Systems and methods for image simulator

Abstract

Systems and methods for simulating a CBCT scan implement or include receiving a set of initial CBCT scan parameters; generating a mono-energetic simulation of the CBCT scan using the initial CBCT scan parameters; storing a slice of the mono-energetic simulation in memory; receiving a set of updated CBCT scan parameters; and modifying the stored slice of the mono-energetic simulation based on the updated CBCT scan parameters.

Description

SYSTEMS AND METHODS FOR IMAGE SIMULATORCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of and priority to U.S. Provisional Patent Application No. 63 / 673,711, filed on July 20, 2024, and entitled “RTIS: Real-Time Image Simulator,” the entire contents of which are herein incorporated by reference for all purposes.TECHNICAL FIELD

[0002] This disclosure relates to the field of computed tomography (CT) techniques. In examples, this disclosure is particularly directed to simulating cone-beam CT images from planning CT images to allow real-time updating of the imaging technique.BACKGROUND

[0003] Cone-beam CT (CBCT) imaging plays a role in radiotherapy treatment planning and delivery, allowing for visualization of patient anatomy during treatment. Likewise, CBCT treatment plans can adapt to daily patient anatomy changes, thereby providing better normal tissue sparing and tumor coverage. However, comparative techniques suffer from challenges in adapting treatments due to CBCT image quality, with issues such as artifacts and nonuniformity in Hounsfield units leading to errors in CBCT dose calculations and organ segmentations. Machine learning techniques, particularly generative adversarial networks (GANs), have been proposed as other comparative techniques for improving CBCT image quality. Still, their effectiveness is hindered by the lack of simultaneously acquired CT and CBCT image pairs, an essential element for training.SUMMARY

[0004] To address these and other limitations in the comparative techniques, the present disclosure presents imaging techniques and modalities for CBCT imaging.

[0005] One example technique is referred to as a GPU Enhanced CT to CBCT Simulator (GECCO) technique, which utilizes a hybrid Monte Carlo (MC) toolkit with the goal of generating highly accurate synthetic CT / CBCT datasets, which may be used, in examples, for training machine learning models, thus facilitating more widespread adoption ofCBCT adaptive radiotherapy. GECCO also serves as a fast and accurate CBCT simulation tool, effecting improvements in CBCT imaging and contributing to the broader field of radiotherapy.

[0006] Another example technique is referred to as a Real-Time Image Simulator (RTIS) technique, which calculates primary projections through GPU raytracing, combining many mono-energetic simulations into one poly-energetic simulation. This example builds on a simulation toolkit to combine mono-energetic projections of voxelized phantoms into a poly- energetic projection by weighting them according to beam energy spectrum and detector energy deposition efficiencies. After calculation, projections are stored and can be re-weighted to simulate different image settings, such as different x-ray tube peak voltage (kVp) or filtration settings. Additionally, noise and scatter levels can be modified using mAs and background scatter level settings.

[0007] As will be set forth in detail below, experimental validation has been performed using anthropomorphic phantoms and anonymized patient data. In examples, the techniques set forth herein achieve a low latency (<40 ms) output that provides real-time estimation of CBCT quality as the user adjusts acquisition parameters. These techniques have a range of uses and effects, including allowing users to prospectively determine target CT acquisition parameters, thus improving scan quality while reducing or eliminating the number of repeat scans needed. These effects may further reduce patient procedure time and imaging dose.BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Some examples of the disclosure are described herein with reference to the accompanying figures. The description, together with the figures, makes apparent to a person having ordinary skill in the art how some examples of the disclosure may be practiced. The figures are for the purpose of illustrative discussion and no attempt is made to show structural details of an example in more detail than is necessary for a fundamental understanding of the teachings of the disclosures. In the drawings:

[0009] FIG. 1 illustrates an example workflow in accordance with various aspects of the present disclosure.

[0010] FIG. 2 illustrates an example validation of a simulation method in accordance with various aspects of the present disclosure.

[0011] FIG. 3 illustrates example simulation images in accordance with various aspects of the present disclosure.

[0012] FIG. 4 illustrates an example validation of a simulation method in accordance with various aspects of the present disclosure.

[0013] FIG. 5 illustrates an example validation of a simulation method in accordance with various aspects of the present disclosure.

[0014] FIG. 6 illustrates an example validation of a simulation method in accordance with various aspects of the present disclosure.

[0015] FIG. 7 illustrates example simulation images in accordance with various aspects of the present disclosure.

[0016] FIG. 8 illustrates example simulation images in accordance with various aspects of the present disclosure.

[0017] FIG. 9 illustrates an example workflow in accordance with various aspects of the present disclosure.

[0018] FIG. 10 illustrates an example method in accordance with various aspects of the present disclosure.

[0019] FIG. 11 illustrates an example computing device in accordance with various aspects of the present disclosure.DETAILED DESCRIPTION

[0020] In the following detailed description, reference is made to the accompanying drawings in which specific examples are shown by way of illustration. These examples are described in sufficient detail to enable those of ordinary skill in the art to practice the disclosure. It should be understood, however, that the detailed description and the specific examples, while indicating examples of embodiments of the disclosure, are given by way of illustration only and not by way of limitation. From this disclosure, various substitutions, modifications, additions rearrangements, or combinations thereof within the scope of the disclosure may be made and will become apparent to those of ordinary skill in the art.

[0021] Unless otherwise indicated, the various features illustrated in the drawings may not be drawn to scale. The illustrations presented herein are not necessarily intended to be actual views of any particular method, device, or system, but are merely idealized representations that are employed to describe various embodiments of the disclosure. Accordingly, the dimensions of the various features as illustrated may be arbitrarily expanded or reduced for clarity. In addition, some of the drawings may be simplified for clarity. Thus,the drawings may not depict all of the components of a given apparatus (e.g., device) or method. In addition, like reference numerals may be used to denote like features throughout the specification and figures.

[0022] It should be understood that any reference to an element herein using a designation such as “first,” “second,” and so forth does not limit the quantity or order of those elements, unless such limitation is explicitly stated. Rather, these designations may be used herein as a convenient method of distinguishing between two or more elements or instances of an element. Thus, a reference to first and second elements does not mean that only two elements may be employed there or that the first element must precede the second element in some manner. Also, unless stated otherwise a set of elements may comprise one or more elements.

[0023] Unless otherwise specified or indicated by context, the terms “a,” “an,” and “the” mean “one or more.” As used herein, unless otherwise limited or defined, “or” indicates a non-exclusive list of components or operations that can be present in any variety of combinations, rather than an exclusive list of components that can be present only as alternatives to each other. For example, a list of “A, B, or C” indicates options of: A; B; C; A and B; A and C; B and C; and A, B, and C. Correspondingly, the term “or” as used herein is intended to indicate exclusive alternatives only when preceded by terms of exclusivity, such as “only one of,” or “exactly one of.” For example, a list of “only one of A, B, or C” indicates options of: A, but not B and C; B, but not A and C; and C, but not A and B. In contrast, a list preceded by “one or more” (and variations thereon) and including “or” to separate listed elements indicates options of one or more of any or all of the listed elements. For example, the phrases “one or more of A, B, or C” and “at least one of A, B, or C” indicate options of: one or more A; one or more B; one or more C; one or more A and one or more B; one or more B and one or more C; one or more A and one or more C; and one or more A, one or more B, and one or more C. Similarly, a list preceded by “a plurality of’ (and variations thereon) and including “or” to separate listed elements indicates options of one or more of each of multiple of the listed elements. For example, the phrases “a plurality of A, B, or C” and “two or more of A, B, or C” indicate options of: one or more A and one or more B; one or more B and one or more C; one or more A and one or more C; and one or more A, one or more B, and one or more C.

[0024] As used herein, “about,” “approximately,” “substantially,” and “significantly” will be understood by persons of ordinary skill in the art and will vary to some extent on the context in which they are used. If there are uses of these terms which are not clear to personsof ordinary skill in the art given the context in which they are used, “about” and “approximately” will mean plus or minus <10% of the particular term and “substantially” and “significantly” will mean plus or minus >10% of the particular term.

[0025] As used herein, the terms “include” and “including” have the same meaning as the terms “comprise” and “comprising” in that these latter terms are “open” transitional terms that do not limit claims only to the recited elements succeeding these transitional terms. The term “consisting of,” while encompassed by the term “comprising,” should be interpreted as a “closed” transitional term that limits claims only to the recited elements succeeding this transitional term. The term “consisting essentially of,” while encompassed by the term “comprising,” should be interpreted as a “partially closed” transitional term which permits additional elements succeeding this transitional term, but only if those additional elements do not materially affect the basic and novel characteristics of the claim.

[0026] As used herein, the term “subject” may be used interchangeably with the term “patient” or “individual” and may include an “animal” and in particular a “mammal.” Mammalian subjects may include humans and other primates, domestic animals, farm animals, and companion animals such as dogs, cats, guinea pigs, rabbits, rats, mice, horses, cattle, cows, and the like.

[0027] As used herein, the terms “treat” or “treatment” encompass both “preventative” and “curative” treatment. “Preventative” treatment is meant to indicate a postponement of development of a disease, a symptom of a disease, or medical condition, suppressing symptoms that may appear, or reducing the risk of developing or recurrence of a disease or symptom. “Curative” treatment includes reducing the severity of or suppressing the worsening of an existing disease, symptom, or condition. Thus, treatment includes ameliorating or preventing the worsening of existing disease symptoms, preventing additional symptoms from occurring, ameliorating or preventing the underlying systemic causes of symptoms, inhibiting the disorder or disease, e.g., arresting the development of the disorder or disease, relieving the disorder or disease, causing regression of the disorder or disease, relieving a condition caused by the disease or disorder, or stopping the symptoms of the disease or disorder.

[0028] CBCT adaptive radiotherapy is a technique that aims to adapt radiotherapy treatments to patient anatomy based on patient CBCTs taken at the time of radiotherapy delivery. This approach has demonstrated benefits in improving tumor targeting and sparing normal tissue. However, the widespread adoption of CBCT adaptive radiotherapy has heightened the demand for enhancing CBCT image quality. The increased scatter and longeracquisition time in CBCT can lead to low image quality characteristics, such as artifacts and non-uniformity in Hounsfield Unit (HU) values, in turn leading to increased errors in CBCT dose calculations and organ segmentations (as compared to CT dose calculations and organ segmentations). Machine learning methods, particularly generative adversarial networks (GANs), have been attempted as solutions for improving CBCT image quality, thereby facilitating better plan adaptation. One such comparative approach involves using matching CT and CBCT image pairs to train a GAN to create an equivalent “synthetic CT” from the CBCT. In other comparative approaches, U-net architectures are sometimes employed for CBCT segmentation, utilizing planning CT as ground truth contours. However, a significant challenge arises from the fact that CT and CBCT images are not acquired simultaneously, resulting in discrepancies in patient anatomy between the image pairs used for training according to these comparative approaches.

[0029] Monte Carlo (MC) simulation is regarded as the gold standard for simulating medical images, particularly for generating CT and CBCT image pairs. Two comparative approaches to achieving this involve either using real experimental phantoms with known composition and dimensions, though with limited relation to real patient data, or employing virtual anthropomorphic phantoms that feature realistic anatomy but do not capture the anatomical variation present in the patient population. However, a third methodology can be leveraged for CBCT simulation: creating a phantom from CT data using elemental decomposition methods and utilizing it as input for the simulation, a technique supported by many MC codes. This approach offers excellent fidelity to patient anatomical structures while being limited by the finite resolution of the patient phantom and the inherent noise, streaking and beam hardening artifacts present in CT imaging. For machine learning applications, this third method holds the added advantage of capturing the variability in human anatomical features observed in the patient population, a characteristic challenging to replicate using purely anthropomorphic phantoms. Although highly accurate, MC simulation is prohibitively computationally expensive, a limitation that can be overcome with emerging graphics processing unit (GPU) and Hybrid MC methods.

[0030] CBCT simulation may be done through MC particle transport, and while MC provides excellent accuracy, MC CBCT simulation in comparative implementations is generally plagued by prolonged computation times. This issue stems from various factors, such as the required number of projections, the handling of high-resolution voxelized phantoms, and the transport of optical photons within the simulator, all of which can detrimentally impactcomputation speed. For instance, one study revealed that simulating a single projection image with a scintillating detector and 1010 primary x-rays consumed 3000 CPU-hours. Conversely, other GPU MC methods have exhibited speedup factors exceeding 1000 times that of traditional MC. Another challenge for MC methods is increased load times, compute times, and memory constraints as the resolution of an anthropomorphic phantom escalates. For example, in another study evaluating the performance of Geant4, MCNP, and PHITS concerning dose calculation speeds in voxelized anthropomorphic phantoms, simulation times increased by an order of magnitude in Geant4 when decreasing the voxel size from 1 mm to 0.1 mm, while MCNP and PHITS could not simulate a 0.1 mm phantom due to the number of phantom voxels exceeding the maximum limit of the simulation codes.

[0031] Other simulation approaches may be used to enhance simulation time. While GPU MC is a method that exhibits speedup factors of 27-90 over CPU-based MC, ray-tracing methods are accessible for both CPU and GPU and generate realistic images in under a second as compared to the minutes and hours for even GPU MC. Although these methods are highly efficient, they cannot account for scatter in phantoms and simulate detector response. To overcome this, hybrid methods exist that use both ray-tracing methods and MC. Hybrid methods have demonstrated success in certain applications. For instance, gDRR simulations exhibited agreement with experimental noise amplitude in CBCT, with a relative difference of 3.8 %.

[0032] A number of comparative hybrid and analytical CBCT simulation tools have been developed, each with its approaches, capabilities, and validation strategies. One comparative hybrid Monte-Carlo approach, Fastcat, was comprehensively evaluated for accuracy across different imaging systems, mitigating potential overfitting concerns, likewise tools like DukeSim, XCIST (Cat-Sim), and gDRR have been well validated but only on a single scanner. VOXSI provides validation at multiple tube potentials, but its reliance on a single source and detector setup can limit generalizability to different imaging systems. Fastcat’s validation utilized the Catphan 504 phantom and imaging metrics, demonstrating increased accuracy in HU values and contrast-to-noise ratio (CNR) compared to VOXSI’ s tissuemimicking phantom. Fastcat also validated detector modulation transfer function (MTF), accounting for realistic scintillator crosstalk beyond nearest neighbors, an aspect overlooked by XCIST and DukeSim’ s simplified spatial resolution estimations. Conversely, DukeSim uses GPU MC to provide scatter, allowing for more accurate and general scatter characterization, whereas Fastcat assumes rotational symmetry of the objects being scanned and usesprecalculated MC scatter instead of calculating scatter based on the phantom itself. This presents a major barrier in the art towards accurate patient CBCT simulation with Fastcat as the image quality of CBCT is very dependent on the scatter associated with the CBCT geometry.

[0033] In examples, the systems and methods set forth start from open-source libraries to build on validation across multiple imaging systems, accuracy, realistic detector modeling, and computational performance. Herein, a MC toolkit that extends and improves upon the comparative techniques is provided. The MC toolkit will be described in terms of two aspects: the GECCO application and the RTIS application. The GECCO application simplifies the analytical calculations and couples them with (e.g., GGEMS) GPU-accelerated MC scatter calculations for CBCT, enabling computationally efficient simulations. Additionally, GECCO incorporates modules to digitize CT data into elemental compositions, which serve as input for the simulations, resulting in perfectly aligned CT / CBCT image pairs for machine-learning applications. Likewise, the RTIS application provides simulations of CBCT generated from planning CT that update in real-time according to user-selected mAs, kVp, and beam filtration.

[0034] In one example, the GECCO technique may be implemented as instructions stored in a computer-readable medium (e.g., a non-transitory computer readable medium) that are executable by a processor of a computing device (e.g., as a software program). For purposes of discussion, the “GECCO software toolkit” is a collection of software (e.g., in Python) for pre / post-processing, simulating, reconstructing, and analyzing CBCT simulations, whereas the “GECCO application” and the “RTIS application” are software systems that connects the different parts of the GECCO toolkit to perform specific simulations.

[0035] The GECCO application employs a hybrid MC approach to generate projection images. An overview of the workflow 100 can be seen in FIG. 1. In the example workflow 100, the workflow components are broadly categorized as pre-processing operations 110, simulation operations 120, and post-processing operations 130. Each set of operations includes a plurality of subroutines or submethods. Of these, boxes identified by a black triangle in the upper-left comer may be implemented as modified GGEMS methods, boxes identified by a black triangle in the upper-right corner may be implemented as modified Fastcat methods, boxes identified by a black triangle in the bottom-left comer may be implemented as native GECCO methods, and boxes identified by a black triangle in the bottom-right corner may be implemented as third-party dependencies. The RTIS application may utilize a workflow similar to the workflow 100, except that the RTIS application does not utilize GGEMS.

[0036] The pre-processing operations 110 involve converting input data from simplified or standard formats into forms compatible with the simulation. For example, a bowtie filter may be defined as a vector of thicknesses versus angle. This must be transformed into a 2D fluence map that weights the primary X-ray fluence in a GECCO simulation. Similarly, DICOM images are decomposed into tissue classifications, and essential metadata such as voxel dimensions and image orientation are reformatted into geometries that GECCO or RTIS can interpret. To improve efficiency, the pre-processing operations 110 may be performed once and the outputs may be saved, allowing simulations to be rerun without repeating the conversion steps.

[0037] The simulation operations 120 are organized top to bottom in chronological order. These begin with the X-ray spectrum, which may be generated or loaded from file. This spectrum influences the detector’s point spread function (PSF), the bowtie filter effects, and other energy-dependent processes. Modified Fastcat and GGEMS operate in parallel (as shown side-by-side) and their outputs are later combined. In the case of RTIS, as noted above, Modified Fastcat may be used without GGEMS. The post-processing operations 130 may be performed after the simulation run (e.g., on the next day), and thus may be separated from the main loop. Tasks such as scatter denoising and interpolation, though shown separately, could also be considered part of post-processing depending on the workflow.

[0038] In this example, GECCO is written in Python 3 (3.10), and simulations are specified as Python scripts. Primary projections are generated through analytical ray tracing, while secondary (scatter) projections are generated using GPU-accel erated Monte Carlo simulations. This hybrid approach aims to reduce the overall simulation time. By decoupling the primary and secondary projections and using analytical simulations for the primary projections and MC simulations for the secondary (scatter) projections, it is possible leverage the properties of scattered radiation. Specifically, because scattered radiation is characterized by low spatial frequencies and slow angular variations, one can calculate fewer secondary projections than primary, thereby reducing the number of primary photons simulated and decreasing computational time. These fewer scatter projections are then denoised and interpolated to provide high-quality estimates of the scatter at all angles.

[0039] Primary projections are calculated using a modified and updated version of the opensource Fastcat Python simulation toolkit. Fastcat uses GPU ray tracing and precalculated Monte Carlo scatter to create fast cone-beam CT simulations. Briefly, Fastcat divides MC simulation into primary and secondary projections. Primary projections undergo analyticalattenuation, whereas scatter is simulated using MC techniques. However, MC scatter is only calculated at one angle per phantom and the phantom is approximated to be rotationally symmetric. To accommodate arbitrary poly-energetic beams, both primary and secondary radiation are simulated across 18 discrete energies ranging from 10 keV to 6 MeV, at intervals of 10 keV up to 100 keV and larger intervals in the MeV range. Image formation entails the weighting of primary radiation and scatter, the addition of quantum noise via Poisson scaling based on user-defined parameters, and subsequent weighting according to the x-ray spectrum and detector energy response. In implementations, only quantum noise is accounted for in Fastcat, as it typically yields a more substantial impact than electronic noise. The simulation of the source’s nominal focal spot involves convolving the detector point spread function (as calculated in Topas) with a Gaussian kernel, where the full width at half maximum (FWHM) is scaled by a magnification factor to reflect simulation geometry.

[0040] Modifications to Fastcat in GECCO include simulating scatter directly in the GGEMs GPU MC software toolkit. The analytical simulation code was also sped up by a factor of 38% and the memory usage decreased four times that of Fastcat by refactoring the code. Additionally, Fastcat uses an empirical fit between total energy released in matter (TERMA) to calculate doses in phantoms while GECCO uses GGEMs to calculate the dose in a phantom directly (with the option of having a voxelized dose report). Because MC is the gold standard for dose calculation, the GECCO approach is superior. Additionally, Spekpy, a software toolkit used for analytical spectrum generation, is integrated into GECCO, replacing XPECGEN, Fastcat’s analytical spectrum generator. Spekpy is faster and more accurate than XPECGEN. An electronic noise model was also included in GECCO, which models the electronic noise using a Poisson distribution with an expectation value based on the standard deviation of successive experimental dark fields. Moreover, GECCO has the additional capability of generating CBCT simulations directly from CT data (as will be described in more detail below), as well as the ability to read and write phantoms from .mhd files, a common phantom file format. GECCO simulations were conducted on a Linux desktop system featuring 64 GB memory, a Nvidia GeForce RTX 4090 GPU, and 32 AMD Ryzen 7 CPU cores.

[0041] GECCO secondary projections were calculated using a modified version of the GPU Geant4-based Monte Carlo Simulations (GGEMs) GPU MC software toolkit; more specifically, GECCO uses a modified GGEMs user code called FastMC which specifies GGEMs simulations using an XML script. FastMC was used to increase the ease of use and readability of GGEMS code. FastMC was developed in-house as a way to make GGEMSsimulations more accessible to users unfamiliar with C++ and was previously used in modelling CBCT.

[0042] While the functionality of GGEMs is to calculate secondary projections, primary projections are also calculated and used as an internal consistency check. The primary projections from GECCO are normalized to the GGEMs projections. Both primary projections are averaged in the y-direction, and the root mean error between the two profiles is then calculated. If the mean error is greater than 1 % of the maximum a flag is raised to the user to indicate a possible inconsistency.

[0043] To model CBCT accurately, additional components are added to the simulation (e.g., bowtie filters and anti-scatter grids) as shown in FIG. 1. Bowtie filters are implemented for both the primary and secondary projections in GECCO. The primary projections were filtered analytically according to a specific thickness of aluminum for each x-position along the detector. For the secondary projections, the GGEMS source code for generating photons from x-ray sources was modified from an isotropic source to a bowtie-filtered source: X-rays were filtered by a varying amount of aluminum filtration according to the angle of the photon relative to the central axis of the beam. In both primary and scatter bowtie filtration, the bowtie was specified as a text file containing the filter material and filter thicknesses as a function of the angle relative to the central axis of the beam.

[0044] To improve computational efficiency, GECCO uses an analytical model of the treatment head instead of full MC transport through components such as the beam hardening filter, bowtie, and inherent filtration. To simulate accurate head scatter, scatter from the x-ray tube and filters was characterized by running FastMC simulations with treatment heads defined in MC for photon energies from 10 to 200 keV in 10 keV increments scatter. These scatter distributions were then stored in GECCO. For poly-energetic beams, head scatter is estimated as a sum of discrete monoenergetic scatter profiles. These are analytically propagated through the phantom and weighted by the poly-energetic spectrum before being convolved with a Gaussian kernel (with FWHM of 2.7 cm). This kernel was empirically determined to provide the best agreement between full MC and analytically modeled simulations for a CTDI phantom.

[0045] Additionally, a model for the kV on-board imager (OBI) anti-scatter grid (ASG) was incorporated. This model draws upon existing measurements, leveraging primary transmission and scatter transmission factors from the 44rl0 ASG. These factors were determined to be 0.76 and 0.37, respectively. In the primary simulation, the primary fluence reaching the detector was adjusted by the primary transmission factor. The scatter transmissionfactor was used to separate secondary radiation into two components: one accepted by the detector and the other assumed to impinge upon the ASG at a mean angle of scatter incidence. The latter component underwent analytical filtration by 173 pm of lead, equivalent to the path length through the 36 pm lamella at the specified angle. The particle counts in these two segments were calculated to ensure that the total scatter transmission factor aligned with theoretical expectations.

[0046] In one use case, GECCO converts CT volumes to CBCT volumes for machine learning applications. To do this, patient phantoms were created through the elemental decomposition of CT images. HU values were converted to elemental compositions. Specifically, a method implemented in Topas MC was used for the conversion: HU values were converted into 25 materials according to set HU bins between -1000 HU and 4000 HU. Each material with a set elemental composition according to the type of material present at that HU value, the lowest HU material being air and the highest HU material being titanium. Additionally, unique HU values were converted to densities: densities D for a given HU value HU were calculated as D = o + (F X (FO+ Dcorr, where 0, F, and Foare theoffset, factor, and factor offset, respectively. These variables are dependent on the HU value with eight unique values for the HU ranges between -1000, -98, 15, 23, 101, 2001, 2995, and 2996, respectively to create a piecewise linear HU to density mapping curve. Dcorris the density correction, an array of density corrections for each HU value between -1000 and 2995. These variables are the same as those values which can be found in the Topas user documentation.

[0047] CT image densities are handled differently for the primary and secondary simulations. For the primary simulation, an array of densities is saved for each voxel of the CT scan. Fastcat typically loads the linear attenuation value at a given energy for each material in the phantom. This approach was modified in GECCO, the mass attenuation coefficient for each material is loaded. This array of material mass attenuation coefficients is then multiplied by the density array for each material at a specific energy, and the resulting array is used for raytracing. In this way, the primary simulation accounts for the density at each voxel.

[0048] For the secondary projections, an .mhd file is created containing an integer array specifying the material at each voxel of the phantom. Additionally, a range file is created containing the material for each binary value, and a material file is created with the material composition and density for each material. In the GGEMS simulation, the density of each material is approximated as the mean density of each material’s HU range. Whileapproximating each material as one density results in some loss of accuracy, GGEMS currently has no method for specifying unique densities for every voxel of a phantom. Nevertheless, these values are only used for calculating scatter and have less effect on the CBCT image than the primary projections, which do include unique densities.

[0049] GECCO uses much fewer secondary projections than primary projections to improve computational efficiency. This is accomplished through denoising and interpolation of the secondary projections. Specifically, secondary projections are average pooled with an 8 x 8 kernel (downsampled by a factor of 8) to remove noise. The pooled image is then fit with a 4th-order spline, and this spline is used to interpolate the image back to its original dimensions. These denoised, smoothed secondary images are then linearly interpolated to estimate secondary projections at each angle used for primary projections.

[0050] All validation simulations were run with a source axis distance (SAD) of 1 m and a source-detector distance (SDD) of 1.5 m. The detector was modelled after a CsI scintillating PaxScan 4030CB detector (Varex Imaging Corporation, Salt Lake City, UT), with a pixel binning of 4 by 4 for increased simulation efficiency. The detector size simulated was 39.73 cm by 39.73 cm with 512 by 512 pixels with 0.776 mm pixel size. All attenuation coefficients used in simulations were obtained from NIST XCOM. The bowtie filter was modelled after a TrueBeam full-fan bowtie filter. For GGEMS secondary simulations, the beam was collimated to a 26.8 cm square at isocenter, an energy cut of 1 keV, and a range cut of 0.1 mm was used for photons in all simulations, Compton, Photoelectric, and Rayleigh processes were activated in all regions.

[0051] A CTDI head phantom (Sun Nuclear Corp., Melbourne, Florida) was imaged using a narrow beam (0.5 cm Y-jaw collimation) to validate the scatter modeling in GECCO. The CBCT scan of the phantom was reconstructed, and a corresponding GECCO model was created. A single projection through the phantom was simulated using the same collimation settings as in the experiment. A region containing only scattered radiation was identified, and the simulated and experimental images were compared over an ROI in this area. In this setup the ROI was taken at 5 cm below the slit to avoid signal from the narrow beam. Both images were detrended and smoothed with a Gaussian filter to suppress noise.

[0052] An XCAT anthropomorphic head phantom was used to validate GECCO on phantom data. The XCAT phantom was simulated using GGEMS and GECCO, and primary projections were compared. For validation, the mean squared error was compared between GGEMS and GECCO projections. The phantom had dimensions of 86 by 512 by 512 voxelwith 3.125 slice thickness and 0.5 mm axial dimensions. Elemental compositions of XCAT materials were defined using equivalent materials from the ICRP tissue database.44,45 The spectrum used in both simulations was simulated as an x-ray tube in Spekpy with a peak voltage of 100 kV and an anode angle of 12 degrees; the spectrum was filtered by 4 mm aluminum. The GECCO projection contained 1011primary photons.

[0053] For sensitometry evaluation, the CTP404 module from a Catphan 504 phantom (The Phantom Laboratories, Salem, NY) was utilized. This 20 cm diameter sensitometry section contains 1.2 cm cylindrical inserts composed of air, acrylic, low-density polyethylene (LDPE), Teflon, polystyrene, Delrin, and polymethylpentene (PMP). The phantom was carefully aligned such that the central axis of the CTP404 module coincided with the isocenter of the imaging system. A digital phantom was created with the same dimensions and compositions and used to compare experimental and simulated CBCTs. For validation, HU values were compared between the GECCO simulation and the experimental acquisition using a TrueBeam on-board imager (OBI). The virtual phantom had dimensions 400 by 526 by 526 voxels with isotropic voxel dimensions of 0.5 mm. Slightly different sensitometry modules were used between GECCO and the experimental images; thus, not all inserts were present in both phantoms.

[0054] A total of 915 projections were acquired / simulated to create both images. Experimental images were acquired on a Truebeam STx linac with a Varian GS-1542 x-ray tube and a 100 kVp x-ray beam. The spectra were filtered by 2.7 mm aluminum inherent filtration as well as a 0.89 mm titanium beam hardening filter. Projections were exported to .xim files in developer mode on the treatment machine. Both sets of images were reconstructed manually in Python with the FDK algorithm with a ram-lak filter, reconstructed images had dimensions of 400 by 526 by 526 voxels with isotropic voxel dimensions of 0.5 mm.

[0055] Registered CBCT and CT volumes from an anonymized central nervous system (CNS) patient were used to evaluate GECCO’s conversion accuracy. A CNS patient was used as the face mask provides a good chance of registering the CT to CBCT of any anatomical site, a factor that can sometimes confound comparison. The CBCT volume was rigidly registered to the CT volume in the 3D Slicer application. The CT volume was acquired on a GE LightSpeed RT16 CT scanner with a tube voltage of 120 kVp and reconstructed using standard reconstruction settings, with a 1.25 mm slice thickness, 0.76 mm pixel spacing, and a soft- tissue reconstruction kernel. The CBCT was acquired on a Truebeam treatment machine withan x-ray tube peak voltage of 100 kV; the image was reconstructed using the standard Varian Truebeam reconstructor, with the FDK algorithm and a ram-lak filter, with a 1 mm slice thickness and 0.51 mm pixel spacing. To assure consistency, the CBCT was linearly interpolated to 0.76 mm pixel spacing and 1.25 mm slice thickness, then Gaussian filtered with a 1 -pixel standard deviation kernel.

[0056] A GECCO simulation with 915 primary and 10 secondary projections with 1010initial histories per projection, a full-fan bowtie filter, and the Truebeam treatment geometry were used to convert the CT to CBCT. For validation profiles through the phantom were compared between the GECCO-simulated CBCT and the patient CBCT. Profiles through the two volumes were compared, disregarding values below -250 HU as low values are removed in TrueBeam postprocessing.

[0057] The RTIS application behaves and produces simulations similar to the primary component of the GECCO application. However, RTIS simplifies and speeds up the analytical calculations for maximum computational efficiency, operating on only one slice of the planning CT. RTIS updates images in real-time according to user-specified kVp, mAs, and scatter, to allow radiation therapists and researchers to evaluate CBCT imaging techniques for a given patient before image acquisition. For validation, profiles through a CNS patient CBCT were compared to RTIS as described below.

[0058] Major approximations of the CBCT process included reducing the number of secondary projections and using a method of material decomposition for elemental compositions. In light of these approximations, one goal was to keep GECCO and RTIS simulation errors within 30 HU of experimental values in soft tissue and 50 HU of experimental values in bony anatomy following the AAPM TG-179 recommendation for kV-CBCT accuracy. Some simulation results in GECCO have been previously validated using Fastcat. For the kV-OBI, contrast agreed within 14 HUs and detector MTF agreed within 4.2%. The contrast to noise ratio (CNR) had a root mean squared error (RMSE) of 2.6% in Catphan sensitometry modules as compared to experimental data from a Truebeam OBI.

[0059] HU accuracy was validated in a Catphan sensitometry module against an experimentally acquired scan of a Catphan module. FIG. 2 illustrates this validation. In the upper-left image of FIG. 2, a virtual Catphan sensitometry module was simulated with GECCO. In the upper-right image, a 100 kVp scan of a Catphan sensitometry module was acquired on a TruBeam OBI. In the graph, HU values are compared between the two phantoms for the labeled inserts. GECCO had an RMSE of 5.7 HU compared to the TrueBeam OBI. Thelargest RMSE was found in the PMP at 20 HU followed by Teflon, acrylic, LDPE, and polystyrene, with errors of 3.6, 2.7, 1.8, and 0.6 HU, respectively. This validation agreed very closely, with similar results to Fastcat which had a largest RMSE of 14 HU and an average RMSE of 6.9 HU. The validation is expected to be similar between Fastcat and GECCO as they use similar analytical simulations with different scatter simulation methods. There is also some variability in HU values for a Catphan sensitometry module experimentally, with errors within 30-50 HU being acceptable. Therefore, the error of within 20 HU shown here is quite acceptable.

[0060] Primary projections were compared of the XCAT phantom between GGEMS primary projections and the GECCO primary projections (these results equally apply to RTIS as well). RMSE between the two projections was 0.2 %, while RMSE was slightly higher (0.5 %) in an ROI of cortical bone.

[0061] Further, GGEMS tended to have slightly higher attenuation coefficients than GECCO in the area where the bowtie has the highest attenuation. This can be seen in FIG. 3, which shows a comparison of the MC projections, as images with 512 by 512, 0.776 mm pixels, of an XCAT phantom generated using GGEMs (at left) and those generated using GECCO (at center). The difference between the two images is also shown (at right) after Gaussian filtration to remove noise. The RMSE was 0.25 % between the GGEMs and GECCO projections. The GGEMs projection was calculated using 1010initial photons. A full-fan bowtie was simulated in both cases, energy deposition was recorded in a 39.73 cm by 39.73 cm CsI detector in both cases and normalized by a simulated air energy deposition. The difference in attenuation coefficients was attributed to the different handling of the bowtie filter in the two simulations. GGEMS has a limited bowtie resolution of one aluminum thickness per degree, while GECCO has an arbitrary angular resolution of the bowtie. To avoid artifacts in the reconstruction, GECCO uses a high-resolution bowtie. Consequently, there is a slight difference in handling bowties, resulting in an error in the primary projections between GECCO and GGEMS. This error is non-existent at the bowtie periphery and peaks with GECCO attenuation coefficients being 0.4 % smaller at the center of the bowtie.

[0062] Particular implementations of GECCO may balance trade-offs between simulation speed and accuracy. In the above-described case, a less than 0.4 % error at the center of the bowtie was estimated to be acceptable for the application of CT to CBCT machine learning. This systematic error could be corrected in the simulation method empirically.

[0063] FIG. 4 shows narrow-beam CTDI phantom scatter validation results. The image in the upper-left is an experimental image of the CTDI phantom, whereas the image in the upper-right is a GECCO simulated image of the CTDI phantom. The graph shows a comparison of profiles through the ROI (present as a horizontal line in the images) averaged in the y- direction. As can be seen from FIG. 4, in a narrow beam image of a CTDI phantom, profiles in a scatter-dominated region showed agreement within 7% of the maximum scatter intensity in ROI with a RMSE of 1.1%. Overall, the scatter distributions had local irregularities that may present modeling difficulties. Some explanations for these local irregularities might be backscatter from detector electronics, gain and offset variations, or inaccuracies in the elemental decomposition algorithm with respect to the treatment couch.

[0064] A GECCO-simulated CBCT reconstruction created using a CNS patient CT scan was compared to a first-week CBCT scan of the same patient, as shown in FIG. 5. For the upper-left image of FIG. 5, a stereotactic brain CT scan was converted to elemental compositions and used as input into the GECCO application to generate a CBCT volume. For the upper-right image, a registered 100 kVp CBCT of the same patient from the TrueBeam OBI was used for validation through comparisons on vertical and horizontal profiles. The lower-left and lower-right graphs show these comparisons, respectively. There was an average error of 14 HU between GECCO and OBI profiles, with some of this error combing from image registration and the material decomposition of the CT. Thus, qualitatively, the images look quite similar, while quantitatively, there are a few differences arising from the material decomposition process, the GECCO simulation modelling accuracy, as well as inherent limitations of generating CBCTs from CT data with a finite resolution and image artifacts.

[0065] FIG. 6 shows an RTIS-simulated CBCT reconstruction created using a CNS patient CT scan as compared to a first-week CBCT scan of the same patient. The upper-left image of FIG. 6 corresponds to a stereotactic brain CT scan that was converted to elemental compositions and used as input into the RTIS simulation tool to generate a CBCT image. In the upper-right image, a registered 100 kVp CBCT of the same patient from the TrueBeam OBI is shown. Two horizontal profiles and one vertical profile are also shown for comparison, with the center-left graph corresponding to the lower horizontal line in the images, the centerright graph corresponding to the vertical line in the images, and the lower-left graph corresponding to the upper horizontal line in the images. There was an average error of 4.0Abetween the profiles, with some of this error coming from image registration and the material decomposition of the CT. As above, substantial effort was made to simulate the exactspecifications of the imaging system; however, some variability in HU values between linacs may contribute to the observed discrepancy. Additionally, a non-uniformity artifact, likely caused by photon starvation due to the metal components of the face mask assembly, was present in the CBCT but absent in the RTIS simulation, as the simulation was based on the planning CT, where these metal objects were outside the field of view.

[0066] Profiles through the CNS patient had an average RMSE of 14 HU and 16 HU between GECCO and the OBI, and RTIS and the OBI, respectively. The agreement between RTIS and GECCO indicated that a uniform scatter approximation was sufficient for simulating this patient. Care was taken to simulate the exact specifications of the imaging system; however, there remains some variability of HU values between linacs, which is a possible source of the discrepancy. This may be due to the method of converting HU to elemental composition, where the poor soft tissue contrast in CT causes many tissues of different compositions to have HU values near that of water. This can be resolved using dual- or multienergy CT to allow for more complete differentiation of tissues; however, single-energy CT data is more readily available clinically and thus the focus of this explanatory example. Likewise, the non-uniformity artifacts in the CBCT were not completely present in the GECCO simulation as can be seen at pixel 250 in the vertical profile. Substantial effort was made to simulate the exact specifications of the imaging system; however, non-uniformity artifacts, sometimes caused by photon starvation due to the metal components of the face mask assembly, were present in the CBCT but absent in the GECCO and RTIS simulations, as the simulations were based on the planning CT, where these metal objects were outside the field of view. Likewise, there is a region of non-uniformity surrounding the eyes in the GECCO simulation that is not present in the CBCT. This artifact is not found in the RTIS simulation, indicating that it is related to differences in the scatter modelling in GECCO.

[0067] The spatial resolution and pixel size of the CT reconstruction may limit the spatial resolution of the CBCT. The CNS patient was reconstructed with a soft tissue kernel and a pixel spacing of 0.76 mm, while the CBCT image was reconstructed with a pixel spacing of 0.51 mm and a ram-lak kernel. Because the spatial resolution of the CT scan was already degraded by the reconstruction kernel, no detector optical spread function kernel was used to further reduce spatial resolution, and a ram-lak filter was used to preserve the spatial resolution. However, as can be seen in the inset of the images in FIG. 5, the spatial resolution of the GECCO simulation is qualitatively worse than the OBI spatial resolution. To address this, clinical workflows could be modified to include higher spatial resolution CT scans. This mayalso be addressed by augmenting data using smoothing and deconvolution, respectively, to degrade and sharpen the input data so that the ML model does not overfit to low-resolution data. Likewise, in the CNS patient, there were differences in the postprocessing between GECCO and clinical images. Unlike for the Catphan, clinical images could not be acquired in developer mode for a patient. Thus, the volume has been processed with the Varian beam hardening correction and scatter reduction algorithm. The GECCO scatter was reduced manually to match the scatter reduction, however, no beam hardening correction was implemented in GECCO. It is estimated that this did not have a huge effect on accuracy as it is believed that the largest errors in accuracy come from the CT to CBCT conversion process and registration of the CT and CBCT images.

[0068] GECCO and RTIS compare favorably to other simulation tools, as shown in Table 1, as they have capabilities and validation that other applications lack. GECCO and RTIS are have been validated using patient data. Likewise, GECCO and RTIS are the capable of simulating the Truebeam half-fan geometry. Finally, with a 36 ms simulation time on a single Nvidia GPU, RTIS performs simulations three orders of magnitude faster than any comparative application.TABLE 1

[0069] The RTIS application may have additional advantages due to its rapid processing speed. Initialization of a simulation on an Nvidia 4050 laptop GPU was completed in 917 ms, with subsequent kVp, mAs, and scatter updates taking only 36 ms, enabling realtime evaluation of imaging acquisition techniques. This is in contrast to comparative approaches, which would require minutes to update on a CPU or, in Fastcat’s case, 71 seconds to simulate a CBCT on GPU. For an MC application, GECCO is also fast, taking 370 seconds per GGEM projection and 63 minutes to simulate for 915 full-resolution projection CNS patients.

[0070] Unlike the baseline GECCO application, which loads the full CT volume, RTIS uses only a single slice. This slice is converted to elemental compositions, and only analytical projections are calculated, with a constant scatter approximation used to estimate secondary projections. Further processing and reconstruction may be performed similarly to the full GECCO application.

[0071] While the GECCO application generates a spectrum analytically from Spekpy or Xpecgen for each new simulation, RTIS uses pre-calculated spectra stored in a spectrumdatabase for faster calculation. Spekpy analytical spectrum generation takes 233 ms for a 140 kVp, 14-degree anode x-ray tube. To enable quicker simulation, RTIS contains pre-calculates analytical spectra in increments of 1 keV from 60-140 kVp, filtered with 2.7 mm aluminum inherent filtration and 0.89 mm titanium beam hardening filter in a database. Additionally, user-specified spectra with other kVp or filtration can be saved to the RTIS spectrum database.

[0072] RTIS can be updated dynamically. RTIS uses the same default mono-energetic energies used for raytracing, and the mono-energetic projections are combined into a poly- energetic simulation. However, the GECCO application does not keep the mono-energetic projection data from each ray tracing. This approach is prohibitively memory intensive for a full CBCT scan; a scan can create up to 18 mono-energetic sets of projection data, each containing around 900 projections with dimensions of up to 1024 x 1024 pixels using 32-bit floats, would consume 67.95 GB of memory. Thus, in Fastcat, to save memory, mono-energetic calculations are added to a partial calculation of the poly-energetic projections and deleted at the end of each mono-energetic calculation step. However, storing the mono-energetic calculations for the single slice used in RTIS is not memory intensive, as the projection data is of a dimension of up to 1 x 1024 pixels.

[0073] By storing the mono-energetic calculations, to update the simulation, one does not have to ray-trace; one has to create a new weighting of the mono-energetic simulations according to the detector response and x-ray tube spectrum. Poisson noise is also added to the projections data according to user-defined mAs. The number of photons per mAs is derived from Spekpy, which generates spectral fluence in photons per mAs per unit energy per unit area [photons / cm2 / mAs / keV], From this fluence, the average number of photons per pixel is derived for the flood field. Poisson statistics are then used to calculate the noise in each pixel of the projection data. Scatter can also be added. To model scatter, a constant amount of background counts are added to the projections, and the user specifies the number of background counts as a percentage of the mean number of photons per pixel.

[0074] Projections can then be convolved with a Gaussian kernel with a FWHM equal to the x-ray tube focal spot and the detector’s optical spread function (precalculated from Fastcat). Projections are log-normalized using the flood field and reconstructed using the Feldkamp, Davis, and Kress (FDK) algorithm48 implemented in the TIGRE reconstruction toolbox. The user interface for RTIS is accessed through a GUI with a slider for kVp, mAs, and percent scatter for ease of use.

[0075] An RTIS simulation with 915 projections, a full-fan Varian bowtie filter, and the Truebeam treatment geometry was used to convert the CT to CBCT. Validation profiles through the phantom were compared between the RTIS-simulated CBCT and the patient CBCT. Additionally, a 2.5 MV RTIS simulation was also simulated. The RTIS model used a beam energy spectrum created in EGSnrc from colliding a 2.5 MeV electron beam with a 2.3 mm Tungsten target. The detector modelled was a Varian asl200 GOS detector with a pixel pitch of 0.784 mm. All simulations were modelled to have 1011initial photons per projection.

[0076] FIG. 7 illustrates the relative differences in a MV and two kV CBCTs, illustrating the potential of using RTIS to identify cases where an MV CBCT might be suitable over a kV CBCT. In FIG. 7, scatter-free reconstructions are shown in the CNS patient for three different settings: an 80 kVp CBCT showcasing increased tissue contrast, a 140 kVp CBCT with reduced metal artifacts, and a demonstration of a 2.5 MV CBCT with reduced metal artifacts and increased noise. Furthermore, the fast updating capability allows for the creation of a GUI that loads a slice of the CT volume, enabling radiation therapists to evaluate the change in the CBCT for different available techniques in real-time, thereby assessing techniques that are appropriate to the patient anatomy. This may be particularly useful in cases involving large patients or irregular anatomies, potentially preventing the need for CBCT reacquisition due to poor image quality in the initial image and thus sparing patient dose and improving patient throughput. This can also facilitate the development of more specific imaging settings tailored to anatomical sites, rather than relying on the few generalized settings available on clinical treatment machines.

[0077] FIG. 8 illustrates one example of the use of the RTIS application to evaluate changes in CBCT settings in real-time. In head and neck imaging, it can be challenging to choose between protocols such as the standard 100 kVp, 75 mAs “Head Standard” setting (shown at left) and higher-energy options that may better visualize the shoulders. RTIS addresses this by determining preferred imaging parameters — here, 125 kVp and 300 mAs (shown at right). Moreover, the preferred imaging parameters may be determined prior to image acquisition, thereby improving image quality while reducing trial-and-error scans, saving both time and patient does.

[0078] While the noise and scatter in the simulations may not be calculated as well as in MC simulation methods, RTIS has real-time updating capabilities that other simulation methods lack. For a therapist trying to quickly assess techniques, RTIS may be especially useful.

[0079] FIG. 9 illustrates an example of the simulation workflow structure, with several files and methods (e.g., among those illustrated in FIG. 1) outlined. The RTIS Python script specifies the initial parameters, including CBCT geometry, spectrum information, and patient CT slice. The RTIS data initialization module coverts the patient CT volume and parameters in files for primary and secondary simulation, including attenuation data, CT density, spectrum data, CT range, and CT .mhd files. These files are also stored along with present files, including the bowtie filter file and detector offset file. RTIS simulations are then run using the initialized files and preset files, and mono-energetic simulations are stored in memory. A GUI may then be opened (or may already be open) to permit an operator to update the simulation settings dynamically. These updated parameters are used to combine the mono-energetic simulations into a poly-energetic simulation, with noise and scatter, and reconstructed.

[0080] In examples, the methodology set forth herein may be implemented as a method of simulating a CBCT scan. FIG. 10 illustrates an example method 1000. In examples, the method 1000 may be used to implement the RTIS application described above. The method 1000 includes an operation 1002 of receiving a set of initial CBCT scan parameters. These parameters may include geometry parameters, spectrum information parameters, CT slice parameters, and the like (e.g., as shown in FIG. 9).

[0081] At operation 1004, a mono-energetic simulation is generated using the initial parameters. Operation 1004 may include converting the parameters into files to be stored in memory for use in primary and secondary simulations. Operation 1004 may further include generating the primary and secondary simulation files themselves, for example using the stored files and preset files from the RTIS database. At operation 1006, mono-energetic simulations are then stored in memory (e.g., as a slice). The mono-energetic simulations may be presented to a user, for example via a GUI. To present the simulations, an image may be generated corresponding to the mono-energetic simulation (e.g., to the slice) and displayed to the user on the GUI. The user may then enter updated scan parameters (e.g., an updated kVp parameter, an updated mAs parameter, an updated scatter parameter, etc.), for example via the same GUI. Using the updated parameters, at operation 1010 the stored slice may be modified using the updated parameters.

[0082] Operation 1010 may include converting the stored slice into a set of elemental compositions, determining a set of primary projections using pre-calculated spectra (e.g., stored in the RTIS database), and estimating a set of secondary projections using a constant scatter approximation. In any example, operation 1010 may include re-weighting the stored sliceaccording to the updated parameters. Thus, the method 1000 need not perform computationally intensive operations such as ray tracing for each simulation. Operation 1010 may further include post-processing or filtering operations (e.g., applying a bowtie filter). Thereafter, the updated simulation may be presented to the user, for example via the same GUI. As with operation 1006, presenting the updated simulation may include generating an updated image corresponding to the mono-energetic simulation as modified in operation 1010, and displaying the updated image to the user via the GUI. This process may be repeated any number of times, thereby to permit the user to input several sets of parameters and ultimately select the preferred set. The updating of the simulation and presentation of the updated results may be performed in real-time. In some examples, a plurality of mono-energetic simulations may be combined into a poly-energetic simulation.

[0083] In an example, the method 1000 (or any of the techniques set forth herein) may be implemented with the use of a computing device, such as the computing device 1100 illustrated in FIG. 11. As illustrated the device 1100 includes at least one processor 1102, a memory 1104, and a set of interface components 1106. The memory 1104 may store a set of instructions (e.g., as a non-transitory computer-readable medium) that are executable by the processor 1102 to perform sets of operations, including the operations described above. In one particular example, the set of operations includes the method 1000 of FIG. 10. The interface components 1106 may provide an interface through which a user may interact with the device 1100. For example, the interface components 1106 may permit the device 1100 to receive communication from a user via an input device (e.g., a mouse, a keyboard, a microphone, an image or motion sensor, a stylus, a touch screen, etc.) and to present information to the user via an output device (e.g., a display, a speaker, a haptic feedback device, etc.). In examples, the interface components 1106 may include a GUI displayed on a display.

[0084] The GECCO application methodology, and its implementations as both GECCO and RTIS, have applications in the radiotherapy landscape. Adaptive radiotherapy is putting pressure on CBCT to fill clinical roles traditionally held by CT, and machine learning is seen as the methodology to facilitate the improvement of CBCT. Models can be trained on registered pairs of CT / CBCT images, which has the advantage of the CBCT being perfectly accurate. However, there will always be changes in the anatomy and errors in the registration between the two volumes. Likewise, anthropomorphic phantoms can be used to generate simulated CBCT images with a complete ground truth rather than a CT, but to give a completedataset, one would need a large variety of these models, and the complexity of tissue heterogeneity remains difficult to model with anthropomorphic phantoms.

[0085] Thus, the GECCO application, a fast MC approach using CT data to model CBCT from CT, can also be used on anthropomorphic phantoms. One benefit of generating CBCT from CT is the large availability of comprehensive CT datasets, whereas registered CT / CBCT datasets are relatively rare, and anthropomorphic phantom datasets have limited variability compared to human data. Large datasets of GECCO-simulated CBCT / CT pairs can be generated, providing advantages as data availability is often a limiting factor in the generalizability of machine learning methods in radiation oncology.

[0086] Furthermore, the 36 ms updating capability of RTIS allows for creating a GUI that loads a slice of the CT volume, enabling radiation therapists to evaluate the change in the CBCT for different available techniques in real-time, thereby assessing techniques that best display the patient’s anatomy. This is particularly useful in cases involving large patients or irregular anatomies, potentially preventing the need for CBCT reacquisition due to poor image quality in the initial image, thus sparing patient dose and improving patient throughput. Additionally, this approach could facilitate the development of more specific imaging settings tailored to anatomical sites rather than relying on the few generalized settings available on clinical treatment machines.

[0087] As set forth above, the present disclosure introduces GECCO, a hybrid MC toolkit for fast CBCT generation containing the GECCO and RTIS applications. GECCO analytical primary projections showed an RMSE of 0.25% of MC, while HU values had an RMSE of 5.7 HU as compared to those acquired experimentally on a Truebeam OBI. CBCTs generated from CT took 63 minutes with RMSE of 14 HU in profiles through the phantom as compared to registered CBCTs. Likewise, RTIS simulations of one slice of the same CNS patient had an error of 16 HU along a horizontal profile and took 917 ms to initialize and 36 ms to update mAs, beam filtration, and kVp.

[0088] Other examples and uses of the disclosed technology will be apparent to those having ordinary skill in the art upon consideration of the specification and practice of the invention disclosed herein. The specification and examples given should be considered exemplary only, and it is contemplated that the appended claims will cover any other such embodiments or modifications as fall within the true scope of the invention.

[0089] The Abstract accompanying this specification is provided to enable the United States Patent and Trademark Office and the public generally to determine quickly from acursory inspection the nature and gist of the technical disclosure and in no way intended for defining, determining, or limiting the present invention or any of its embodiments.

Claims

CLAIMSWhat is claimed is:

1. A method of simulating a cone-beam computed tomography (CBCT) scan, the method comprising: receiving a set of initial CBCT scan parameters; generating a mono-energetic simulation of the CBCT scan using the initial CBCT scan parameters; storing a slice of the mono-energetic simulation in memory; receiving a set of updated CBCT scan parameters; and modifying the stored slice of the mono-energetic simulation based on the updated CBCT scan parameters.

2. The method of claim 1, wherein operation of modifying the stored slice of the mono-energetic simulation is performed in real-time.

3. The method of claim 1 or claim 2, further comprising: generating an image corresponding to the slice of the mono-energetic simulation; displaying the image to a user via a graphical user interface (GUI); receiving the set of updated CBCT scan parameters via the GUI; generating an updated image corresponding to the modified slice of the mono-energetic simulation; and displaying the updated image to the user via the GUI.

4. The method of any one of claims 1 to 3, wherein modifying the stored slice of the mono-energetic simulation includes: converting the stored slice into a set of elemental compositions; determining a set of primary projections using pre-calculated spectra stored in a spectrum database; and estimating a set of secondary projections using a constant scatter approximation.

5. The method of claim 4, wherein modifying the stored slice of the mono- energetic simulation further includes applying a bowtie filter to the set of primary projections and the set of secondary projections.

6. The method of any one of claims 1 to 5, further comprising combining a plurality of the mono-energetic simulations into a poly-energetic simulation.

7. The method of any one of claims 1 to 6, wherein modifying the stored slice of the mono-energetic simulation includes re-weighting the stored slice of the mono-energetic simulation according to the updated CBCT scan parameters.

8. The method of any one of claims 1 to 7, wherein the updated CBCT scan parameters include at least one of an updated peak voltage (kVp) parameter, an updated tube current-exposure time (mAs) parameter, and an updated percent scatter parameter.

9. A computing device, comprising: a display configured to present a graphical user interface (GUI); a memory; and at least one processor configured to execute instructions, thereby causing the computing device to perform operations comprising: receiving a set of initial cone-beam computed tomography (CBCT) scan parameters, generating a mono-energetic simulation of the CBCT scan using the initial CBCT scan parameters, storing a slice of the mono-energetic simulation in the memory, generating an image corresponding to the slice of the mono-energetic simulation, displaying the image via the GUI, receiving a set of updated CBCT scan parameters via the GUI, and modifying the stored slice of the mono-energetic simulation based on the updated CBCT scan parameters.

10. The computing device of claim 9, the operations further comprising:generating an updated image corresponding to the modified slice of the mono-energetic simulation; and displaying the updated image to the user via the GUI.

11. The computing device of claim 9 or claim 10, wherein generating the mono- energetic simulation includes applying a ray-tracing method.

12. The computing device of any one of claims 9 to 11, wherein modifying the stored slice of the mono-energetic simulation further includes applying a bowtie filter to the set of primary projections and the set of secondary projections.

13. The computing device of any one of claims 9 to 12, the operations further comprising combining a plurality of the mono-energetic simulations into a poly-energetic simulation.

14. The computing device of any one of claims 9 to 13, wherein modifying the stored slice of the mono-energetic simulation includes re-weighting the stored slice of the mono-energetic simulation according to the updated CBCT scan parameters.

15. The computing device of any one of claims 9 to 14, wherein the updated CBCT scan parameters include at least one of an updated peak voltage (kVp) parameter, an updated tube current-exposure time (mAs) parameter, and an updated percent scatter parameter.

16. The computing device of any one of claims 9 to 15, the operations further comprising repeating the operations of receiving the set of updated CBCT scan parameters and modifying the stored slice of the mono-energetic simulation until a user approval indication is received.

17. A non-transitory computer-readable medium storing instructions that, when executed by at least one processor of a cone-beam computed tomography (CBCT) simulation system, cause the system to perform operations comprising: receiving a set of initial CBCT scan parameters; generating a mono-energetic simulation of the CBCT scan using the initial CBCT scan parameters;storing a slice of the mono-energetic simulation in memory; receiving a set of updated CBCT scan parameters; and modifying the stored slice of the mono-energetic simulation based on the updated CBCT scan parameters.

18. The non-transitory computer-readable medium of claim 17, the operations further comprising: displaying an image corresponding to the slice of the mono-energetic simulation to a user via a graphical user interface (GUI); receiving the set of updated CBCT scan parameters via the GUI; and displaying an updated image corresponding to the modified slice of the mono-energetic simulation to the user via the GUI.

19. The non-transitory computer-readable medium of claim 17 or claim 18, wherein the operation of modifying the stored slice of the mono-energetic simulation includes: converting the stored slice into a set of elemental compositions; determining a set of primary projections using pre-calculated spectra stored in a spectrum database; and estimating a set of secondary projections using a constant scatter approximation.

20. The non-transitory computer-readable medium of claim 19, wherein the operation of modifying the stored slice of the mono-energetic simulation further includes applying a bowtie filter to the set of primary projections and the set of secondary projections.

Images