GPU optimization techniques to accelerate optiGAN-a particle simulation GAN.

The demand for specialized hardware to train AI models has increased in tandem with the increase in the model complexity over the recent years. Graphics processing unit (GPU) is one such hardware that is capable of parallelizing operations performed on a large chunk of data. Companies like Nvidia, AMD, and Google have been constantly scaling-up the hardware performance as fast as they can. Nevertheless, there is still a gap between the required processing power and processing capacity of the hardware. To increase the hardware utilization, the software has to be optimized too. In this paper, we present some general GPU optimization techniques we used to efficiently train the optiGAN model, a Generative Adversarial Network that is capable of generating multidimensional probability distributions of optical photons at the photodetector face in radiation detectors, on an 8GB Nvidia Quadro RTX 4000 GPU. We analyze and compare the performances of all the optimizations based on the execution time and the memory consumed using the Nvidia Nsight Systems profiler tool. The optimizations gave approximately a 4.5x increase in the runtime performance when compared to a naive training on the GPU, without compromising the model performance. Finally we discuss optiGANs future work and how we are planning to scale the model on GPUs.

Potential of Depth-of-Interaction-Based Detection Time Correction in Cherenkov Emitter Crystals for TOF-PET.

Cherenkov light can improve the timing resolution of Positron Emission Tomography (PET) radiation detectors, thanks to its prompt emission. Coincidence time resolutions (CTR) of ~30 ps were recently reported when using 3.2 mm-thick Cherenkov emitters. However, sufficient detection efficiency requires thicker crystals, causing the timing resolution to be degraded by the optical propagation inside the crystal. We report on depth-of-interaction (DOI) correction to mitigate the time-jitter due to the photon time spread in Cherenkov-based radiation detectors. We simulated the Cherenkov and scintillation light generation and propagation in 3 × 3 mm2 lead fluoride, lutetium oxyorthosilicate, bismuth germanate, thallium chloride, and thallium bromide. Crystal thicknesses varied from 9 to 18 mm with a 3-mm step. A DOI-based time correction showed a 2-to-2.5-fold reduction of the photon time spread across all materials and thicknesses. Results showed that highly refractive crystals, though producing more Cherenkov photons, were limited by an experimentally obtained high-cutoff wavelength and refractive index, restricting the propagation and extraction of Cherenkov photons mainly emitted at shorter wavelengths. Correcting the detection time using DOI information shows a high potential to mitigate the photon time spread. These simulations highlight the complexity of Cherenkov-based detectors and the competing factors in improving timing resolution.

A mobile high-resolution gamma camera for therapeutic dose control during radionuclide therapy.

Objective.Molecular radiotherapy is the most used treatment modality against malign and benign diseases of thyroid. In that context, the large heterogeneity of therapeutic doses in patients and the range of effects observed show that individualized dosimetry is essential for optimizing treatments according to the targeted clinical outcome.Approach.We developed a high-resolution mobile gamma camera specifically designed to improve the quantitative assessment of the distribution and biokinetics of131I at patients's bedside after treatment of thyroid diseases. The first prototype has a field of view of 5 × 5 cm2and consists of a high-energy parallel-hole collimator made of 3D-printed tungsten, coupled to a 6 mm thick CeBr3scintillator readout by an array of silicon photomultiplier detectors. The intrinsic and overall imaging performance of the camera was evaluated with133Ba and131I sources. In order to test its quantification capability in realistic clinical conditions, two different 3D-printed thyroid phantoms homogeneously filled with131I were used. Both single view and conjugate view approaches have been applied, with and without scatter correction technique.Main Results.The camera exhibits high imaging performance with an overall energy resolution of 7.68 ± 0.01%, a submillimetric intrinsic spatial resolution of 0.74 ± 0.28 mm and a very low spatial distortion 0.15 ± 0.10 mm. The complete calibration of the camera shows an overall spatial resolution of 3.14 ± 0.03 mm at a distance of 5 cm and a corresponding sensitivity of 1.23 ± 0.01 cps/MBq, which decreases with distance and slightly changes with source size due to the influence of scattering. Activity recovery factors better than 97% were found with the thyroid phantoms.Significance.These preliminary results are very encouraging for the use of our camera as a tool for accurate quantification of absorbed doses and currently motivates the development of a fully operational clinical camera with a 10 × 10 cm2field of view and improved imaging capabilities.

The Accuracy of Cerenkov Photons Simulation in Geant4/Gate Depends on the Parameterization of Primary Electron Propagation.

Energetic electrons traveling in a dispersive medium can produce Cerenkov radiation. Cerenkov photons' prompt emission, combined with their predominantly forward emission direction with respect to the parent electron, makes them extremely promising to improve radiation detector timing resolution. Triggering gamma detections based on Cerenkov photons to achieve superior timing resolution is challenging due to the low number of photons produced per interaction. Monte Carlo simulations are fundamental to understanding their behavior and optimizing their pathway to detection. Therefore, accurately modeling the electron propagation and Cerenkov photons emission is crucial for reliable simulation results. In this work, we investigated the physics characteristics of the primary electrons (velocity, energy) and those of all emitted Cerenkov photons (spatial and timing distributions) generated by 511 keV photoelectric interactions in a bismuth germanate crystal using simulations with Geant4/GATE. Geant4 uses a stepwise particle tracking approach, and users can limit the electron velocity change per step. Without limiting it (default Geant4 settings), an electron mean step length of ~250 µm was obtained, providing only macroscopic modeling of electron transport, with all Cerenkov photons emitted in the forward direction with respect to the incident gamma direction. Limiting the electron velocity change per step reduced the electron mean step length (~0.200 µm), leading to a microscopic approach to its transport which more accurately modeled the electron physical properties in BGO at 511 keV. The electron and Cerenkov photons rapidly lost directionality, affecting Cerenkov photons' transport and, ultimately, their detection. Results suggested that a deep understanding of low energy physics is crucial to perform accurate optical Monte Carlo simulations and ultimately use them in TOF PET detectors.

Integration of polarization in the LUTDavis model for optical Monte Carlo simulation in radiation detectors.

OBJECTIVE: Cerenkov photons have distinctive features from scintillation photons. Among them is their polarization: their electric field is always perpendicular to the direction of propagation of light and parallel to the plane of incidence. Scintillation photons are instead considered unpolarized. APPROACH: This study aims at understanding and optimizing the reflectance of polarized Cerenkov photons for optical Monte Carlo simulation of scintillation detectors with Geant4/GATE. First, the Cerenkov emission spectrum and polarization were implemented in the previously developed look-up-table Davis model of crystal reflectance. Next, we modified Geant4/GATE source code to account for scintillation and Cerenkov photons LUTs simultaneously. Then, we performed optical Monte Carlo simulations in BGO using GATE to show the effect of Cerenkov features on the photons' momentum at the photodetector face, using two surface finishes, with and without reflector. MAIN RESULTS: In this work, we describe the new features added to the algorithm and GATE. We showed that Cerenkov characteristics affect their probability to be reflected/refracted and thus their travel path within a material. SIGNIFICANCE: We showed the importance of accounting for accurate Cerenkov photons reflectance while performing advanced optical Monte Carlo simulations.

Optimization of scintillator-reflector optical interfaces for the LUT Davis model.

PURPOSE: Designing and optimizing scintillator-based gamma detector using Monte Carlo simulation is of great importance in nuclear medicine and high energy physics. In scintillation detectors, understanding the light transport in the scintillator and the light collection by the photodetector plays a crucial role in achieving high performance. Thus, accurately modeling them is critical. METHODS: In previous works, we developed a model to compute crystal reflectance from the crystal 3D surface measurement and store it in look-up tables to be used in the Monte Carlo simulation software GATE. The relative light output comparison showed excellent agreement between simulations and experiments for both polished and rough surfaces in several configurations, that is, without and with reflector. However, when comparing them at the irradiation depth closest to the photodetector face, rough crystals with a reflector overestimated the predicted light output. Investigating the cause of this overestimation, we optimized the LUT algorithm to improve the reflectance computation accuracy, especially for rough surfaces. However, optical Monte Carlo simulations carried out with these newly generated LUTs still overestimate the light output. Based on previous observations, one probable cause is the erroneous assumption of perfect couplings between the reflector and crystal and between the crystal and photodetector, which likely results in an important overestimation of the light output compared to experimental values. In practice, several factors could degrade it. Here, we investigated possible suboptimal optical experimental configurations that could lead to a degraded light collection when using Teflon or ESR reflectors coupled to the crystal with air or grease. We generated look-up tables with a mixture of air and grease and showed the effect of three possible sources of light loss: the presence of a small gap between the crystal and the reflector edges close to the photodetector face, the infiltration of grease in the crystal-reflector coupling, and the presence of inhomogeneities in the photodetector-crystal interface. RESULTS: The strongest effect is linked to the presence of a small gap of grease between the edges of the reflector material and the crystal (light loss of 10%-12% for 0.2 mm gap). The optical grease infiltrating the crystal-reflector air coupling decreases the light output, depending on the infiltration's extent and the amount of grease infiltrated. Five percent of air in the crystal-photodetector coupling can cause a light output decrease of 2% to 4%. The individual and combined effect of these advanced models can explain the discrepancy of the relative light output obtained with ESR in simulations and experiments. With Teflon, the study indicates that the light output loss strongly depends on the reflectance deterioration caused by grease absorption. CONCLUSIONS: Our results indicate that when studying scintillation detector performance with different finishes, performing simulations in ideal coupling conditions can lead to light output overestimation. To perform an accurate light output comparison and ultimately have a reliable detector performance estimation, all potential sources of practical limitations must be carefully considered. To broadly enable high-fidelity modeling, we developed an interface for users to compute their own LUTs, using their surface, scintillator, and reflector characteristics.

Advanced Monte Carlo simulations of emission tomography imaging systems with GATE.

Built on top of the Geant4 toolkit, GATE is collaboratively developed for more than 15 years to design Monte Carlo simulations of nuclear-based imaging systems. It is, in particular, used by researchers and industrials to design, optimize, understand and create innovative emission tomography systems. In this paper, we reviewed the recent developments that have been proposed to simulate modern detectors and provide a comprehensive report on imaging systems that have been simulated and evaluated in GATE. Additionally, some methodological developments that are not specific for imaging but that can improve detector modeling and provide computation time gains, such as Variance Reduction Techniques and Artificial Intelligence integration, are described and discussed.

Technical Note: Standalone application to generate custom reflectance Look-Up Table for advanced optical Monte Carlo simulation in GATE/Geant4.

PURPOSE: The need for high-fidelity modeling of radiation detectors to perform reliable detector performance optimization using Monte Carlo simulations requires to accurately simulate the light transport in the scintillator and the light collection by the photodetector. In this work, we implement our well-validated crystal reflectance model computed from three-dimensional (3D) crystal surface measurement in a standalone open-source application to allow researchers to generate fully customized crystal reflectance look-up-tables (LUTs) to be used in optical Monte Carlo simulation. METHODS: The LUTDavisModel application can be installed in a few minutes on Windows, macOS, and Linux, using 26 MB of space. MATLAB Runtime is required and is automatically installed with the application. The core algorithm has been previously validated experimentally and implemented in GATE v8.0. The standalone is divided into five panels, each of which performing a specific task: generate LUTs from a combination of surface type, scintillator, and coupling medium available in the database (such as LSO or BGO) or custom; compute LUTs with the reflectors available and custom coupling thickness; create a mixture of coupling media to account for possible defects in the optical coupling; plot precomputed LUTs for visual comparison. Tooltips and errors/warnings facilitate the navigation. The reported computational times were obtained with an Intel Core i7 MacBook Pro. RESULTS: LUTs can be generated with computational time ranging from a few minutes to several hours depending on the selected surface, sampling, and computational power. A longer time is needed when using rough surfaces and thick coupling media (hundreds of µ m ) due to increased photon tracking. CONCLUSIONS: We developed a user-friendly standalone application to generate LUTs that can be used inside GATE Monte Carlo simulations. It can be easily downloaded, installed, and used. Future optimizations will expand the database, decrease the computational time through greater parallelization, and include the generation of LUTs to study Cerenkov photons transport.