Jan De Beenhouwer

Fast and memory‐efficient Monte Carlo‐based image reconstruction for whole‐body PET

PURPOSEnSeveral studies have shown the benefit of an accurate system modeling using Monte Carlo techniques. For state-of-the-art whole-body positron emission tomography (PET) scanners, Monte Carlo-based image reconstruction is associated with a significant computational cost to calculate the system matrix as well as a large memory capacity to store it. In this article, the authors present a simulation-reconstruction framework to solve these problems on the Philips Gemini GS PET scanner.nnnMETHODSnA fast, realistic system matrix simulation module was developed using egs_pet, which is an efficient PET simulation code based on EGSnrc. The generated system matrix was then used in a rotator-based ordered subset expectation maximization (OS-EM) algorithm, which exploits the rotational symmetry of a cylindrical PET scanner. The system matrix was further compressed by using sparse storage techniques.nnnRESULTSnThe system matrix simulation took five days on 50 cores of Xeon 2.66 GHz, resulting in a system matrix of 2.01 GB. The entire system matrix could be stored in the main memory of a standard personal computer. The image quality in terms of contrast-noise trade-offs was considerably improved compared to a standard OS-EM algorithm. The image quality was also compared to the clinical software on the scanner using routine parameter settings. The contrast recovery coefficient of small hot spheres and cold spheres was significantly improved.nnnCONCLUSIONSnThe results indicated that the proposed framework could be used for this PET scanner with improved image quality. This method could also be applied to other state-of-the-art whole-body PET scanners and preclinical PET scanners with a similar shape.

Acceleration of GATE SPECT simulations.

GEANT4 application for tomographic emission (GATE) is a geometry and tracking 4 (GEANT4) application toolkit for accurate simulation of positron emission tomography (PET) and single photon emission computed tomography (SPECT) scanners. GATE simulations with realistic count levels are very CPU-intensive as they take up to several days with single-CPU computers. Therefore, we implemented both standard (FD) and convolution based forced detection (CFD) with multiple projection sampling (MPS) which allows the simulation of all projections simultaneously in GATE. In addition, a FD and CFD specialized Geant4 navigator was developed to overcome the detailed but slow tracking algorithms in Geant4. This paper is focussed on the implementation and validation of these aforementioned developments. The results show a good agreement between the FD and CFD versus analog GATE simulations. These combined developments accelerate GATE from three to six orders of magnitude and render realistic simulations feasible within clinically acceptable simulation times.

Fast simulation of yttrium-90 bremsstrahlung photons with GATE

PURPOSEnMultiple investigators have recently reported the use of yttrium-90 (Y90) bremsstrahlung single photon emission computed tomography (SPECT) imaging for the dosimetry of targeted radionuclide therapies. Because Monte Carlo (MC) simulations are useful for studying SPECT imaging, this study investigates the MC simulation of Y90 bremsstrahlung photons in SPECT. To overcome the computationally expensive simulation of electrons, the authors propose a fast way to simulate the emission of Y90 bremsstrahlung photons based on prerecorded bremsstrahlung photon probability density functions (PDFs).nnnMETHODSnThe accuracy of bremsstrahlung photon simulation is evaluated in two steps. First, the validity of the fast bremsstrahlung photon generator is checked. To that end, fast and analog simulations of photons emitted from a Y90 point source in a water phantom are compared. The same setup is then used to verify the accuracy of the bremsstrahlung photon simulations, comparing the results obtained with PDFs generated from both simulated and measured data to measurements. In both cases, the energy spectra and point spread functions of the photons detected in a scintillation camera are used.nnnRESULTSnResults show that the fast simulation method is responsible for a 5% overestimation of the low-energy fluence (below 75 keV) of the bremsstrahlung photons detected using a scintillation camera. The spatial distribution of the detected photons is, however, accurately reproduced with the fast method and a computational acceleration of ∼17-fold is achieved. When measured PDFs are used in the simulations, the simulated energy spectrum of photons emitted from a point source of Y90 in a water phantom and detected in a scintillation camera closely approximates the measured spectrum. The PSF of the photons imaged in the 50-300 keV energy window is also accurately estimated with a 12.4% underestimation of the full width at half maximum and 4.5% underestimation of the full width at tenth maximum.nnnCONCLUSIONSnDespite its limited accuracy, the fast bremsstrahlung photon generator is well suited for the simulation of bremsstrahlung photons emitted in large homogeneous organs, such as the liver, and detected in a scintillation camera. The computational acceleration makes it very useful for future investigations of Y90 bremsstrahlung SPECT imaging.

Cluster computing software for GATE simulations.

Geometry and tracking (GEANT4) is a Monte Carlo package designed for high energy physics experiments. It is used as the basis layer for Monte Carlo simulations of nuclear medicine acquisition systems in GEANT4 Application for Tomographic Emission (GATE). GATE allows the user to realistically model experiments using accurate physics models and time synchronization for detector movement through a script language contained in a macro file. The downside of this high accuracy is long computation time. This paper describes a platform independent computing approach for running GATE simulations on a cluster of computers in order to reduce the overall simulation time. Our software automatically creates fully resolved, nonparametrized macros accompanied with an on-the-fly generated cluster specific submit file used to launch the simulations. The scalability of GATE simulations on a cluster is investigated for two imaging modalities, positron emission tomography (PET) and single photon emission computed tomography (SPECT). Due to a higher sensitivity, PET simulations are characterized by relatively high data output rates that create rather large output files. SPECT simulations, on the other hand, have lower data output rates but require a long collimator setup time. Both of these characteristics hamper scalability as a function of the number of CPUs. The scalability of PET simulations is improved here by the development of a fast output merger. The scalability of SPECT simulations is improved by greatly reducing the collimator setup time. Accordingly, these two new developments result in higher scalability for both PET and SPECT simulations and reduce the computation time to more practical values.

Physics process level discrimination of detections for GATE: Assessment of contamination in SPECT and spurious activity in PET

The GEANT4 application for tomographic emission (GATE) is one of the most detailed Monte Carlo simulation tools for SPECT and PET. It allows for realistic phantoms, complex decay schemes, and a large variety of detector geometries. However, only a fraction of the information in each particle history is available for postprocessing. In order to extend the analysis capabilities of GATE, a flexible framework was developed. This framework allows all detected events to be subdivided according to their type: In PET, true coincidences from others, and in SPECT, geometrically collimated photons from others. The framework of the authors can be applied to any isotope, phantom, and detector geometry available in GATE. It is designed to enhance the usability of GATE for the study of contamination and for the investigation of the properties of current and future prototype detectors. The authors apply the framework to a case study of Bexxar, first assuming labeling with 124I, then with 131I. It is shown that with 124I PET, results with an optimized window improve upon those with the standard window but achieve less than half of the ideal improvement. Nevertheless, 124I PET shows improved resolution compared to 131I SPECT with triple-energy-window scatter correction.

Fast 3D iterative image reconstruction for SPECT with rotating slat collimators

As an alternative to the use of traditional parallel hole collimators, SPECT imaging can be performed using rotating slat collimators. While maintaining the spatial resolution, a gain in image quality could be expected from the higher photon collection efficiency of this type of collimator. However, the use of iterative methods to do fully three-dimensional (3D) reconstruction is computationally much more expensive and furthermore involves slow convergence compared to a classical SPECT reconstruction. It has been proposed to do 3D reconstruction by splitting the system matrix into two separate matrices, forcing the reconstruction to first estimate the sinograms from the rotating slat SPECT data before estimating the image. While alleviating the computational load by one order of magnitude, this split matrix approach would result in fast computation of the projections in an iterative algorithm, but does not solve the problem of slow convergence. There is thus a need for an algorithm which speeds up convergence while maintaining image quality for rotating slat collimated SPECT cameras. Therefore, we developed a reconstruction algorithm based on the split matrix approach which allows both a fast calculation of the forward and backward projection and a fast convergence. In this work, an algorithm of the maximum likelihood expectation maximization (MLEM) type, obtained from a split system matrix MLEM reconstruction, is proposed as a reconstruction method for rotating slat collimated SPECT data. Here, we compare this new algorithm to the conventional split system matrix MLEM method and to a gold standard fully 3D MLEM reconstruction algorithm on the basis of computational load, convergence and contrast-to-noise. Furthermore, ordered subsets expectation maximization (OSEM) implementations of these three algorithms are compared. Calculation of computational load and convergence for the different algorithms shows a speedup for the new method of 38 and 426 compared to the split matrix MLEM approach and the fully 3D MLEM respectively and a speedup of 16 and 21 compared to the split matrix OSEM and the fully 3D OSEM respectively. A contrast-to-noise study based on simulated data shows that our new approach has comparable accuracy as the fully 3D reconstruction method. The algorithm developed in this study allows iterative image reconstruction of rotating slat collimated SPECT data with equal image quality in a comparable amount of computation time as a classical SPECT reconstruction.

Fan beam forced detection in Gate

Fan beam collimators can obtain a higher sensitivity without loss in resolution at the cost of a reduced field of view. The geometric response has been studied both analytically and with numerical and Monte Carlo simulations, but a fast and accurate Monte Carlo simulator for fan beam geometry is not available. The goal of this work is therefore to accelerate 99mTc fan beam simulations in Gate, with full MC modeling of the collimator and detector in order to retain the characteristic hexagonal hole pattern of the collimator. To this end, two problems need to be solved: the long calculation time of particle transport in fan beam collimator geometry and the lack of dedicated variance reduction techniques. The first problem is solved by a dedicated tracking algorithm for fan beam collimator. The second problem is solved by the introduction of fan beam forced detection with variable solid angles. Our methods were validated with both analog Gate simulations and measurements. A good agreement was found for the hexagonal hole pattern, energy spectra, spatial resolution and sensitivity.

Fast GATE fan beam SPECT projector

Fan beam collimation offers a higher sensitivity than parallel beam collimation at the cost of a reduced field of view. The spatially varying collimator-detector response has so far been studied analytically and with Monte Carlo simulations in order to improve reconstruction quality. Similarly to parallel beam collimation, it may be possible to improve the reconstruction quality further by using accurate Monte Carlo-based scatter estimates. In this work, we have developed a fast hybrid CPU-GPU accelerated Monte Carlo projector for fan beam collimation that can be used for this purpose. A pre-simulated collimator-detector response was combined with a parallelized GEANT4 implementation and GPU accelerated convolution-based forced detection. Our method was validated with a GATE model of the Prism 3000XP equiped with low energy high resolution fan beam collimators. Projections of an image quality phantom showed an excellent agreement with GATE. Almost noiseless projections can be obtained in under 30 seconds.

Fast GATE multi-pinhole SPECT simulations

Multi-pinhole collimation is increasinly being used in SPECT imaging. A wide variety of geometric designs has recently been introduced, which can easily be modeled with the Monte Carlo simulator GATE. However multi-pinhole simxadulations are still very inefficient with GATE due to the lack of a dedicated variance reduction technique. In this work, we introduced a pinhole forced detection method which allows for fast 99mTc GATE multi-pinhole simulations. An excellent agreement with analog GATE simulations was found in terms of spatial resolution, energy spectra, sensitivity and collimator penetration. As the collimator-detector response is modeled with full Monte Carlo — without making any assumptions on the detector configuration — our method can be of use in multi-pinhole collimator design, development of compensation methods and image reconstruction.

Scatter effects of MR components in PET-MR inserts

System design research for upcoming PET-MR scanners has mainly focussed on the effect of the high magnetic field on PET performance and on the influence of the PET scanner inside the MR bore on MR image quality. However, the presence of MR components close to the PET detectors could also have an influence on PET performance. We have investigated these effects in a simulation study of the preclinical PET-MR insert and of the proposed integrated whole-body system of the HYPERimage project. Simulations were performed with the ProcessGATE extension of the GATE simulation framework, which makes it possible to determine the fractions of total scatter caused by different components. The preclinical insert was simulated inside a clinical MR scanner. All components of the clinical system and the preclinical insert were modeled in realistic dimensions and materials. The PET detector consisted of 10 detector blocks on a 100 mm radius cylinder, each containing a 44 (tangential) by 72 (axial) array of LYSO crystals. The crystal dimensions were 1.3 × 1.3 × 10 mm. The energy window was set to 250 - 750 keV. The integrated whole-body system was modeled as the same clinical MR system with a split gradient coil and PET detector blocks between both parts of the split gradient coil. The PET detector blocks in the whole-body system consisted of 22 detector blocks on a 35 cm radius cylinder containing 22 (tangential) by 43 (axial) LYSO crystals. The crystal dimensions were 4 × 4 × 22 mm. The energy window in this configuration was 410 - 700 keV. A uniform cylinder (radius 5 mm, length 100 mm) filled with 1 MBq of 18F was simulated in both the preclinical insert and the whole-body system. The simulated time was 1s yielding one million simulated decays. In the preclinical insert only 47 % of detected singles were unscattered. The clinical system and precinical insert accounted for respectively 38 % and 15 % of scattered photons. On the coincidences level the influence of the clinical system was much smaller (17 %), while the scatter effect of the insert increased (20%). In the clinical system the gradient coils scatter the largest fraction of photons (58 %). In the insert over 65 % of scatter is caused by the table and the RF screen. In the integrated whole-body system 44 % of detected singles were scattered. At coincidence level this fraction was reduced to 34 %. The largest amount of scattered coincidences is caused by the RF screen. In conclusion, it is clear that putting MR components within or close to the FOV of a PET scanner can cause significant scatter. The scattering effect of the MR components should be taken into account in the design phase.