Small beams, fast predictions: a comparison of machine learning dose prediction models for proton minibeam therapy.

BACKGROUND: Dose calculations for novel radiotherapy cancer treatments such as proton minibeam radiation therapy is often done using full Monte Carlo (MC) simulations. As MC simulations can be very time consuming for this kind of application, deep learning models have been considered to accelerate dose estimation in cancer patients. PURPOSE: This work systematically evaluates the dose prediction accuracy, speed and generalization performance of three selected state-of-the-art deep learning models for dose prediction applied to the proton minibeam therapy. The strengths and weaknesses of those models are thoroughly investigated, helping other researchers to decide on a viable algorithm for their own application. METHODS: The following recently published models are compared: first, a 3D U-Net model trained as a regression network, second, a 3D U-Net trained as a generator of a generative adversarial network (GAN) and third, a dose transformer model which interprets the dose prediction as a sequence translation task. These models are trained to emulate the result of MC simulations. The dose depositions of a proton minibeam with a diameter of 800µm and an energy of 20-100 MeV inside a simple head phantom calculated by full Geant4 MC simulations are used as a case study for this comparison. The spatial resolution is 0.5 mm. Special attention is put on the evaluation of the generalization performance of the investigated models. RESULTS: Dose predictions with all models are produced in the order of a second on a GPU, the 3D U-Net models being fastest with an average of 130 ms. An investigated 3D U-Net regression model is found to show the strongest performance with overall 61.0 % ± $\%\pm$ 0.5% of all voxels exhibiting a deviation in energy deposition prediction of less than 3% compared to full MC simulations with no spatial deviation allowed. The 3D U-Net models are observed to show better generalization performance for target geometry variations, while the transformer-based model shows better generalization with regard to the proton energy. CONCLUSIONS: This paper reveals that (1) all studied deep learning models are significantly faster than non-machine learning approaches predicting the dose in the order of seconds compared to hours for MC, (2) all models provide reasonable accuracy, and (3) the regression-trained 3D U-Net provides the most accurate predictions.

Characterisation of a well-type NaI(Tl) detector by means of a Monte Carlo simulation for radionuclide metrology application.

A well-type NaI(Tl) detector was modelled and characterised by means of a Monte Carlo simulation, as part of a project to develop a 4πß (Plastic Scintillator)-4πγ instrument to be used for the primary standardisation of radionuclides at ANSTO. The simulation based on GEANT4 was used to characterise the 4πγ detector in terms of potential dead layer/inactive materials, full energy peak efficiency, coincidence-summing correction, and energy resolution. An excellent agreement was obtained between the simulation results and the experimental measurements.

The dose magnifying glass quality assurance system for daily proton therapy range verification.

Proton therapy has a distinct dosimetric advantage over conventional photon therapy due to its Bragg peak profile. This allows greater accuracy in dose delivery and dose conformation to the target, however it requires greater precision in setup, delivery and for quality assurance (QA) procedures. The AAPM TG 224 report recommends daily range and spot position checks with tolerance on the order of a millimetre. Daily QA systems must therefore be efficient for daily use and be capable of sub-millimetre precision however few suitable commercial systems are available. In this work, a compact, real-time daily QA system is optimised and characterised for proton range verification using an ad-hoc Geant4 simulation. The system is comprised of a monolithic silicon diode array detector embedded in a perspex phantom. The detector is orientated at an angular offset to the incident proton beam to allow range in perspex to be determined for flat proton fields. The accuracy of the system for proton range in perspex measurements was experimentally evaluated over the full range of clinical proton energies. The meanR100,R90andR80ranges measured with the system were accurate within ±0.6 mm of simulated ranges in a perspex phantom for all energies assessed. This system allows real-time read-out of individual detector channels also making it appropriate for temporal beam delivery diagnostics and for spot position monitoring along one axis. The system presented provides a suitable, economical and efficient alternative for daily QA in proton therapy.

Study of the X-ray radiation interaction with a multislit collimator for the creation of microbeams in radiation therapy.

Microbeam radiation therapy (MRT) is a developing radiotherapy, based on the use of beams only a few tens of micrometres wide, generated by synchrotron X-ray sources. The spatial fractionation of the homogeneous beam into an array of microbeams is possible using a multislit collimator (MSC), i.e. a machined metal block with regular apertures. Dosimetry in MRT is challenging and previous works still show differences between calculated and experimental dose profiles of 10-30%, which are not acceptable for a clinical implementation of treatment. The interaction of the X-rays with the MSC may contribute to the observed discrepancies; the present study therefore investigates the dose contribution due to radiation interaction with the MSC inner walls and radiation leakage of the MSC. Dose distributions inside a water-equivalent phantom were evaluated for different field sizes and three typical spectra used for MRT studies at the European Synchrotron Biomedical beamline ID17. Film dosimetry was utilized to determine the contribution of radiation interaction with the MSC inner walls; Monte Carlo simulations were implemented to calculate the radiation leakage contribution. Both factors turned out to be relevant for the dose deposition, especially for small fields. Photons interacting with the MSC walls may bring up to 16% more dose in the valley regions, between the microbeams. Depending on the chosen spectrum, the radiation leakage close to the phantom surface can contribute up to 50% of the valley dose for a 5 mm × 5 mm field. The current study underlines that a detailed characterization of the MSC must be performed systematically and accurate MRT dosimetry protocols must include the contribution of radiation leakage and radiation interaction with the MSC in order to avoid significant errors in the dose evaluation at the micrometric scale.

Report on G4-Med, a Geant4 benchmarking system for medical physics applications developed by the Geant4 Medical Simulation Benchmarking Group.

BACKGROUND: Geant4 is a Monte Carlo code extensively used in medical physics for a wide range of applications, such as dosimetry, micro- and nanodosimetry, imaging, radiation protection, and nuclear medicine. Geant4 is continuously evolving, so it is crucial to have a system that benchmarks this Monte Carlo code for medical physics against reference data and to perform regression testing. AIMS: To respond to these needs, we developed G4-Med, a benchmarking and regression testing system of Geant4 for medical physics. MATERIALS AND METHODS: G4-Med currently includes 18 tests. They range from the benchmarking of fundamental physics quantities to the testing of Monte Carlo simulation setups typical of medical physics applications. Both electromagnetic and hadronic physics processes and models within the prebuilt Geant4 physics lists are tested. The tests included in G4-Med are executed on the CERN computing infrastructure via the use of the geant-val web application, developed at CERN for Geant4 testing. The physical observables can be compared to reference data for benchmarking and to results of previous Geant4 versions for regression testing purposes. RESULTS: This paper describes the tests included in G4-Med and shows the results derived from the benchmarking of Geant4 10.5 against reference data. DISCUSSION: Our results indicate that the Geant4 electromagnetic physics constructor G4EmStandardPhysics_option4 gives a good agreement with the reference data for all the tests. The QGSP_BIC_HP physics list provided an overall adequate description of the physics involved in hadron therapy, including proton and carbon ion therapy. New tests should be included in the next stage of the project to extend the benchmarking to other physical quantities and application scenarios of interest for medical physics. CONCLUSION: The results presented and discussed in this paper will aid users in tailoring physics lists to their particular application.

Microdosimetry of a therapeutic proton beam with a mini-TEPC and a MicroPlus-Bridge detector for RBE assessment.

Proton beams are widely used worldwide to treat localized tumours, the lower entrance dose and no exit dose, thus sparing surrounding normal tissues, being the main advantage of this treatment modality compared to conventional photon techniques. Clinical proton beam therapy treatment planning is based on the use of a general relative biological effectiveness (RBE) of 1.1 along the whole beam penetration depth, without taking into account the documented increase in RBE at the end of the depth dose profile, in the Bragg peak and beyond. However, an inaccurate estimation of the RBE can cause both underdose or overdose, in particular it can cause the unfavourable situation of underdosing the tumour and overdosing the normal tissue just beyond the tumour, which limits the treatment success and increases the risk of complications. In view of a more precise dose delivery that takes into account the variation of RBE, experimental microdosimetry offers valuable tools for the quality assurance of LET or RBE-based treatment planning systems. The purpose of this work is to compare the response of two different microdosimetry systems: the mini-TEPC and the MicroPlus-Bridge detector. Microdosimetric spectra were measured across the 62 MeV spread out Bragg peak of CATANA with the mini-TEPC and with the Bridge microdosimeter. The frequency and dose distributions of lineal energy were compared and the different contributions to the spectra were analysed, discussing the effects of different site sizes and chord length distributions. The shape of the lineal energy distributions measured with the two detectors are markedly different, due to the different water-equivalent sizes of the sensitive volumes: 0.85 µm for the TEPC and 17.3 µm for the silicon detector. When the Loncol's biological weighting function is applied to calculate the microdosimetric assessment of the RBE, both detectors lead to results that are consistent with biological survival data for glioma U87 cells. Both the mini-TEPC and the MicroPlus-Bridge detector can be used to assess the RBE variation of a 62 MeV modulated proton beam along its penetration depth. The microdosimetric assessment of the RBE based on the Loncol's weighting function is in good agreement with radiobiological results when the 10% biological uncertainty is taken into account.

Validation of Geant4 for silicon microdosimetry in heavy ion therapy.

Microdosimetry is a particularly powerful method to estimate the relative biological effectiveness (RBE) of any mixed radiation field. This is particularly convenient for therapeutic heavy ion therapy (HIT) beams, referring to ions larger than protons, where the RBE of the beam can vary significantly along the Bragg curve. Additionally, due to the sharp dose gradients at the end of the Bragg peak (BP), or spread out BP, to make accurate measurements and estimations of the biological properties of a beam a high spatial resolution is required, less than a millimetre. This requirement makes silicon microdosimetry particularly attractive due to the thicknesses of the sensitive volumes commonly being ∼10 [Formula: see text]m or less. Monte Carlo (MC) codes are widely used to study the complex mixed HIT radiation field as well as to model the response of novel microdosimeter detectors when irradiated with HIT beams. Therefore it is essential to validate MC codes against experimental measurements. This work compares measurements performed with a silicon microdosimeter in mono-energetic [Formula: see text], [Formula: see text] and [Formula: see text] ion beams of therapeutic energies, against simulation results calculated with the Geant4 toolkit. Experimental and simulation results were compared in terms of microdosimetric spectra (dose lineal energy, [Formula: see text]), the dose mean lineal energy, y  D and the RBE10, as estimated by the microdosimetric kinetic model (MKM). Overall Geant4 showed reasonable agreement with experimental measurements. Before the distal edge of the BP, simulation and experiment agreed within ∼10% for y  D and ∼2% for RBE10. Downstream of the BP less agreement was observed between simulation and experiment, particularly for the [Formula: see text] and [Formula: see text] beams. Simulation results downstream of the BP had lower values of y  D and RBE10 compared to the experiment due to a higher contribution from lighter fragments compared to heavier fragments.

The impact of sensitive volume thickness for silicon on insulator microdosimeters in hadron therapy.

Compact silicon on insulator (SOI) microdosimeters have been used to characterise the radiation field of many different hadron therapy beams. SOI devices are particularly attractive in hadron therapy fields due to their spatial resolution being well suited to the sharp dose gradients at the end of the primary beam's range. Due to the small size of SOI's sensitive volumes (SVs), which are usually ∼1-10 [Formula: see text]m thick, the fabrication of these devices can present challenges which are not as common for more conventional thickness silicon devices such as silicon spectroscopy detectors. Microdosimetry is the study of the energy deposition in micrometre sized volumes representing biological sites and is a powerful approach to estimate the biological effect of radiation on the micron-scale level, in a cell. However, cell sizes vary extensively translating in different energy deposition spectra. This work studies SV thicknesses between 1 and 100 [Formula: see text]m using Geant4 and examines the impact of SV dimensions on microdosimetric quantities. The quantities studied were the frequency mean lineal energy, [Formula: see text], and the dose mean lineal energy, [Formula: see text]. Additionally the relative biological effectiveness (RBE), estimated by the microdosimetric kinetic model (MKM), is also investigated. To study the impact of the SV thickness, SOI microdosimeters were irradiated with proton, [Formula: see text] and [Formula: see text] ion beams with ranges of ∼160 mm, with the microdosimeter being set at various positions along the Bragg curve. It was found that [Formula: see text] was influenced the least in proton beams and increased for heavier ion beams. Conversely, [Formula: see text] was impacted by the SV thickness the most in proton beams and [Formula: see text] was the least. Similar to [Formula: see text], protons were impacted the most by the SV thickness when estimating the RBE using the MKM. The cause of these differences was largely due to the different densities of the delta electron track structure for the case of [Formula: see text] and the energy transferred to the medium from the primary beam for [Formula: see text].

Influence of track structure and condensed history physics models of Geant4 to nanoscale electron transport in liquid water.

The Geant4 toolkit offers a range of electromagnetic (EM) models for simulating the transport of charged particles down to sub-keV energies. They can be divided to condensed-history (CH) models (like the Livermore and Penelope models) and the track-structure (TS) models included in the Geant4-DNA low-energy extension of Geant4. Although TS models are considered the state-of-the-art for nanoscale electron transport, they are difficult to develop, computationally intensive, and commonly tailored to a single medium (e.g., water) which prohibits their use in a wide range of applications. Thus, the use of CH models down to sub-keV energies is particularly intriguing in the context of general-purpose Monte Carlo codes. The aim of the present work is to compare the performance of the CH models of Geant4 against the recently implemented TS models of Geant4-DNA for nanoscale electron transport. Calculations are presented for two fundamental quantities, the dose-point-kernel and the microdosimetric lineal energy. The influence of user-defined simulation parameters (tracking and production cuts, and maximum step size) on the above calculations is also examined. It is shown that Livermore offers the best performance among the CH models of Geant4 for nanoscale electron transport. However, even under optimally-chosen simulation parameters, the differences between the CH and TS models examined may be sizeable for low energy electrons (<1 keV) and/or nanometer size targets (<100 nm).

A New Standard DNA Damage (SDD) Data Format.

Our understanding of radiation-induced cellular damage has greatly improved over the past few decades. Despite this progress, there are still many obstacles to fully understand how radiation interacts with biologically relevant cellular components, such as DNA, to cause observable end points such as cell killing. Damage in DNA is identified as a major route of cell killing. One hurdle when modeling biological effects is the difficulty in directly comparing results generated by members of different research groups. Multiple Monte Carlo codes have been developed to simulate damage induction at the DNA scale, while at the same time various groups have developed models that describe DNA repair processes with varying levels of detail. These repair models are intrinsically linked to the damage model employed in their development, making it difficult to disentangle systematic effects in either part of the modeling chain. These modeling chains typically consist of track-structure Monte Carlo simulations of the physical interactions creating direct damages to DNA, followed by simulations of the production and initial reactions of chemical species causing so-called "indirect" damages. After the induction of DNA damage, DNA repair models combine the simulated damage patterns with biological models to determine the biological consequences of the damage. To date, the effect of the environment, such as molecular oxygen (normoxic vs. hypoxic), has been poorly considered. We propose a new standard DNA damage (SDD) data format to unify the interface between the simulation of damage induction in DNA and the biological modeling of DNA repair processes, and introduce the effect of the environment (molecular oxygen or other compounds) as a flexible parameter. Such a standard greatly facilitates inter-model comparisons, providing an ideal environment to tease out model assumptions and identify persistent, underlying mechanisms. Through inter-model comparisons, this unified standard has the potential to greatly advance our understanding of the underlying mechanisms of radiation-induced DNA damage and the resulting observable biological effects when radiation parameters and/or environmental conditions change.

SOI microdosimetry and modified MKM for evaluation of relative biological effectiveness for a passive proton therapy radiation field.

With more patients receiving external beam radiation therapy with protons, it becomes increasingly important to refine the clinical understanding of the relative biological effectiveness (RBE) for dose delivered during treatment. Treatment planning systems used in clinics typically implement a constant RBE of 1.1 for proton fields irrespective of their highly heterogeneous linear energy transfer (LET). Quality assurance tools that can measure beam characteristics and quantify or be indicative of biological outcomes become necessary in the transition towards more sophisticated RBE weighted treatment planning and for verification of the Monte Carlo and analytical based models they use. In this study the RBE for the CHO-K1 cell line in a passively delivered clinical proton spread out Bragg peak (SOBP) is determined both in vitro and using a silicon-on-insulator (SOI) microdosimetry method paired with the modified microdosimetric kinetic model. The RBE along the central axis of a SOBP with 2 Gy delivered at the middle of the treatment field was found to vary between 1.11-1.98 and the RBE for 10% cell survival between 1.07-1.58 with a 250 kVp x-ray reference radiation and between 1.19-2.34 and 0.95-1.41, respectively, for a Co60 reference. Good agreement was found between RBE values calculated from the SOI-microdosimetry-MKM approach and in vitro. A strong correlation between proton lineal energy and RBE was observed particularly in the distal end and falloff of the SOBP.

Optimisation of the design of SOI microdosimeters for hadron therapy quality assurance.

Silicon-on-insulator (SOI) microdosimeters offer a promising method for routine quality assurance (QA) for hadron therapy due to their ease of operation and high spatial resolution. However, one complication which has been shown previously is that the traditional use of the mean chord length, [Formula: see text], calculated using Cauchy's formula, for SOI devices in clinical carbon ion fields is not appropriate due to the strong directionality of the radiation field. In a previous study, we demonstrated that the mean path length, [Formula: see text], which is the mean path of charged particles in the sensitive volume (SV), is a more appropriate method to obtain microdosimetric quantities and biological relevant values, namely the relative biological effectiveness (RBE) by means of the microdosimetric kinetic model. The previous work, which was limited to mono-energetic [Formula: see text] ion beams typical of heavy ion therapy (HIT), is extended here to investigate the [Formula: see text] in a pristine proton beam as well as for spread out Bragg peaks (SOBP) for both proton and carbon ion clinical beams. In addition, the angular dependence of the SOI device for a number of different SV designs is also investigated to quantify the effects which the alignment has on the [Formula: see text]. It is demonstrated that the [Formula: see text] can be accurately estimated along the depth of a pristine or SOBP using the energy deposition spectra for both proton and [Formula: see text] ion beams. This observation allows a quick and accurate estimation of the [Formula: see text] for experimental use.

Geant4-DNA example applications for track structure simulations in liquid water: A report from the Geant4-DNA Project.

This Special Report presents a description of Geant4-DNA user applications dedicated to the simulation of track structures (TS) in liquid water and associated physical quantities (e.g., range, stopping power, mean free path…). These example applications are included in the Geant4 Monte Carlo toolkit and are available in open access. Each application is described and comparisons to recent international recommendations are shown (e.g., ICRU, MIRD), when available. The influence of physics models available in Geant4-DNA for the simulation of electron interactions in liquid water is discussed. Thanks to these applications, the authors show that the most recent sets of physics models available in Geant4-DNA (the so-called "option4" and "option 6" sets) enable more accurate simulation of stopping powers, dose point kernels, and W-values in liquid water, than the default set of models ("option 2") initially provided in Geant4-DNA. They also serve as reference applications for Geant4-DNA users interested in TS simulations.

A novel high-resolution 2D silicon array detector for small field dosimetry with FFF photon beams.

PURPOSE: Flattening filter free (FFF) beams are increasingly being considered for stereotactic radiotherapy (SRT). For the first time, the performance of a monolithic silicon array detector under 6 and 10 MV FFF beams was evaluated. The dosimeter, named "Octa" and designed by the Centre for Medical Radiation Physics (CMRP), was tested also under flattened beams for comparison. METHODS: Output factors (OFs), percentage depth-dose (PDD), dose profiles (DPs) and dose per pulse (DPP) dependence were investigated. Results were benchmarked against commercially available detectors for small field dosimetry. RESULTS: The dosimeter was shown to be a 'correction-free' silicon array detector for OFs and PDD measurements for all the beam qualities investigated. Measured OFs were accurate within 3% and PDD values within 2% compared against the benchmarks. Cross-plane, in-plane and diagonal DPs were measured simultaneously with high spatial resolution (0.3 mm) and real time read-out. A DPP dependence (24% at 0.021 mGy/pulse relative to 0.278 mGy/pulse) was found and could be easily corrected for in the case of machine specific quality assurance applications. CONCLUSIONS: Results were consistent with those for monolithic silicon array detectors designed by the CMRP and previously characterized under flattened beams only, supporting the robustness of this technology for relative dosimetry for a wide range of beam qualities and dose per pulses. In contrast to its predecessors, the design of the Octa offers an exhaustive high-resolution 2D dose map characterization, making it a unique real-time radiation detector for small field dosimetry for field sizes up to 3 cm side.

MICRODOSIMETRIC APPLICATIONS IN PROTON AND HEAVY ION THERAPY USING SILICON MICRODOSIMETERS.

Using the CMRP 'bridge' µ+ probe, microdosimetric measurements were undertaken out-of-field using a therapeutic scanning proton pencil beam and in-field using a 12C ion therapy field. These measurements were undertaken at Mayo Clinic, Rochester, USA and at HIMAC, Chiba, Japan, respectively. For a typical proton field used in the treatment of deep-seated tumors, we observed dose-equivalent values ranging from 0.62 to 0.99 mSv/Gy at locations downstream of the distal edge. Lateral measurements at depths close to the entrance and along the SOBP plateau were found to reach maximum values of 3.1 mSv/Gy and 5.3 mSv/Gy at 10 mm from the field edge, respectively, and decreased to ~0.04 mSv/Gy 120 mm from the field edge. The ability to measure the dose-equivalent with high spatial resolution is particularly relevant to healthy tissue dose calculations in hadron therapy treatments. We have also shown qualitatively and quantitively the effects critical organ motion would have in treatment using microdosimetric spectra. Large differences in spectra and RBE10 were observed for treatments where miscalculations of 12C ion range would result in critical structures being irradiated, showing the importance of motion management.

Development of a high resolution voxelised head phantom for medical physics applications.

Computational anthropomorphic phantoms have become an important investigation tool for medical imaging and dosimetry for radiotherapy and radiation protection. The development of computational phantoms with realistic anatomical features contribute significantly to the development of novel methods in medical physics. For many applications, it is desirable that such computational phantoms have a real-world physical counterpart in order to verify the obtained results. In this work, we report the development of a voxelised phantom, the HIGH_RES_HEAD, modelling a paediatric head based on the commercial phantom 715-HN (CIRS). HIGH_RES_HEAD is unique for its anatomical details and high spatial resolution (0.18×0.18mm2 pixel size). The development of such a phantom was required to investigate the performance of a new proton computed tomography (pCT) system, in terms of detector technology and image reconstruction algorithms. The HIGH_RES_HEAD was used in an ad-hoc Geant4 simulation modelling the pCT system. The simulation application was previously validated with respect to experimental results. When compared to a standard spatial resolution voxelised phantom of the same paediatric head, it was shown that in pCT reconstruction studies, the use of the HIGH_RES_HEAD translates into a reduction from 2% to 0.7% of the average relative stopping power difference between experimental and simulated results thus improving the overall quality of the head phantom simulation. The HIGH_RES_HEAD can also be used for other medical physics applications such as treatment planning studies. A second version of the voxelised phantom was created that contains a prototypic base of skull tumour and surrounding organs at risk.

Development of a Geant4 application to characterise a prototype neutron detector based on three orthogonal 3He tubes inside an HDPE sphere.

The Bonner Sphere Spectrometer (BSS) system is a well-established technique for neutron dosimetry that involves detection of thermal neutrons within a range of hydrogenous moderators. BSS detectors are often used to perform neutron field surveys in order to determine the ambient dose equivalent H*(10) and estimate health risk to personnel. There is a potential limitation of existing neutron survey techniques, since some detectors do not consider the direction of the neutron field, which can result in overly conservative estimates of dose in neutron fields. This paper shows the development of a Geant4 simulation application to characterise a prototype neutron detector based on three orthogonal 3He tubes inside a single HDPE sphere built at the Australian Nuclear Science and Technology Organisation (ANSTO). The Geant4 simulation has been validated with respect to experimental measurements performed with an Am-Be source.