Please note! All times in the online program are given in New York - America (GMT -04:00) times!

New York - America ()
Jan 26, 2022, 3:21:51 PM
Your time ()
Tokyo - Asia ()
Jan 27, 2022, 5:21:51 AM
Our exhibitors and sponsors – click on name to visit booth:

To search for a specific ID please enter the hash sign followed by the ID number (e.g. #123).

Computer Modeling I

Session chair: Clifford , J. (SNL, USA)
Shortcut: N-01
Date: Tuesday, 19 October, 2021, 9:15 AM - 11:15 AM
Room: NSS - 1
Session type: NSS Session


Click on an contribution to preview the abstract content.

9:15 AM N-01-01

Deep Image Prior (DIP) based Material Decomposition for Spectral CT (#90)

M. Nakazawa1, X. Zhimin1, T. Kobayashi1, K. Tanabe1, K. Kitamura1

1 Shimadzu Corporation, Technology Research Laboratory, Kyoto, Japan


In spectral (photon-counting) CT, photon statistics of each energy window is reduced by splitting an energy window into multiple narrow windows. Although regularized iterative image reconstruction such as total variation (TV) can reduce the statistical image noise, hyperparameters must be determined in advance depending on the imaging object. Also, TV reconstruction works well only for piece-wise uniform objects. Those might be barriers in adapting it to industrial CT equipment.

In this study, we adopted deep image prior (DIP) for noise reduction of parameter-free but noisy filtered-backprojection (FBP) images. The advantage of DIP is that the convolutional deep neural network structure itself provides noise reduction effect, and users does not need to set parameters in advance. In addition, it does not need the prior training dataset which is required for general deep learning processing.

We evaluated through simulation studies the decomposition performance for a low-contrast polymer composites phantom (minimum effective atomic number difference about 0.2). Material decomposition was performed by principal component analysis (PCA) of each energy image followed by k-means clustering.  We confirmed that FBP + DIP processing had the same decomposition performance compared to TV-based reconstruction. By using DIP, it is expected that spectral CT imaging will be possible for various objects like composite materials without tangled parameter turning.
Keywords: Spectral CT, photon counting CT, deep image prior, material decomposition
9:30 AM N-01-02

Geant4 Simulation for the ClearMind Project and Reconstruction of the Gamma Conversion (#135)

C. - H. Sung1, L. Cappellugola2, M. Follin1, M. Dupont2, C. Morel2, D. Yvon1, 3, V. Sharyy1, 3

1 Université Paris-Saclay, IRFU, CEA, Gif-sur-Yvette, France
2 Aix-Marseille Université, CNRS/IN2P3, CPPM, Marseille, France
3 Université Paris-Saclay, BioMAPs, Service Hospitalier Frédéric Joliot, CEA, CNRS, Inserm, Orsay, France


ClearMind project aims to develop the TOF PET detection module providing a short coincidence time resolution, good spatial resolution, and high detection efficiency. ClearMind project uses a monolithic PbWO4 scintillating crystal for the position-sensitive detector, and the bialkali photo-electric layer deposited on the crystal. The 511 keV gamma conversion produces both scintillating and Cherenkov photons. Photoelectrons generated at the photocathode are amplified by the micro-channel plate, and signals induced on the pixelized anode are collected through the transmission lines readout and digitized by the SAMPIC module. In this work, we present a realistic Geant4 simulation of such a module including the propagation of the optical photons in the crystal, realistic response of the photocathode, simulation of the PMT response, and propagation of the electrical signals over the transmission lines. The reconstruction of the gamma conversion in the detector volume is performed from the signal registered at both ends of transmission lines. We compare the reconstruction precision of the simple statistical algorithms using mean, standard deviation, with machine learning algorithms developed using the TVMA package. We estimate to reach a spatial resolution down to a few mm3 (FWHM) and a CTR of the order of 100 ps (FWHM).

Keywords: lead tungstate, monolithic crystal, MCP-PMT, event reconstruction, machine learning
9:45 AM N-01-03

A model for a linear a-Se detector in simulated x-ray breast imaging with Monte Carlo software (#211)

A. Sarno1, R. M. Tucciariello4, 3, M. E. Fantacci4, 3, A. C. Traino5, G. Mettivier2, 1, C. Valero6, M. Stasi6, P. Russo2, 1

1 INFN, Sez. di Napoli, Napoli, Italy
2 Università degli studi di Napoli "Federico II", Dip di Fisica "E Pancini", Napoli, Italy
3 INFN, Sez. di Pisa, Pisa, Italy
4 Università di Pisa, Pisa, Italy
5 Azienda Ospedaliero-Universitaria Pisana, Pisa, Italy
6 A.O. Ordine Mauriziano, Torino, Italy


As alternative to high-cost and long clinical trials on patient population for testing x-ray breast imaging techniques, the AGATA project proposes the use of in-silico clinical trials with digital patient models and simulated apparatuses. In the in-silico reproduction, the detector model assumes great importance, and its performance has to reflect that of the real one. In this work, we simulated an a-Se linear detector as an absorber of known thickness. The detector response curve (DRC), modulation transfer function (MTF) and normalized noise power spectrum (NNPS) were measured on a clinical mammographic unit. The same tests were replicated in-silico via the AGATA Monte Carlo software. In this case, coherent and incoherent scatter and photoelectric interactions as well as fluorescence in a-Se detector were simulated. The relation between the measured DRCs (pixel value vs air kerma) and the simulated ones (pixel dose vs air kerma) will permit to convert simulated pixel values to realistic values. The differences between simulated MTF (spatial frequency for MTF at 50% of its maximum, MTF50% =8.8±0.2 mm-1 @25 kV) and measured MTF (MTF50% =8.1±0.5 mm-1 @25 kV), also extended to several tube voltages, will permit to define linear filters for spatial resolution tuning in simulated projections. The comparison between simulated and measured NNPS will support in strategies for defining a suitable noise model, in particular for the estimates of neglected thermal noise and negleted noise due to electrons and electro-hole pairs tracking. These comparisons will be the fundamental for tuning the simulated detector noise and spatial resolution on the basis of the linear-space invariant system hypothesis for a fair in-silico clinical trials reproduction.

AcknowledgmentThis work is funded by the INFN (Italy) under the AGATA_GR5 project. Measurements were performed at the Azienda Ospedaliero Universitaria Pisana (Pisa, Italy) and at the Ospedale Mauriziano Umberto I (Torino, Italy).
Keywords: Digital breast tomosynthesis, Mammography, Monte Carlo, virtual clinical trials
10:00 AM N-01-04

Evaluation of Block sequential regularized expectation maximization (BSREM) Reconstruction Algorithm using relative difference penalties compared to OSEM on a PET/CT scanner: Phantom and Clinical study (#441)

F. Sadeghi1, 2, P. Sheykhzadeh3, S. Farzanehfar3, M. Ay1, 2

1 Tehran University of Medical Sciences, Department of Medical Physics and Biomedical Engineering, Tehran, Iran (Islamic Republic of)
2 Tehran University of Medical Sciences, Research Center for Molecular and Cellular Imaging, Tehran, Iran (Islamic Republic of)
3 Tehran University of Medical Sciences, Department of Nuclear Medicine, Imam Khomeini Hospital complex, Tehran, Iran (Islamic Republic of)


In this study, we compare image quality parameters extracted from BSREM newly reconstruction algorithm with various strength of noise penalization and conventional OSEM. NEMA IEC phantom and fifteen clinical whole-body F-FDG examinations were scanned on a Discovery IQ PET/CT Scanner. In phantom study: SUVmax and SUVmean increased with decreasing β-factor and sphere diameter so the lowest and highest level of SUV was reached with OSEM and β-value= 100 respectively. By increasing β-value from 100 to 500, noise decreases from 9.44 to 3.94, 7.4 to 3.05 and 9.63 to 4.52 in LBR of 2 and 4 and 8 respectively. Contrast decreased with increasing β factor but contrast for OSEM was significantly worse compare with BRSEM results. All BRSEM findings showed considerably less LE than OSEM. The contrast recovery increases when reducing the β-value and in all sphere size. β-value 500 provided the highest CNR and β-value 100 along with OSEM has lowest CNR. In clinical study, the lowest and highest level of SBR and SNR were reached with OSEM and β-value 100 respectively and β-value 400 was the highest ranked in most evaluated parameters.

In summary, for the phantom and clinical study, SUVmax of lesions as the most important parameter for diagnosing and staging diseases, in BSREM is more accurate than OSEM. BSREM provide superior quantitative evaluation parameter such as contrast, SNR, CNR by controlling image noise through regularized reconstruction.

AcknowledgmentThe authors would like to thank Shayan Monsef for him help with phantom acquisitions
Keywords: BSREM, OSEM, reconstruction algorithm, PET/CT, penalization factor
10:15 AM N-01-05

Experimental validation of a neural network for PET event positioning and timestamping in monolithic crystals (#476)

P. Carra1, 2, M. G. Bisogni1, 2, E. Ciarrocchi3, M. Morrocchi1, 2, V. Rosso1, 2, G. Sportelli1, 2, N. Belcari1, 2

1 Università di Pisa, Physics, Pisa, Italy
2 INFN, Pisa section, Pisa, Italy
3 Università di Pisa, Translational Research and New Surgical and Medical Technologies, Pisa, Italy


This contribtuion presents the experimental validation of a neural network for simultaneous PET event positioning and timestamping in monolithic scintillators coupled to SiPM matrices. The algorithm is tested on a 25 mm x 25 mm x 8 mm LYSO crystal coupled to an 8x8 matrix of last generation 3 mm Hamamatsu MPPCs. This work reports a description of the method and the experimental results obtained. We also observe that the inclusion of simulated data is essential for obtaining the best performance from the algorithm and for shortening the calibration procedure, which is one of the major obstacles for the adoption of monolithic crystals in full PET scanners. In our laboratory setup we achieve a spatial resolution below 1 mm full-width-at-half-maximum (FWHM) in the (x,y) plane and below 2 mm FWHM in the DOI coordinate. The coincidence time resolution (CTR) obtained is below 200 ps. The algorithm also has been implemented in a modern, low-cost FPGA that could fit in a detector, providing real-time data processing at a rate of approximately 2 Mcps.


The research leading to these results has received funding from the the European Union’s Horizon 2020 research and innovation programme under grant agreement No 688735 (Photonics Based Sensing ERA-NET Cofund), Regione Toscana and VLAIO.

Keywords: Monolithic scintillators, PET, Detectors, Neural networks, Time-of-flight
10:30 AM N-01-06

A data-driven fragmentation model for GPU fastMonte Carlo dose recalculation for carbon therapy (#356)

M. De Simoni1, 4, P. De Maria3, M. Fischetti2, 4, G. Franciosini1, 4, M. Marafini6, 4, V. Patera2, 4, A. Sarti2, 4, A. Sciubba2, 5, M. Toppi2, 5, G. Traini4, A. Trigilio1, 4, A. Schiavi2, 4

1 Sapienza, Physics department, Rome, Italy
2 Sapienza, Scienze di Base Applicate all'Ingegneria (SBAI) department, Rome, Italy
3 Sapienza, Scienze e biotechnologie medico-chirurgiche department, Rome, Italy
4 INFN, Section of Rome, Rome, Italy
5 INFN, Section of Frascati, Frascati, Italy
6 Museo Storico della Fisica e Centro Studi e Ricerche E. Fermi, Rome, Italy


The advent of general programming Graphics Processing Units (GPU) has prompted the development of MC codes that can dramatically reduce the plan recalculation time with respect to standard MC codes in CPU hardware. FRED (Fast paRticle thErapy Dose evaluator) is a software that exploits the GPU power to recalculate and optimize ion beam treatment plans. Rapidly recalculating a complete treatment plan within minutes, instead of hours, paves the way for many clinical applications where the time-factor is important. When developing the core algorithms, the goal is to balance accuracy in the energy range of particle therapy, calculation time and GPU execution guidelines.  For what concerns proton beams, FRED is already used as a quality assurance tool in the clinical center of Maastricht and Krakow and as a research tool at several clinical and research centers in Europe (Krakow, Trento, Maastricht, Lyon and PSI). Carbon ion, electron and photon beams have been introduced inside FRED to allow fast optimization of a treatment plan also in carbon ion therapy, photon radiotherapy and IORT (IntraOperative Radiation Therapy). Those implementations are under development and in this article the new data-driven tracking model of carbon ion will be described.  The carbon version of FRED is now in the phase of porting on GPU. A scaping from the proton version allows estimating that the tracing kernel, running on GPU hardware, can achieve an order of million primaries per second on a single card.

Keywords: Particle Therapy, Fast-MC, Carbon ions, Fragmentation, GPU
10:45 AM N-01-07

An artificial neural network capable of multi-radioisotope identification in various shielding scenarios using sodium iodide gamma spectra (#517)

L. Lee-Brewin1, D. Read2, C. Shenton-Taylor1

1 University of Surrey, Physics, Guildford, United Kingdom
2 University of Surrey, Chemistry, Guildford, United Kingdom


Machine learning algorithms designed to identify isotopes present within gamma spectra are often extremely sensitive to environmental conditions such as shielding between the source and detector. In this contribution, a machine learning algorithm capable of identifying multiple radioisotopes across a range of shielded environments is presented. Five isotopes were selected for a training set; small sets of single isotope spectra were collected using a sodium iodide detector in a variety of environments with different levels of shielding. Each of these small sets of spectra were augmented to create multi-isotope spectra and combined to form a training set. Seven testing sets were independently collected through experiment using sources of the same isotopes with different activities. Each testing set contained all possible combinations of the 5 isotopes and did not use any augmentation techniques. Three of the testing sets were collected using the same shielding environments as the training sets and 3 sets used new shielding environments the network had not been trained on. The final set was collected using a new shielding environment and a different sodium iodide detector. Results indicate a high detection accuracy with no significant loss in accuracy for multi-isotope detection or for spectra collected in shielded environments the algorithm had not directly been trained on.  The network was tested on spectra collected from a different sodium iodide detector showing no significant loss of accuracy.

AcknowledgmentThis work was supported by funding from the Nuclear Decommissioning Authority within the Engineering & Physical Science Research Council project EP/S01019X/1 (TRANSCEND). The authors wish to thank Jack Armitage of Magnox Ltd. and Dr. James Graham of the National Nuclear Laboratory for advice and support.
Keywords: artificial neural network, gamma spectroscopy, isotope identification, machine learning, nuclear decommissioning
11:00 AM N-01-08

Radkit: Python Packages for Radiological and Contextual Data Processing and Fusion (#1058)

D. Hellfeld1, J. C. Curtis1, M. Salathe1, M. S. Bandstra1, M. Folsom1, J. Vavrek1, A. Moran1, B. J. Quiter1, R. J. Cooper1, T. H. Y. Joshi1

1 Lawrence Berkeley National Laboratory, Nuclear Science Division, Berkeley, California, United States of America


The radkit software suite is a set of python packages that enable nuclear Scene Data Fusion (SDF) by utilizing data recorded from contextual sensors and fusing them with radiological data. The contextual data enable Simultaneous Localization and Mapping (SLAM), which provides accurate pose estimates and a 3D scene model. These are leveraged in various 3D gamma-ray imaging modalities to produce a visual representation of the unknown distribution of radioactive material in the vicinity of the system. The radkit suite consists of three primary packages: trajan, curie, and mfdf. The trajan package provides the tools to analyze and manipulate data obtained from contextual sensors and SLAM algorithms. The curie package comprises tools for radiation data and analyses such as manipulation, plotting and slicing of listmode and binmode data, radiological source detection and identification data structures, correlations between source encounters, and definitions of energy-dependent angular detector response functions. The mfdf package is capable of fusing data from trajan and curie, and implements 3D gamma-ray image reconstruction algorithms. The tools are flexible for offline analyses as well as online deployments with associated Robot Operating System (ROS) packages.


This work has been supported by the US Department of Homeland Security, Countering Weapons of Mass Destruction office, under competitively awarded contract/IAA 70RDND18K00000005 and by the Defense Threat Reduction Agency under HDTRA 10027-30529 and 13081-36239. This support does not constitute an express or implied endorsement on the part of the Government.

Keywords: gamma-ray imaging, software, python, data fusion

Our exhibitors and sponsors – click on name to visit booth: