. This approach is motivated by the fields of neuroscience and psychiatry, where PET is used to investigate differences in activation patterns between different groups of participants, requiring all images to be registered to a common spatial atlas. Currently, image registration is performed after image reconstruction which introduces interpolation effects into the final image. Furthermore, motion correction (also requiring registration) introduces a further level of interpolation, and the overall result of these operations can lead to resolution degradation and possibly artifacts. It is important to note that performing such operations on a post-reconstruction basis means, strictly speaking, that the final images are not ones which maximize the desired objective function (e.g. maximum likelihood (ML), or maximum a posteriori reconstruction (MAP)). To correctly seek parameter estimates in the desired spatial atlas which are in accordance with the chosen reconstruction objective function, it is necessary to include the transformation parameters for both motion correction and registration within the system modeling stage of image reconstruction. Such an approach not only respects the statistically chosen objective function (e.g. ML or MAP), but furthermore should serve to reduce the interpolation effects. To evaluate the proposed method, this work investigates registration (including motion correction) using 2D and 3D simulations based on the high resolution research tomograph (HRRT) PET scanner geometry, with and without resolution modeling, using the ML expectation maximization (MLEM) reconstruction algorithm. The quality of reconstruction was assessed using bias-variance and root mean squared error analyses, comparing the proposed method to conventional post-reconstruction registration methods. An overall reduction in bias (for a cold region: from 41% down to 31% (2D) and 97% down to 65% (3D), and for a hot region: from 11% down to 8% (2D) and from 16% down to 14% (3D)) and in root mean squared error analyses (for a cold region: from 43% to 37% (2D) and from 97% to 65% (3D), and for a hot region: from 11% to 9% (2D) and from 16% down to 14% (3D)) in reconstructed regional mean activities (full regions of interest; all with statistical significance: p < 5 × 10−10) is found when including the motion correction and registration in the system matrix of the MLEM reconstruction, with resolution modeling. However, this improvement in performance comes with an extra computational cost of about 40 min. In this context, this work constitutes an important step toward the goal of estimating parameters of interest directly from the raw Poisson-distributed PET data, and hence toward the complete elimination of post-processing steps.