We develop multiscale models and simulation tools for personalized medicine, with an emphasis on medical physics and nuclear medicine. Our models are informed with clinical images and experimental data obtained from our close collaborators. We create in silico models of radiation detectors to advance the technology of positron emission tomography (PET). Projects include new optical Monte Carlo simulation models, studying the physics of ultrafast detectors, optimization of detector design for time-of-flight PET. * Check out our optiGAN project!
*Our lab is a member of the openGATE collaboration and happy to report that GATE 10, the first python-based GATE, is now released and published!
We develop quantitative dosimetry for radiopharmaceutical therapy, with a special interest in liver radioembolization with yttrium-90 microspheres and lutetium-177 therapies. Our projects include digital twins for treatment planning, computational fluid dynamics to model the microsphere transport, yttrium-90 imaging, and multiscale dosimetry. We aim to improve the outcome of cancer patients through advanced treatment planning and monitoring.
Transarterial radioembolization is a radionuclide therapy based on the delivery of radioactive Y-90 microspheres to liver tumors. Accurate pretreatment dosimetry is necessary to determine the Y-90 activity to inject to optimize the dose to the tumor while sparing the rest of the liver. Combined with monitoring of the patient’s response to treatment, it could result in much greater increase in patient survival than what recent clinical trials have demonstrated.
We are developing a liver digital twin using computational fluid dynamics (CFD) simulation and Y-90 microsphere decay physics to estimate the dose distribution in the liver. Using multiscale modeling, we carry out CFD simulations for each patient’s anatomy, estimating the microsphere transport and the radiation dose distribution.
Our long-term goal is developing a tool to assist physicians with optimizing the quantity and injection site of Y-90 microspheres during radioembolization planning. Current research includes personalizing CFD simulations, developing strategies to speed up the computation, optimizing dose calculation using multiscale modeling.
Y-90 radioembolization delivers β-emitting microspheres to liver tumors through the hepatic arterial tree. How those microspheres distribute, and how much radiation reaches the tumor, is governed by patient-specific blood flow. Accurate hemodynamic modeling is therefore essential for treatment planning, but traditional CFD solvers are too computationally expensive for routine clinical use. This project develops deep learning surrogates to bridge that gap. A Physics-Informed Neural Network (PINN) embeds the governing fluid equations directly into training, enabling accurate flow prediction in complex patient-specific hepatic geometries with minimal CFD supervision. A complementary multi-geometry neural operator learns a generalizable mapping from vessel geometry to the full 3D flow field, allowing predictions on new patient anatomies without retraining. Together, these tools aim to enable fast, patient-specific hemodynamic prediction as a foundation for prospective Y-90 dose planning.
NK Panneerselvam, G Mummaneni, E Roncali
"Toward Digital Twins for Optimal Radioembolization", PET clinics 21 (1), 153-167
A. Taebi et al, "On the impact of injection distance to bifurcations on yttrium-90 distribution in liver cancer radioembolization", Journal of Vascular Interventional Radiology 33(6) 2022
A. Taebi et al, "Realistic boundary conditions in SimVascular through inlet catheter modeling", BMC Res Notes 2021 14(1), 215 (2021)
A. Taebi et al, “Multi-scale computational fluid dynamics modeling for personalized liver cancer radioembolization dosimetry”, Journal of Biomechanics (2020)
A. Taebi et al, Computational modeling of the liver arterial blood flow for microsphere therapy: Effect of boundary conditions. Bioengineering 7(3), 64 (2020)
A. Taebi et al., "Estimation of Yttrium-90 Distribution in Liver Radioembolization using Computational Fluid Dynamics and Deep Neural Networks", 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) 2020 P4974-77
E. Roncali et al., “Personalized Dosimetry for Liver Cancer Y-90 Radioembolization Using Computational Fluid Dynamics and Monte Carlo Simulation”, Annals of Biomedical Engineering, 48:1499–1510 2020
A. Taebi et al, Hepatic arterial tree segmentation: Towards patient-specific dosimetry for liver cancer radioembolization, JNM 60 (supp1), 122
E. Roncali et al., Personalized dosimetry for liver cancer radioembolization using fluid dynamics, JNM 58 (supp1), 603
NIH R21 CA237686 (NCI ITCR)
CCSG P30 (NCI P30CA093373)
Radiopharmaceutical therapy is a type of radiation therapy based on radiolabeled molecules or particles injected to target tumors with short-range radiation. If attached to a tumor marker, the radionuclide has the potential to reach disseminated tumors with limited radiotoxicity, which makes it attractive and has recently accelerated the development of therapeutic radiopharmaceuticals.
Radiopharmaceutical therapy uniquely sits between external beam radiation therapy and chemotherapy and is typically utilized with little consideration for the dosimetry. Our lab combines engineering principles and translational research to develop quantitative dosimetry methods personalized for each patient. We focus on two types of radionuclide therapies administered at UC Davis Health: yttrium-90 radioembolization and lutetium 177 based therapies.
NIH R21 CA237686 (NCI ITCR)
CCSG P30 (NCI P30CA093373)
In parallel to dose prediction, we are also developing quantitative imaging to measure the Y-90 microsphere distribution after treatment, using Positron Emission Tomography (PET). This translational research is carried out within the Department of Radiology at UC Davis Health including the new EXPLORER Molecular Imaging Center where we are evaluating the impact of high-sensitivity PET on Y-90 quantification using total-body PET.
Problems we are solving include optimizing the reconstruction of the low signal Y-90 PET images, developing quantitative comparison of the Y-90 PET dose distribution with CFDose estimates, incorporating anatomical information from other imaging modalities such as CT to segment tumors and improve quantification.
Recently, we focused on contrast-enhanced CT to predict the dose distribution in liver segments based on liver perfusion and, have demonstrated a strong correlation between perfused tumor volumes and lung shunt fraction, indicating contrast-enhanced CT could be used for radioembolization treatement planning.
We are using pre-treatment 4-phase liver CT with contrast-enhancement to estimate the dose and the lung shunt fraction for a given injection location. The goal of this work is to simplify the radioembolization workup by exploring alternative solutions to the 99mTc-MAA SPECT exam.
Mehadji B., Marx T., Carter A., Goldman R.E., Vu C.T."Contrast-enhanced CT as a non-invasive alternative for lung shunt fraction estimation in hepatic transarterial radioembolization", Radiology Advances Vol. 2(4) umaf025 2025
Mehadji B., Ruvalcaba C. A., Hernandez A. M., Abdelhafez Y. G., Goldman R. E., and Roncali E.
“Translating contrast enhanced computed tomography images to liver radioembolization dose distribution for more comprehensively indicating patients”, Physics in Medicine & Biology Vol. 69 Issue 16 2024
C. Ruvalcaba, A. Rajamani and E. Roncali
"Characteristics of catheter injection for predictive particle transport modeling in Y-90 transarterial radioembolization procedures", Bulletin of the American Physical Society 2023 Vol. 11/21/2023
G. C. Costa et al.,"Radioembolization Dosimetry with Total-Body Y-90 PET", The Journal of Nuclear Medicine (2021)
E. Roncali et al., "Overview of the First NRG Oncology–National Cancer Institute Workshop on Dosimetry of Systemic Radiopharmaceutical Therapy", The Journal of Nuclear Medicine (2021)
A. Taebi et al.,
"Realistic boundary conditions in SimVascular through inlet catheter modeling", BMC Res Notes 14, 215 (2021).
A. Taebi et al., "Multi-scale computational fluid dynamics modeling for personalized liver cancer radioembolization dosimetry", J. Biomech. Eng., 2020.
A. Taebi et al., "Computational modeling of the liver arterial blood flow for microsphere therapy: Effect of boundary conditions". Bioengineering 7, 2020.
E. Roncali et al., "Personalized dosimetry for liver cancer Y-90 radioembolization using computational fluid dynamics and monte carlo simulation", ABMES 2020
E. Roncali et al., “Comparison of Y-90 liver dose distribution predicted with fluid dynamics with Y-90 PET”, Journal of Nuclear Medicine 61 (supplement 1), 1308-1308
A. Taebi, C. T. Vu and E. Roncali, "Estimation of Yttrium-90 Distribution in Liver Radioembolization using Computational Fluid Dynamics and Deep Neural Networks," 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Montreal, QC, Canada, 2020, pp. 4974-4977.
NIH U01 CA289068 (NCI ITCR) 2023-2027
NIH R21 CA237686 (NCI ITCR) 2018-2021
CCSG P30 (NCI P30CA093373)
Optical Monte Carlo Simulation for Radiation Detector Design:
Designing and optimizing radiation detectors for medical imaging requires accurate simulation of light propagation through detector components. In scintillator-based detectors, a gamma-ray interaction produces thousands of optical photons, whose spatial and temporal distributions on the photodetector directly determine the detector's spatial, energy, and timing resolutions. Simulating this optical transport is therefore central to predicting and improving detector performance. Monte Carlo simulation tools available to the nuclear medicine community were inherited from high-energy physics and included only simplistic optical models. To address this, our lab developed the LUT Davis model, an optical simulation framework based on measured three-dimensional crystal surface profiles and ray tracing. Rather than relying on idealized surface parameterizations, the model stores realistic reflectance data in optical look-up tables (LUTs) derived from actual surface measurements. The LUT Davis model was validated in a series of benchtop experiments with PET detectors and integrated into GATE, one of the most widely used Monte Carlo simulation platforms in medical physics (with over 2,000 users worldwide), in 2017. We have also released a standalone application, LUTDavisModel, that allows users to generate customized LUTs for their own crystal and reflector configurations. The model continues to expand in scope and capability.
Our goal is to provide the community with free, open computational tools that allow scientists to independently design and test optical surfaces tailored to their detector application. Our lab is one of the 27 international partners of the openGATE collaboration, which develops and maintains GATE as open-source software and organizes workshops and training events. In November 2024, the collaboration released GATE 10, the first Python-based version of the platform. For more information, visit opengatecollaboration.org. We welcome students interested in contributing to this project; if you would like to learn more, please reach out directly.
Stephan Naunheim, PH.D. | snaunheim@ucdavis.edu
Clara Ansel, MSc | cansel@ucdavis.edu
D. Sarrut et al.,
GATE 10 Monte Carlo particle transport simulation: I. Development and new features, Phys. Med. Biol. 2026
N. Krah et al.,
GATE 10 Monte Carlo particle transport simulation: II. Architecture and innovations, Phys. Med. Biol. 2026
Trigila et al.,
Intercrystal Optical Crosstalk in Radiation Detectors: Monte Carlo Modeling and Experimental Validation, Phys. Med. Biol. 2024
D. Sarrut et al., "Advanced Monte Carlo simulations of emission tomography imaging systems with GATE", Phys. Med. Biol. 2021
C. Trigila and E. Roncali, "Integration of polarization in the LUTDavis model for optical Monte Carlo simulation in radiation detectors", Phys. Med. Biol. 2021
C. Trigila and E. Roncali, "Optimization of scintillator–reflector optical interfaces for the LUT Davis model", Med. Phhys. 2021
C. Trigila et al.., "Standalone application to generate custom reflectance Look-Up Table for advanced optical Monte Carlo simulation in GATE/Geant4", Med. Phys. 2021
M. Stockhoff et al., “Advanced optical simulation of scintillation detectors in GATE V8.0: first implementation of a reflectance model based on measured data,” Phys. Med. Biol., 2017. PMB Highlight 2017
E Roncali et al., "An integrated model of scintillator-reflector properties for advanced simulations of optical transport", Phys. Med. Biol., 2017
E Roncali et al., “Modelling the transport of optical photons in scintillation detectors for diagnostic and radiotherapy imaging”, Phys. Med. Biol. 2017. PMB Highlight 2017
E. Berg et al., “Optimizing light transport in scintillation crystals for time-of-flight PET: an experimental and optical Monte Carlo simulation study,” Biomed. Opt. Express, 2015
E. Roncali et al., “Predicting the timing properties of phosphor-coated scintillators using Monte Carlo light transport simulation,” Phys. Med. Biol., 2014
E. Roncali et al., “Simulation of light transport in scintillators based on 3D characterization of crystal surfaces,” Phys. Med. Biol., 2013.
NIH R03 EB025533 2015-2017
NIH R01EB027130 2019-2022
The Motivation:
In scintillator-based radiation detectors, every measurable signal (spatial, energy, and timing resolution) is shaped by how optical photons travel through the crystal and reach the photodetector. Understanding this optical transport is essential for the fundamental characterization of signal formation, for optimizing detector instrumentation, and for answering design questions in application-specific PET scanners such as dedicated brain or breast imaging systems. Monte Carlo simulation of optical photon transport using frameworks such as GATE/Geant4 is the standard approach, but tracking millions of individual photons through complex crystal geometries is extremely time-consuming. At the system level, where hundreds or thousands of crystals must be modeled, this becomes effectively prohibitive.
Figure 1: PET detector array consisting of 3x3 scintillation crystals. One inital gamma-photon interaction leads to the emission of thousands of optical photons (red lines) which are subsequently detected with photosensors (green). Image taken from Naunheim et al. (DOI: https://doi.org/10.48550/arXiv.2601.18780)
Artificial Intelligence Combined with Physics:
Replacing these simulations with a deep learning model is not simply a matter of reproducing the right-looking output distributions. A generative model must also respect the underlying physical laws governing optical transport, including conservation of energy, Fresnel reflection and transmission at interfaces, and the geometric constraints imposed by the crystal and reflector geometry. Purely data-driven models can achieve good distributional agreement on average while violating these physical constraints for individual events, which limits their reliability for detector design studies. This makes optiGAN a problem in physics-informed generative modeling: the network must learn not just the statistical output of the simulation, but the physics that produces it.
optiGAN:
optiGAN uses conditional Generative Adversarial Networks (GANs) to learn the mapping from gamma interaction parameters, including the three-dimensional emission point within the scintillation crystal, to the full multidimensional distribution of detected optical photons (spatial coordinates, timing, and wavelength). We have demonstrated the functionality of this approach both for single crystals and for crystal arrays, where inter-crystal optical crosstalk introduces additional complexity that the model must capture. In both settings, optiGAN shows excellent agreement with full Monte Carlo simulations, including for emission points outside the training dataset.
We welcome students interested in contributing to this project; if you would like to learn more, please reach out directly.
Stephan Naunheim, PH.D. | snaunheim@ucdavis.edu
Brandon Pardi, B.Sc. | bmpardi@ucdavis.edu
Trigila C., Mummaneni G., B Mehadji B., Pardi B., E Roncali"Towards large nuclear imaging system optical simulations with optiGAN, a generative adversarial network", Physics in Medicine & Biology 2024 Vol. 10(12) Pages 125002 DOI 10.1088/1361-6560/adde0c
Mummaneni G., Trigila, C Krah N, Sarrut D, E Roncali"optiGAN: a deep learning-based alternative to optical photon tracking in Python-based GATE (10+)", Physics in Medicine and Biology 2025 Vol. 70(13) Pages 135009 DOI 10.1088/1361-6560/ade2b5
Srikanth A. , Trigila C., and E. Roncali
"GPU Optimization Techniques to accelerate optiGAN - a particle simulation GAN", Machine Learning: Science and Technology 2024 Vol. 5 Pages 027001 DOI: 10.1088/2632-2153/ad51c9
Trigila C, Srikanth A, Roncali E.
"A generative adversarial network to speed up optical Monte Carlo simulations." Machine Learning: Science and Technology. 2023 Apr 12;4(2):025005.
Cerenkov emission in scintillators only takes ~10 ps, much faster than scintillation (10-300 ns). It is a promising alternative for timing triggering to achieve ultra-fast detectors with timing resolution of less than 100 ps, but very few photons are produced (10-20 per 511 keV photoelectric interaction). Cerenkov-based detectors must minimize the travel time in the crystal, maximize the photon collection, and employ photodetectors that generate a fast signal to trigger on these prompt photons (collaboration with Dr. Sun Il Kwon). We are applying our unique optical modeling tools to study these combined factors and have demonstrated that our simulations can explain complex timing patterns measured reported by several international research groups.
We extended the LUT Davis model to complex nanostructures at the interface of scintillation crystals and photodetectors, such as photonic crystals. Photonic crystals, due to their nanosize, cannot be modeled with geometric optics like conventional crystal-photodetector interfaces. We combined wave and geometrics optics to develop LUTs modeling BGO and LYSO interfaces with TiO2 based photonic crystals, which were used in GATE detector simulations to study light collection, timing resolution, and energy resolution.
These tools are critical to develop the next generation of ultrafast detectors. We are now pushing the frontiers of Cerenkov-based detectors by studying and optimizing the use of Cerenkov light in semi-conductor radiation detectors such as Thallium Bromide and Thallium Chloride (collaboration with Dr. Gerard Ariño-Estrada).
Xuzhi He et al.,
Experimental validation of Davis LUT module with photonic crystals for GATE simulations, Physics in Medicine & Biology, Volume 71, Number 12
G. Arino-Estrada et al.,
Current Status of Cherenkov-Based Gamma Detectors for TOF-PET and Proton Range Verification, IEEE Transactions on Radiation and Plasma Medical Sciences, 2025
X. He, C. Trigila and E. Roncali,
"Implementation of Photonic Crystals into Davis LUT Module for GATE simulation", IEEE Transactions on Radiation and Plasma Medical Sciences 2024
Trigila C., Kratochwil N., Mehadji B., Ariño-Estrada G., and Roncali E.
"Inter-Crystal Optical Crosstalk in Radiation Detectors: Monte Carlo Modeling and Experimental Validation”, IEEE Transactions on Radiation and Plasma Medical Sciences 2024C. Trigila et al., "The Accuracy of Cerenkov Photons Simulation in Geant4/Gate Depends on the Parameterization of Primary Electron Propagation", Frontiers in Physics 10 (2022)
G. Terragni et al., "Time Resolution Studies of Thallium Based Cherenkov Semiconductors" Frontiers in Physics 10 (2022)
G. Arino-Estrada et al.,"Study of Čerenkov Light Emission in the Semiconductors TlBr and TlCl for TOF-PET", IEEE TRPMS 2020
E. Roncali et al, “Cerenkov light transport in scintillation crystals explained: realistic simulation with GATE,” Biomed.Opt. Express, 2019
S. I. Kwon et al., “Dual-ended readout of bismuth germanate to improve timing resolution in time-of-flight PET,” Phys Med Biol, 2019
NIH R01 EB034475 2023-2027
NIH R01 EB027130 2019-2022
NIH R03 EB025533 2015-2017