Simulation apparatus for reproducing 3-dimensional image using artificial intelligence and operating method thereof
The hybrid simulation method addresses CBCT image distortion by combining Monte Carlo and Forward Projection, achieving fast and accurate 3D image generation for AI training, overcoming computational bottlenecks and ethical data constraints.
Patent Information
- Application Number
- KR1020250198915
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-12-15
AI Technical Summary
Existing Cone Beam Computed Tomography (CBCT) systems suffer from reduced image contrast and distortion due to scattered photons, and conventional simulators either lack physical accuracy or are computationally inefficient, while Monte Carlo simulations are too slow for practical use.
A hybrid simulation method combining Monte Carlo and Forward Projection simulations, utilizing a reinforcement learning model to optimize grid specification and noise removal, enabling fast and accurate generation of high-resolution 3D medical images.
This approach significantly reduces computation time while maintaining physical accuracy, allowing for efficient simulation of scattering phenomena and generating hyper-realistic medical images suitable for training AI models without ethical constraints.
Smart Images

Figure 112025141385238-PAT00003_ABST
Abstract
Description
Technology Field
[0001] The present disclosure relates to a simulation device for improving three-dimensional images without a grid using artificial intelligence and a method of operating the same. Background Technology
[0003] In the field of medical imaging diagnostics, a Cone Beam Computed Tomography (CBCT) system may be a device that acquires three-dimensional images of a subject by irradiating X-rays in the form of a broad cone beam. However, problems may arise in such CBCT systems where scattered photons reach the detector due to scattering phenomena that occur as the X-rays interact with materials inside the human body or objects during the irradiation process. Consequently, the acquired images may suffer from reduced contrast or distortion of Hounsfield Unit (HU) values, and various artifacts such as cupping or streaks may be induced, which can cause a decrease in diagnostic accuracy.
[0004] Conventionally, clinical patient data has been utilized to analyze these scattering phenomena or to train artificial intelligence (AI) models; however, it may be difficult to construct a sufficient dataset due to safety concerns regarding radiation exposure and ethical constraints such as the protection of personal information. As an alternative, simulators that generate images in a virtual environment are being used; however, simple linear transmission (ray-tracing) based simulators have limitations in that they cannot accurately reproduce the physical characteristics of actual images, as they ignore complex scattering effects or apply only simple empirical corrections, despite their fast computational speed.
[0005] Meanwhile, Monte Carlo simulations, which probabilistically calculate the interaction between photons and matter, can provide high reliability as they can physically simulate Compton scattering and photoelectric effects with high precision. However, since the Monte Carlo method requires individually tracking collision locations, scattering angles, and energy losses for hundreds of millions of photons, the computational load is massive, which can lead to fundamental bottlenecks in processing speed. For example, simulating a single scan takes several hours, and obtaining hundreds of scan data required for CT reconstruction takes several days, presenting inefficiencies. Therefore, there is a need for technology that can dramatically improve computational speed while maintaining physical accuracy. means of solving the problem
[0007] A method of operation of a simulation device including a processor and a memory according to one embodiment of the present disclosure comprises the steps of: the processor acquiring input grid specification information based on instructions stored in memory; acquiring an input radiation image corresponding to the input grid specification information; applying state information including the input grid specification information and the input radiation image to a grid specification determination reinforcement learning model to acquire output grid specification information; acquiring an output radiation image based on the output grid specification information; and outputting the output radiation image.
[0008] A method of operation of a simulation device according to one embodiment of the present disclosure further comprises the steps of: acquiring learning input grid specification information; acquiring a learning input radiation image corresponding to the learning input grid specification information; applying learning state information including the learning input grid specification information and the learning input radiation image to a grid specification determination reinforcement learning model to acquire learning output grid specification information; acquiring a learning output radiation image based on the learning output grid specification information; acquiring noise information related to the degree of noise appearing in the learning output radiation image based on the learning output radiation image; acquiring a higher reward score as the noise information is smaller; and training a grid specification determination reinforcement learning model to maximize the reward score.
[0009] A step of acquiring an output radiation image based on output grid specification information of a method for operating a simulation device according to one embodiment of the present disclosure includes: acquiring information of a virtual subject including at least one of the material, shape, and component information of the subject; acquiring radiation source information related to at least one of the location or intensity of a virtual radiation source; acquiring detector information for a detector receiving radiation from a virtual radiation source; and acquiring an output radiation image through radiation image simulation based on output grid specification information, information of the subject, radiation source information, and detector information.
[0010] The radiation imaging simulation of the operation method of the simulation device according to one embodiment of the present disclosure is based on at least one of Monte Carlo and Forward Projection.
[0011] A method of operation of a simulation device according to one embodiment of the present disclosure further includes the steps of: acquiring an actual radiation image by capturing an actual radiation image based on information of a subject and detector information; acquiring actual noise level information of the actual radiation image; acquiring virtual noise level information of an output radiation image; and modifying radiation source information so that the virtual noise level information becomes identical to the actual noise level information.
[0012] The step of acquiring an output radiation image through radiation image simulation of a method of operation of a simulation device according to one embodiment of the present disclosure comprises: acquiring a first preliminary radiation image based on Monte Carlo at a predetermined first resolution; acquiring a second preliminary radiation image based on Forward Projection at a predetermined second resolution which is higher than the first resolution; acquiring a third preliminary radiation image by converting the second preliminary radiation image to a resolution lower than the second resolution; acquiring a fourth preliminary radiation image by resizing the image converted to the lower resolution back to the second resolution; acquiring a residual image containing high-frequency components based on the fourth preliminary radiation image and the second preliminary radiation image; acquiring a weighted average image by weighting the third preliminary radiation image and the first preliminary radiation image based on predetermined weights; and acquiring an output radiation image of the second resolution based on the weighted average image and the residual image. Includes steps.
[0013] The step of acquiring an output radiation image based on output grid specification information of the operation method of a simulation device according to one embodiment of the present disclosure includes the step of acquiring an output radiation image by applying input grid specification information, an input radiation image, and output grid specification information to an output image machine learning model.
[0014] A method of operation of a simulation device according to one embodiment of the present disclosure further comprises the steps of: acquiring learning input grid specification information; acquiring a learning input radiation image corresponding to the learning input grid specification information; acquiring learning output grid specification information different from the learning input grid specification information; acquiring a learning output radiation image corresponding to the learning output grid specification information; and generating an output image machine learning model by machine learning the correlation between the learning input grid specification information, the learning input radiation image, and the learning output radiation image with respect to the learning output grid specification information.
[0015] A method of operation of a simulation device including a processor and a memory according to an embodiment of the present disclosure comprises: a step in which the processor, based on instructions stored in memory, acquires information of a virtual subject including at least one of the material, shape, and component information of the subject; a step in which radiation source information related to at least one of the location or intensity of a virtual radiation source; a step in which detector information for a detector receiving radiation from a virtual radiation source is acquired; a step in which a first preliminary radiation image is acquired based on Monte Carlo at a predetermined first resolution based on the information of the subject, the radiation source information, and the detector information; a step in which a second preliminary radiation image is acquired based on Forward Projection at a predetermined second resolution which is higher than the first resolution based on the information of the subject, the radiation source information, and the detector information; a step in which the second preliminary radiation image is converted to a resolution lower than the second resolution to acquire a third preliminary radiation image; and a step in which the image converted to a lower resolution is resized back to the second resolution to obtain a fourth The method includes the steps of: acquiring a preliminary radiation image; acquiring a residual image containing high-frequency components based on the fourth preliminary radiation image and the second preliminary radiation image; acquiring a weighted average image by weighting the third preliminary radiation image and the first preliminary radiation image based on predetermined weights; and acquiring an output radiation image of a second resolution based on the weighted average image and the residual image.
[0016] A method of operation of a simulation device according to one embodiment of the present disclosure further includes the steps of: acquiring an actual radiation image by capturing an actual radiation image based on information of a subject and detector information; acquiring actual noise level information of the actual radiation image; acquiring virtual noise level information of an output radiation image; and modifying radiation source information so that the virtual noise level information becomes identical to the actual noise level information.
[0017] A method of operation of a simulation device according to one embodiment of the present disclosure further includes the step of generating a machine-learned scattering simulation machine learning model to receive a second preliminary radiation image and output an output radiation image.
[0018] A method of operation of a simulation device according to one embodiment of the present disclosure further includes the steps of: acquiring a plurality of output radiation images according to the position of the virtual radiation source, wherein the virtual radiation source is rotatable around a predetermined rotation axis, and acquiring a three-dimensional output radiation image based on the plurality of output radiation images.
[0019] A method of operation of a simulation device according to one embodiment of the present disclosure further includes the step of generating a machine-learned 3D scattering simulation machine learning model to receive a 3D preliminary radiation image and output a 3D output radiation image.
[0020] A method of operation of a simulation device according to one embodiment of the present disclosure further includes the steps of acquiring a plurality of second preliminary radiation images according to the location of a virtual radiation source, acquiring a three-dimensional preliminary radiation image based on the plurality of second preliminary radiation images, and generating a machine-learned three-dimensional noise removal machine learning model to receive a three-dimensional output radiation image and output a three-dimensional preliminary radiation image.
[0021] A method of operation of a simulation device according to one embodiment of the present disclosure further includes the step of receiving an output radiation image and generating a machine-learned noise removal machine learning model to output a second preliminary radiation image.
[0022] A method of operation of a simulation device including a processor and a memory according to one embodiment of the present disclosure includes the steps of: the processor acquiring an actual radiation image by irradiating an actual subject with radiation based on instructions stored in memory; and the processor acquiring an improved image with reduced noise by applying a noise removal machine learning model to the actual radiation image.
[0023] In addition, a program for implementing the operation method of the simulation device as described above can be recorded on a computer-readable recording medium. Effects of the invention
[0024] According to the simulation device and method of operation of the present disclosure, by applying a hybrid method that combines the physical accuracy of Monte Carlo simulation with the computational speed of forward projection simulation, the problem of excessive computation time associated with the conventional Monte Carlo method can be resolved, while simultaneously supplementing the accuracy of simulating scattering phenomena that was lacking in the forward projection method. Through this, high-resolution 3D medical images can be generated quickly, enabling efficient simulation even under large field of view (FOV) conditions, and securing highly reliable data at a level similar to actual clinical images.
[0025] Furthermore, simulations allow for the free construction of vast training datasets, enabling the acquisition of data necessary for training AI models without ethical constraints such as radiation exposure to actual patients or privacy protection. By refining machine learning models for scattering simulation, noise removal, or artifact correction based on this acquired data, high-quality images can be obtained even in low-dose imaging environments, and image distortion issues such as beam hardening or metal artifacts can be effectively improved through software.
[0026] Furthermore, by adjusting the photon count reaching the detector to reflect the noise characteristics of actual equipment and precisely synthesizing low-resolution scattering information with high-resolution geometric information using the image pyramid technique, hyper-realistic textures can be imparted to the simulation results. This can contribute to enabling the model to possess high adaptability to actual clinical environments when training sparse-view CT reconstruction or super-resolution models using simulation images.
[0027] However, the effects of the medical imaging device of the present disclosure are not limited to the above effects. Brief explanation of the drawing
[0028] FIG. 1 is a schematic diagram showing a simulation device according to one embodiment of the present disclosure. FIG. 2 is a drawing for explaining the configuration of a simulation system according to one embodiment of the present disclosure. FIG. 3 is a drawing for explaining the operation method of a simulation device according to one embodiment of the present disclosure. FIG. 4 is a drawing for explaining the operation method of a simulation device according to one embodiment of the present disclosure. FIG. 5 is a drawing for explaining the operation method of a simulation device according to one embodiment of the present disclosure. FIG. 6 is a drawing for explaining the effects of a simulation device according to one embodiment of the present disclosure. FIG. 7 is a drawing for explaining the effects of another simulation device in one embodiment of the present disclosure. FIG. 8 is a drawing showing a method of operation of a simulation device according to one embodiment of the present disclosure. FIG. 9 is a flowchart illustrating the operation method of a simulation device according to one embodiment of the present disclosure. FIG. 10 is a flowchart illustrating the operation method of a simulation device according to one embodiment of the present disclosure. FIG. 11 is a drawing showing a method of operation of a simulation device according to one embodiment of the present disclosure. Specific details for implementing the invention
[0029] The advantages and features of the disclosed embodiments and the methods for achieving them will become clear by referring to the embodiments described below in conjunction with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below but may be implemented in various different forms, and the embodiments are provided merely to make the present disclosure complete and to fully inform those skilled in the art of the scope of the invention.
[0030] The terms used in this specification will be briefly explained, and the disclosed embodiments will be described in detail.
[0031] The terms used in this specification have been selected to be as widely used and general as possible, taking into account their functions in the present disclosure; however, these may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been selected at the applicant's discretion, and in such cases, their meanings will be described in detail in the relevant description of the invention. Therefore, terms used in this disclosure should be defined not merely by their names, but based on their meanings and the content throughout the present disclosure.
[0032] In this specification, singular expressions include plural expressions unless the context clearly indicates that they are singular. Additionally, plural expressions include singular expressions unless the context clearly indicates that they are plural.
[0033] When a part of a specification is described as "including" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.
[0034] Additionally, the term “part” as used in the specification refers to a software or hardware component, and the “part” performs certain roles. However, the meaning of “part” is not limited to software or hardware. The “part” may be configured to reside in an addressable storage medium or configured to operate one or more processors. Thus, by example, the “part” includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and “parts” may be combined into a smaller number of components and “parts” or further separated into additional components and “parts.”
[0035] According to one embodiment of the present disclosure, the “part” may be implemented as a processor and memory. The term “processor” should be broadly interpreted to include a general-purpose processor, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a controller, a microcontroller, a state machine, etc. In some environments, “processor” may refer to an application-specific integrated circuit (ASIC), a programmable logic device (PLD), a field programmable gate array (FPGA), etc. The term “processor” may also refer to a combination of processing devices, such as, for example, a combination of a DSP and a microprocessor, a combination of multiple microprocessors, a combination of one or more microprocessors combined with a DSP core, or any other combination of such configurations.
[0036] The term "memory" should be interpreted broadly to include any electronic component capable of storing electronic information. The term memory may also refer to various types of processor-readable media, such as Random Access Memory (RAM), Read-Only Memory (ROM), Non-Volatile Random Access Memory (NVRAM), Programmable Read-Only Memory (PROM), Erasable-Programmable Read-Only Memory (EPROM), Electrically Erasable PROM (EEPROM), Flash Memory, Magnetic or Optical Data Storage Devices, Registers, etc. If a processor can read information from memory and / or write information to memory, memory is said to be in an electronic communication state with the processor. Memory integrated into a processor is in an electronic communication state with the processor.
[0037] Below, embodiments are described in detail with reference to the attached drawings so that those skilled in the art can easily implement the present disclosure. In addition, parts of the drawings that are irrelevant to the description are omitted to clearly explain the present disclosure.
[0038] FIG. 1 is a schematic diagram showing a simulation device according to one embodiment of the present disclosure.
[0039] A simulation device (100) is a computing device that numerically simulates and calculates physical phenomena of a medical imaging system in a virtual environment and may include a processor (110) and memory (120). The processor (110) may be a central processing unit (CPU) or a graphics processing unit (GPU) that controls the overall operation of the simulation device (100), and the memory (120) may be a storage medium that stores instructions, data, and machine learning models to be executed by the processor (110). The simulation device (100) may be a server, PC, tablet, or laptop, etc.
[0040] The processor (110) can perform operations based on a program or instruction stored in memory (120). For example, the processor (110) can perform a hybrid operation combining Monte Carlo simulation and Forward Projection simulation. The processor (110) can generate radiation projection data using at least one of information on a virtual subject, radiation source information, and detector information. The processor (110) may set the physical characteristics of the grid based on input grid specification information and generate radiation projection data using information on a virtual subject, radiation source information, and detector information. The processor (110) can subsequently generate a medical image of grid-applied quality using a medical image without an actual grid applied. At this time, the processor (110) can obtain scattering information by applying a Monte Carlo method at low resolution and geometric information by applying a forward projection method at high resolution to increase the computation speed, and then synthesize them through Image Pyramid Processing to obtain a final output radiation image.
[0041] The memory (120) can store various data and models required for this computation process. For example, the memory (120) may store weight parameters of artificial intelligence models such as a grid specification determination reinforcement learning model, a scattering simulation machine learning model, or a noise removal machine learning model. Additionally, the memory (120) may temporarily or permanently store intermediate results generated during the simulation process, such as first to fourth preliminary radiographic images, residual images, and weighted average images, so that the processor (110) can refer to them.
[0042] Through the configuration of such a simulation device (100), it is possible to efficiently process three-dimensional medical image simulations that require a vast amount of computation. In particular, through the organic combination of the processor (110) and the memory (120), the load of the Monte Carlo method, which requires calculating the behavior of hundreds of millions of photons, is reduced, while utilizing the fast speed of the forward projection method, a processing speed that can be utilized in actual clinical settings can be achieved. Furthermore, by continuously improving the machine learning model based on the simulation data accumulated in the memory (120), it is possible to provide the effect of improving the accuracy of image quality prediction.
[0043] FIG. 2 is a drawing for explaining the configuration of a simulation system according to one embodiment of the present disclosure.
[0044] The simulation system (200) may be configured to include a simulation device (100) that performs calculations and a user terminal (210) that provides an interface with a user. The simulation device (100) and the user terminal (210) may be connected to exchange data through a wired or wireless communication network. In some cases, the simulation device (100) may be implemented in the form of a server and the user terminal (210) may be implemented in the form of a client to provide cloud-based simulation services.
[0045] The user terminal (210) may be an input / output device that receives various parameters required for simulation from the user and visually displays the simulation results. For example, the user terminal (210) may receive or output at least one of information on the subject used in the simulation, radiation source information, or detector information. Additionally, the user may input input grid specification information including grid type, strip density, grid ratio, etc. through the user terminal (210). The simulation device (100) can satisfy the quality of the image based on the input grid specification information without a grid. For example, a noise removal machine learning model or a 3D noise removal machine learning model obtained through the simulation device (100) can simulate the quality of the image when a grid is present, even without an actual grid. Additionally, the user terminal (210) can display output radiation images, noise information, or artifact analysis graphs received from the simulation device (100) on a display screen so that the user can intuitively check the image quality.
[0046] The simulation system (200) has the effect of allowing the user to immediately check results for various shooting conditions by simply changing settings on the user terminal (210) without directly replacing the physical grid or operating the equipment. This provides an environment where medical device developers or researchers can access simulation resources and collaborate even from a remote location, and consequently contributes to shortening the product development cycle and maximizing research efficiency. In addition, quantitative analysis data (e.g., noise level comparison, compensation score, etc.) provided through the user terminal (210) can help the user objectively judge and determine the optimal grid specifications.
[0047] The operation of the simulation device (100) is described in more detail below. At least some of the operation of the simulation device (100) may be performed at the user terminal (210). The simulation system (200) may be implemented with the user terminal (210) and the simulation device (100) organically exchanging information. However, it is not limited to this, and the simulation device (100) may operate independently without exchanging information with other devices.
[0048] FIG. 3 is a drawing for explaining a method of operation of a simulation device according to one embodiment of the present disclosure. FIG. 4 is a drawing for explaining a method of operation of a simulation device according to one embodiment of the present disclosure. FIG. 5 is a drawing for explaining a method of operation of a simulation device according to one embodiment of the present disclosure.
[0049] The simulation device (100) may perform a step (301) of acquiring information of a virtual subject including at least one of the material, shape, and component information of the subject. The simulation device (100) may perform a step (301) of acquiring information of a subject that defines the physical properties of the subject to be located in a virtual shooting space. In this step, the simulation device (100) may load or generate data of the three-dimensional spatial distribution of the material to interact with X-rays. The information of the subject may include not only the external shape but also the density of the material, the atomic number, and detailed component information constituting the interior that determine the absorption and scattering of X-rays. The simulation device (100) may acquire this information in the form of three-dimensional voxel data, mesh data, or a phantom model defined by a mathematical function.
[0050] The subject information may be information about a virtual human body phantom that mimics the anatomical structure of the human body. For example, chest or head phantom data containing information on the X-ray attenuation coefficients of human tissues such as bone, muscle, fat, and air may be used. Additionally, the subject information may include information about a quality control phantom with a standardized shape (e.g., ACR 464 phantom) for quantitative performance evaluation. In this case, the component information is information about specific materials located within the subject, and may specifically define the location and material information of, for example, internal organs, blood vessels, artificially inserted metal implants, or contrast agents.
[0051] By acquiring such precise subject information, it is possible to safely reproduce various clinical conditions or extreme situations in a virtual environment that are difficult to simulate with actual patients. Since rare disease cases or environments inducing metal artifacts can be simulated without ethical constraints, high-quality data necessary to verify or improve the performance of diagnostic algorithms can be secured. Furthermore, by accurately defining the internal material information of the subject, it lays the foundation for physically calculating the location of scattered radiation or beam hardening phenomena during Monte Carlo simulations.
[0052] Information about the subject may not be acquired. As previously explained, the simulation device (100) may simulate a signal in which radiation from a radiation source passes through a grid and reaches a detector. That is, the simulation device (100) may simulate an 'Air Scan' or 'Flat-field' image. In this case, information about the subject may not be required. The simulation device (100) may require input grid specification information instead of information about the subject.
[0053] The simulation device (100) may perform a step (302) of acquiring radiation source information related to at least one of the location of a virtual radiation source or the intensity of radiation. The simulation device (100) may perform a step (302) of acquiring radiation source information that defines the physical characteristics of the radiation source, which is a virtual X-ray generator. In this step, the simulation device (100) may set the geometric location of the X-ray tube and the energy characteristics of the emitted radiation. The radiation source information is an important variable that determines the 'input energy' of the simulation and may include information on the focal position in three-dimensional space, the direction of irradiation, the cone angle of the beam, and the spectrum of the emitted X-rays.
[0054] The location of the radiation source included in the radiation source information may refer to the absolute or relative coordinates of the starting point from which X-rays are emitted within the three-dimensional space of the simulation. The simulation device (100) may set the location of the radiation source to define the geometric relationship between the radiation source, the subject, and the detector. For example, the location of the detector may be determined based on the location of the radiation source, and the location of the radiation source may be determined based on the location of the detector. However, this is not limited thereto, and the location of the detector and the location of the radiation source may be determined independently. The location information of the radiation source may include the Source-Axis Distance (SAD), which is the distance from the radiation source to the center axis of rotation, and the Source-Detector Distance (SDD), which is the distance from the radiation source to the surface of the detector. This distance information may serve as a key factor in determining the magnification ratio of the projected image.
[0055] Additionally, the position of the radiation source is not limited to a single fixed coordinate but may include a series of trajectory information that changes according to the flow of time or the shooting sequence. For example, to simulate cone-beam CT (CBCT) imaging, the position of the radiation source may be defined as a set of coordinates on a circular trajectory or a spiral trajectory that rotates by a predetermined angle (e.g., 360 degrees) around the subject. The simulation device (100) can obtain accurate radiation source angles and position coordinates at each projection point by taking into account the rotation speed of the gantry and the frame rate.
[0056] Furthermore, the location of the radiation source may include not only a location idealized as a point source, but also area or volume information considering the size and shape of the focal spot of the actual X-ray tube. The actual radiation source is not a perfect single point but has a focal area of a certain size (e.g., 0.6 mm, 1.2 mm, etc.), which may cause a penumbra phenomenon where the boundaries of the image are blurred. To reflect this physical phenomenon, the simulation device (100) may model the location of the radiation source as a set of multiple fine point sources or define it as a position distribution function according to the focal size.
[0057] By establishing precise radiation source location information, the geometry of the actual equipment can be perfectly reproduced, effectively enabling accurate simulation of image magnification and distortion characteristics. In particular, by setting the SAD and SDD to match the actual equipment, it is possible to ensure that the simulated image has the same field of view (FOV) and perspective as the actual captured footage; this is essential for maintaining spatial consistency during subsequent grid optimization or artifact correction.
[0058] Furthermore, by incorporating positional information that takes into account the focal size, it is possible to quantitatively predict the impact on image clarity and spatial resolution. Since it allows for the simulation of a more realistic blur effect than when assuming a point light source, it can contribute to verifying image quality degradation due to tube focal size in advance and selecting optimal tube specifications during the development of high-resolution imaging systems.
[0059] Radiation source information may include information related to radiation intensity, such as X-ray tube voltage (kVp), tube current (mA), and exposure time (ms). In particular, in this disclosure, to enhance the realism of the simulation image, the number of photons reaching the detector (Photon Count) can be adjusted as a parameter related to radiation intensity among the radiation source information. By setting the number of virtual light source particles to match the dose conditions of the actual equipment, the simulation device (100) can accurately simulate the rough texture of quantum noise that occurs when the dose is low or the smooth image quality characteristics when the dose is high.
[0060] By precisely configuring radiation source information, it is possible to virtually predict and optimize changes in image quality according to various imaging protocols (low vs. high dose). In particular, by reproducing noise levels identical to actual images through photon count control, it enables simulation data to seamlessly integrate with real clinical data. This resolves the issue of virtual data lacking realism during AI model training and contributes to creating models that operate robustly in actual diagnostic environments.
[0061] The simulation device (100) may perform the step (303) of acquiring detector information for a detector that receives radiation from a virtual radiation source. The simulation device (100) may perform the step (303) of acquiring detector information for a detector that receives radiation emitted from a virtual radiation source and passed through a subject and a grid. In this step, the simulation device (100) may load the specifications and physical characteristics of a light receiving unit that converts X-ray signals into digital images. The detector information is a parameter that determines the final result of the simulation and may include information on at least one of the position of the detector, the overall size of the detector, resolution, pixel pitch, and X-ray detection efficiency (DQE).
[0062] Detector information may include not only simple geometric values but also information on the material (CsI, GOS, etc.) or thickness of the scintillator that determines energy reactivity. In addition, the unique characteristics of the detector analyzed through the 'input radiation image' obtained in the previous step, such as the sensitivity deviation between pixels (Gain Map), the location of dead pixels, or the level of electronic noise (Dark Noise), may be reflected in the detector information. Through this, the simulation device (100) can perform a simulation that reflects the actual state of the equipment being used by the user, rather than an ideal detector.
[0063] By acquiring detector information and reflecting it in a simulation, it is possible to derive a grid solution optimized for specific hardware specifications. For example, the moiré phenomenon caused by frequency interference between the pixel pitch of a detector and the strip density of a grid can only be predicted if the detector information is accurate. By accurately reflecting the physical characteristics of such detectors, this disclosure can prevent image quality degradation issues that may occur during actual equipment installation and recommend grid specifications that bring out the maximum performance achievable by the detector.
[0064] The simulation device (100) may perform a step (310) of acquiring a first preliminary radiation image (410) based on Monte Carlo at a predetermined first resolution. To precisely simulate the physical scattering phenomenon of radiation, the simulation device (100) may perform a step (310) of performing a Monte Carlo simulation at a predetermined first resolution and acquiring the first preliminary radiation image (410) as a result. In step (310), the simulation device (100) may calculate the interaction paths, such as scattering, absorption, and transmission, that occur when hundreds of millions of photons emitted from a virtual light source collide with a subject and grid material. The first resolution may be set to a lower resolution than the resolution of the output radiation image to consider computational efficiency. For example, the first resolution may be 48x48 or 192x192 pixels. However, it is not limited thereto.
[0065] Although the first preliminary radiation image (410) has low resolution, it may be an image that most accurately reflects the complex scattered radiation distribution and noise characteristics occurring in the actual shooting environment. The simulation device (100) generates the first preliminary radiation image (410) by converting photon arrival information obtained through Monte Carlo simulation into pixel values, which is then used as a key element to impart realistic texture and scattering effects to the high-resolution image.
[0066] By applying the Monte Carlo method to the first resolution, which is a low-resolution region, it is possible to drastically reduce the overall computation time of the simulation while maintaining the accuracy of scattering simulation. This is based on the observation that scattering distribution has low-frequency characteristics and can be represented sufficiently accurately even at low resolutions, thereby enabling high-precision scattering calculations that previously took days to complete within hours or minutes, ensuring practicality in clinical settings.
[0067] The simulation device (100) may perform a step (320) of acquiring a second preliminary radiation image (420) based on Forward Projection at a predetermined second resolution that is higher in resolution than the first resolution. The simulation device (100) may perform a step (320) of acquiring a second preliminary radiation image (420) that is the result of performing a Forward Projection simulation at a second resolution that is higher in resolution than the first resolution in order to clearly represent the geometric structure of the subject. In step (320), the simulation device (100) may calculate the path of X-rays emitted from a virtual light source passing through the subject in a straight line and reaching the detector, and generate a projection image based on the integral value of the attenuation coefficient of the material inside the subject. The second resolution may be set to a high resolution that determines the quality of the output radiation image. For example, the second resolution may be 1536x1536 pixels. The number of pixels included in the second resolution may be 53 times or more and 1111 times or less the number of pixels included in the first resolution.
[0068] The second preliminary radiographic image (420) may be an image in which high-frequency components, such as the bones or organ boundaries of the subject, are very clearly visible, even though scattering phenomena are not reflected. The simulation device (100) can extract details and edge information of the image in a later stage based on this image.
[0069] By rapidly generating high-resolution images using a forward projection method, it is possible to ensure that the final output image possesses high spatial resolution. While obtaining high-resolution images using the Monte Carlo method alone requires a significant amount of time, combining it with the forward projection method efficiently secures geometric clarity, thereby enabling the provision of high-quality medical images capable of clearly identifying minute lesions or tissue structures.
[0070] The simulation device (100) may perform the step (330) of converting a second preliminary radiation image to a resolution lower than the second resolution to obtain a third preliminary radiation image (430). The resolution lower than the second resolution may include the first resolution. That is, the resolution lower than the second resolution may be the same as the first resolution or may not be the same.
[0071] In step (330), the simulation device (100) can sequentially lower the resolution of the second preliminary radiographic image using a Gaussian Pyramid method. For example, the simulation device (100) can apply a Gaussian blur to the second preliminary radiographic image (420) and downsample it to obtain a preliminary radiographic image (431) of intermediate resolution, and then downsample it again to obtain a preliminary radiographic image (432), and finally obtain a third preliminary radiographic image (430) having the same size as the first resolution, which is a Monte Carlo result. That is, the third preliminary radiographic image (430) can have the same size as the first preliminary radiographic image.
[0072] Step (330) is a step of extracting only the low-frequency components from the geometric information of the second preliminary radiographic image (420) generated by forward projection, and preparing to compare or synthesize it at the same dimension as the first preliminary radiographic image (410), which is the result of Monte Carlo simulation. The third preliminary radiographic image (430) maintains the structural features of the second preliminary radiographic image (420), but takes the form of having detailed features omitted.
[0073] Step (330) has the effect of establishing a foundation for fusing the first preliminary radiographic image (410) and the second preliminary radiographic image (420), which were generated in different ways, at the same resolution level by systematically converting the high-resolution image to a low-resolution image using a Gaussian pyramid. This allows scattering information and geometric information to be naturally mixed through the weighted average of the two images in a subsequent step, and at the same time provides a mathematical foundation for precisely separating only the high-frequency components (residual image) through the calculation of the difference from the high-resolution original image.
[0074] The simulation device (100) may perform a step (340) of obtaining a fourth preliminary radiographic image (440) by resizing an image converted to a low resolution back to a second resolution. Here, the image converted to a low resolution may include at least one of low-resolution images (431, 432) and a third preliminary radiographic image (430). In step (340), the simulation device (100) may take at least one of the third preliminary radiographic image (430) obtained in the preceding step (330) or the intermediate low-resolution images (431, 432) generated in the Gaussian pyramid process as input and perform an image processing process of upscaling it to the size of the original high-resolution second resolution (e.g., 1536x1536 pixels). Here, the fourth preliminary radiographic image (440) has the same pixel resolution as the second preliminary radiographic image (420), but since it is a low-resolution image that has been forcibly enlarged, the edges may be blurred and the overall image quality may be blurry.
[0075] Specifically, the simulation device (100) can perform resizing using various interpolation methods. For example, the simulation device (100) can expand the third preliminary radiographic image (430) to the second resolution by applying bilinear interpolation, which fills empty pixels by averaging the values of adjacent pixels, or bicubic interpolation, which uses a cubic function. This process can be understood as a process of removing high-frequency components and mapping only low-frequency components into a high-resolution space, and may correspond to the step of expanding an upper-level image to a lower level in the Laplacian Pyramid technique.
[0076] Step (340) has the effect of creating a state where pixel-to-pixel operations with the original second preliminary image are possible by resizing the low-resolution image to a high resolution to obtain the fourth preliminary image. This is an essential preprocessing step for calculating the difference from the original image in a subsequent step to precisely extract only high-frequency components. In addition, by generating an image that appears to be intentionally blurred, it provides a basis for excluding noise components within the image and separating only the overall structural information (low-frequency components) of the subject.
[0077] The simulation device (100) can perform the step (350) of acquiring a residual image (450) containing high-frequency components based on the fourth preliminary radiation image (440) and the second preliminary radiation image (420).
[0078] The simulation device (100) can acquire residual images (451, 452, 453) in stages using the Laplacian pyramid generation principle. For example, referring to FIG. 4, the simulation device (100) can extract a residual image (451) at the lowest resolution stage, a residual image (452) at an intermediate resolution stage, and a residual image (453) at the final second resolution stage, respectively.
[0079] More specifically, in step (350), the simulation device (100) can obtain a residual image (453) by performing an operation to subtract a fourth preliminary radiographic image (440), which is a blurry image composed only of low-frequency components, from a second preliminary radiographic image (420), which is a clear original image generated by forward projection. The residual image (453), which is the result of this subtraction operation, contains only high-frequency components with abrupt changes in brightness, such as the outline of a subject, the microstructure of a bone, and the boundary of an organ. Similarly, a residual image (451) can be obtained by subtracting a preliminary radiographic image (441), which is a third preliminary radiographic image (430) converted to high resolution, from a preliminary radiographic image (432) that is a second preliminary radiographic image (420) converted to low resolution. Additionally, a residual image (452) can be obtained by subtracting the preliminary image (442), which is a high-resolution conversion of the third preliminary image (430), from the preliminary image (431), which is a low-resolution conversion of the second preliminary image (420). At least one of the residual images (451, 452, and 453) obtained as described above may be referred to as a residual image (450). In the present disclosure, the residual image (450) may refer to the residual image (453), which is detail information at the highest resolution.
[0080] By separately acquiring a residual image (450), it is possible to obtain high-resolution geometric detail information that is difficult to obtain through Monte Carlo simulation. Although the Monte Carlo method provides accurate scattering information, it may be blurry due to low resolution; however, by using this residual image extracted from the forward projection method, 'clarity' can be maintained. In other words, by processing scattering information and shape information separately, an efficient data structure is established that combines the advantages of both types of information.
[0081] The simulation device (100) can perform the step (360) of obtaining a weighted average image (460) by weighting the third preliminary radiation image (430) and the first preliminary radiation image (410) based on predetermined weights. The simulation device (100) can perform the step (360) of obtaining a weighted average image (460) by weighting the third preliminary radiation image (430) and the first preliminary radiation image (410) based on predetermined weights. Step (360) is a process of fusing two low-resolution images generated by different simulation techniques into one. The first preliminary radiation image (410) is rich in scattering and physical noise information based on Monte Carlo, and the third preliminary radiation image (430) is a low-frequency image with geometric information reflected based on forward projection. The simulation device (100) can generate a weighted average image (460) that will become the base image by weighting and summing these two images pixel by pixel.
[0082] The simulation device (100) may assign equal weights (e.g., 0.5 to 0.5) to the two images, or set different weights depending on the purpose of the simulation. For example, if you want to emphasize the scattering effect more, you can assign a higher weight to the first preliminary radiographic image (410). The weighted average image (460) is generated as an image that simultaneously possesses the physical accuracy of Monte Carlo and the structural stability of forward projection, even though it is a first resolution which is low resolution.
[0083] By weighted averaging the Monte Carlo and forward projection results in low-resolution areas, it is possible to mitigate the particle noise characteristic of Monte Carlo simulations while preserving scattering information. Since the forward projection image is a clean, noise-free image, mixing it with the Monte Carlo image improves the overall signal-to-noise ratio (SNR) of the image while ensuring physical realism. This is a key balancing process that ensures the final image possesses image quality suitable for diagnosis without appearing overly artificial.
[0084] The simulation device (100) can perform a step (370) of acquiring an output radiation image (510) of a second resolution based on a weighted average image (460) and a residual image (450). The step (370) may be a process of recombining the 'low-frequency base information (scattering and transmission)' and 'high-frequency detail information (edges and boundaries)' that were separated and processed in the preceding steps to ultimately reconstruct the target high-quality image. The simulation device (100) can generate an output radiation image (510) by adding a residual image (450) corresponding to each resolution step while gradually expanding or upscaling the low-resolution weighted average image (460).
[0085] The simulation device (100) can perform the Laplacian Pyramid restoration algorithm in reverse order. Referring to FIG. 5, the simulation device (100) can generate an expanded image (522) by expanding the lowest resolution weighted average image (460) to a higher resolution, and obtain an intermediate restored image (512) by summing the corresponding resolution level residual image (451) in pixel units. Subsequently, the simulation device (100) can generate an expanded image (521) by expanding the restored image (512) again, and obtain a higher resolution restored image (511) by summing the next level residual image (452). Finally, the simulation device (100) can obtain a final output radiation image (510) by expanding the restored image (511) immediately before the final stage to a second resolution size to generate an expanded image (520), and by adding a residual image (453) containing the highest resolution detail information to it. The output radiation image (510) generated at this time may be in a state where the structural features of a clear subject (bones, blood vessels, etc.) derived from the forward projection method are maintained, while the physically accurate scattering distribution and texture derived from the Monte Carlo method are naturally fused. In addition, at this stage, the simulation device (100) can further increase the consistency with the actual equipment by additionally injecting quantum noise following a Poisson distribution into the entire image based on radiation source information (number of photons).
[0086] In this way, by synthesizing a high-resolution residual image with a low-resolution-based weighted average image, it is possible to simultaneously achieve two conflicting goals: computational speed and image quality. Although Monte Carlo simulation has high accuracy in simulating scattering, it requires a massive amount of time for high-resolution computation. The present disclosure limits this to the low-resolution region (weighted average image) and replaces the high-resolution information with a forward-projection-based residual image that is faster to compute, thereby drastically reducing the total simulation time from days to hours or minutes.
[0087] In addition, by not only increasing the resolution but also sequentially restoring detail information (residuals) for each resolution layer, it is possible to generate natural, clear, high-definition medical images without artificial artifacts. This enables efficient simulation even in environments with large data volumes, such as CBCT (Cone Beam CT) using large-area detectors, and consequently contributes to reducing the cost of repetitive testing in the medical device development process and shortening the time to market.
[0088] FIG. 6 is a drawing for explaining the effects of a simulation device according to one embodiment of the present disclosure.
[0089] FIG. 6 is a diagram comparing actual CBCT images with result images obtained from various simulation techniques (forward projection, Monte Carlo, hybrid) to verify the performance of a simulation device according to one embodiment of the present disclosure. An ACR 464 phantom, widely used for quantitative image quality evaluation, was used as the experimental subject. The top shows the Z-axis profile for the lateral projection image of the phantom, and the bottom shows the X-axis profile for the cross-sectional image.
[0090] First, the actual CBCT graph is ground truth data captured using actual medical imaging equipment, showing changes in pixel brightness due to density differences within the phantom. The peaks at both ends of the graph indicate cupping artifacts and brightness distortion caused by scattered radiation.
[0091] The Forward Projection graph is the result of a forward projection simulation; while the overall brightness trend is similar to the actual image, it exhibits a smooth shape with almost no noise. This is because probabilistic phenomena such as scattering effects or quantum noise are not reflected. On the other hand, although the Monte Carlo graph physically simulates scattering phenomena, it tends to show significant noise and reduced overall structural clarity due to low-resolution conditions or a limited number of photons.
[0092] Finally, the graph of Monte Carlo + Forward Projection is the result of applying the hybrid method of the present disclosure. This graph includes both the sharp structural features of forward projection and the scattering and noise characteristics of Monte Carlo. In particular, when compared with the profile of an actual CBCT image, it can be confirmed that the amplitude of the noise and the overall curve shape are very similar. This demonstrates that the simulation device (100) of the present disclosure reproduces the physical characteristics of the actual equipment with great precision.
[0093] FIG. 7 is a drawing for explaining the effects of another simulation device in one embodiment of the present disclosure.
[0094] FIG. 7 is a figure quantitatively comparing the X-axis brightness profiles between actual CBCT images and the hybrid simulation images of the present disclosure, with respect to a water phantom (25 cm). The water phantom has attenuation characteristics similar to human tissue and is composed of a uniform material, making it suitable for evaluating image degradation caused by beam hardening or scattered radiation.
[0095] Looking at the actual CBCT image and profile on the left, cupping artifacts are clearly observed where the edges of the image are bright and the center is dark, and irregular noise patterns appear throughout the profile. The Monte Carlo + Forward Projection result on the right is the image and profile generated by the simulation device (100) of the present disclosure. When comparing the two profiles, the difference in brightness between the center and the periphery (degree of cupping) is almost identical, and the texture of the fine noise also shows a very similar pattern.
[0096] These results suggest that the simulation device (100) does not simply calculate geometric projections, but physically models the interaction between X-rays and matter (scattering, absorption, beam hardening). Therefore, the user can trust these simulation results and perform grid specification optimization or dose reduction research based on them.
[0097] FIG. 8 is a drawing showing a method of operation of a simulation device according to one embodiment of the present disclosure.
[0098] The steps of FIG. 8 may be performed after step (370). Additionally, FIG. 3 may be performed again based on the radiation source information derived by the steps of FIG. 8 after the steps of FIG. 8 have been performed.
[0099] The simulation device (100) can perform the step (810) of acquiring an actual radiation image by taking an actual radiation image based on information about the subject and detector information.
[0100] The simulation device (100) may perform a step (810) of acquiring actual radiation images based on subject information and detector information in order to obtain a standard for matching a virtual simulation environment with an actual physical environment. In this step, the simulation device (100) may acquire original data by using actual medical imaging equipment (e.g., CBCT, X-ray, etc.) owned by the user to capture a subject (e.g., ACR 464 phantom, human body model, etc.) corresponding to the information of the subject to be simulated. At this time, the shooting conditions may be maintained identically to the geometric structure (SID, SOD, etc.) to be set in the simulation. Additionally, shooting may be performed using a detector corresponding to the detector information. The acquired image may be raw data to which noise correction or post-processing has not been applied, or projection data to which linearity is maintained.
[0101] To obtain actual radiation images, the simulation device (100) can capture multiple sets of images taken under various dose conditions. For example, an image with a coarse noise pattern under low dose conditions and a smooth image under high dose conditions can be acquired and used as basic data to identify noise behavior according to dose changes. The simulation device (100) can determine one image selected by the user from the multiple sets of images as the actual radiation image.
[0102] Actual radiographic images may consist of 'flat field' images made of uniform material or include complex phantom images containing a mixture of bone and muscle-equivalent materials. Therefore, actual radiographic images may comprehensively contain not only simple noise but also image quality degradation characteristics caused by scattered radiation.
[0103] The simulation device (100) can perform the step (820) of acquiring actual noise level information of an actual radiation image. The simulation device (100) can perform the step (820) of acquiring actual noise level information by analyzing the acquired actual radiation image and quantifying the image quality characteristics unique to the equipment. In the step (820), the simulation device (100) can identify the intensity and type of noise by setting a region of interest (ROI) within the image and performing statistical analysis. The actual noise level information may include at least one of general random noise indicators such as standard deviation or variance representing the roughness of the image, and intensity information of specific artifacts caused by structural defects or physical phenomena of the system.
[0104] The types of noise used to calculate actual noise level information may include cupping artifacts, where the center of the image appears dark due to beam hardening, and moiré, which is a periodic wave pattern caused by interference between the grid and detector pixels. The simulation device (100) can quantify the severity of cupping artifacts by comparing the brightness profiles of the center and periphery of the image, or extract the intensity of the frequency components of the moiré pattern through frequency analysis (Fourier Transform) and store this as actual noise level information. This information goes beyond simple 'noise' and serves as a fingerprint representing the overall image quality status of the corresponding shooting system.
[0105] When acquiring actual noise level information, the simulation device (100) may apply a method of normalizing each noise value and calculating a weighted average to objectively integrate multiple noise components with different units and ranges existing within the image. Random noise, expressed as the standard deviation of pixel brightness in the image, Moire patterns expressed as frequency intensity, and Cupping artifacts expressed as brightness gradients may not be possible to simply sum because they differ in physical dimensions and numerical scales. Therefore, the simulation device (100) may determine the final actual noise level information by dividing each noise component by a predetermined maximum value or reference value to normalize it to a value between 0 and 1, and multiplying it by a weighting factor that reflects clinical importance. This can be expressed as Equation 1 as follows.
[0107] Real_Noise_Level = w_r * (N_random / Max_random) + w_m * (N_moire / Max_moire) + w_c * (N_cupping / Max_cupping) (Equation 1)
[0109] In Equation 1 above, N_random, N_moire, and N_cupping are the raw values of the measured random noise, moire intensity, and cupping artifact index, respectively, and Max_random, Max_moire, and Max_cupping represent the theoretical or empirical maximum values that each noise component can have. Max_random, Max_moire, and Max_cupping may be predetermined values. The operation within the parentheses represents a normalization process that converts each noise value to the same scale. Additionally, w_r, w_m, and w_c are weighting coefficients assigned to each noise component, and their sum can be set to 1. Also, w_r, w_m, and w_c may be predetermined values.
[0110] By distinguishing various types of noise and acquiring actual noise level information, it is possible to define the target that the simulation device must model in a specific and multifaceted manner. By analyzing not only the quantity of noise but also its quality and form, the virtual images to be generated subsequently can precisely reproduce the textures and artifact characteristics of the actual images. This enables the accurate prediction of potential problems that may occur in actual clinical environments (e.g., intensified cupping in specific patient body types) when evaluating grid performance through simulation.
[0111] The simulation device (100) can perform the step (830) of acquiring virtual noise level information of the output radiation image. The simulation device (100) can analyze the output radiation image generated based on the output grid specification information, the subject information, the radiation source information, and the detector information, and perform the step (830) of acquiring virtual noise level information thereon. This step is a process of performing image quality analysis in the same manner on the virtual image in order to compare it with the actual image analyzed earlier (step 820). The simulation device (100) can calculate virtual noise level information by setting an ROI at the same location as the actual image in the virtual image generated through Monte Carlo and forward projection hybrid simulation, and by performing standard deviation, profile analysis, frequency analysis, etc. The process of acquiring virtual noise level information may be the same as the process of acquiring actual noise level information. Therefore, redundant explanations are omitted.
[0112] The simulation device (100) can perform a step (840) of modifying radiation source information so that virtual noise level information becomes identical to actual noise level information. Step (840) is a key correction process that narrows the gap between the virtual world and the real world by fine-tuning simulation parameters. The simulation device (100) can adjust radiation source information, particularly the number of photons reaching the detector, in a direction that minimizes the difference value (Error) between the two noise level information.
[0113] Specifically, noise (quantum noise) in X-ray images has a Poisson distribution characteristic that is inversely proportional to the square root of the number of arriving photons. Therefore, the simulation device (100) can modify the radiation source information by reducing the number of photons in the simulation to increase the noise if the noise in the virtual image is less than in reality, and conversely, by increasing the number of photons to smooth the image if the noise is too much. In addition, the average energy of the beam can be adjusted by modifying the energy spectrum distribution of the radiation source to reduce the difference in cupping artifacts.
[0114] By modifying radiation source information, particularly the number of photons, to match the noise level, the output radiation image becomes visually and statistically very similar to the actual radiation image. This imparts a hyper-realistic texture to the simulation result, enabling the achievement of a high quality where medical professionals might mistake the virtual image for an actual clinical image. Such high-quality simulation images have high value as data for AI training and can be utilized as a highly reliable verification tool that can completely replace actual imaging during the equipment development phase. Furthermore, the radiation source information, once corrected, is continuously utilized in the subsequent steps of Fig. 8, enabling simulations with consistently high accuracy even under various grid change conditions.
[0115] Step (840) may further include the following process. The simulation device (100) may not only adjust the number of photons as in Step (840) but also modify the weights in Step (360). More specifically, in Step (360) of FIG. 3, the radiation source information was modified to match the actual noise level information with the virtual noise level information. However, it is not limited to this. The simulation device (100) may match the actual noise level information with the virtual noise level information by modifying predetermined weights. For example, the higher the actual noise level information is compared to the virtual noise level, the simulation device (100) may be controlled to assign a higher weight to the first preliminary radiation image (410). Conversely, the lower the actual noise level information is compared to the virtual noise level, the simulation device (100) may be controlled to assign a lower weight to the first preliminary radiation image (410). The actual noise level information and the virtual noise level can be matched by adjusting the weights.
[0116] After the steps of Fig. 8 are modified, the steps of Fig. 3 can be performed using the modified radiation source information or weights. The output radiation image obtained by the steps of Fig. 3 performed after Fig. 8 can be very similar to the actual radiation image.
[0117] Through the steps of FIG. 3, the simulation device (100) can obtain training data for machine learning. For example, the second preliminary radiation image can be input data used to generate a machine learning model, and the output radiation image can be ground truth data used to generate a machine learning model. The simulation device (100) can perform the step of receiving the second preliminary radiation image as input and generating a machine-learned scattering simulation machine learning model to output an output radiation image.
[0118] More specifically, the simulation device (100) may receive a second preliminary radiation image based on forward projection and perform the step of generating a scattering simulation machine learning model to output an output radiation image that reflects scattering information based on Monte Carlo. In this step, the simulation device (100) may obtain a large number of image pairs generated through the preceding hybrid simulation process (steps of FIG. 3) as a training dataset. Specifically, a 'second preliminary radiation image' that is clear but has no scattering effect and contains only geometric information may be set as input data, and an 'output radiation image' that physically reflects scattering and noise effects through Monte Carlo simulation and synthesis processes may be set as ground truth data.
[0119] The creation and learning process of a scattering simulation machine learning model may include forward propagation and backpropagation processes. First, in the forward propagation process, the simulation device (100) inputs a second preliminary radiographic image into an initial model to generate a prediction image, and can calculate the difference between this prediction image and the output radiographic image, which is the ground truth data, using a loss function. As a loss function, mean squared error (MSE) or structural similarity (SSIM) that calculates the difference between pixels may be used. Subsequently, in the backpropagation process, the simulation device (100) can update the weights and biases within the neural network in a direction that minimizes the calculated loss value. At this time, Gradient Descent or the Adam optimization algorithm may be used so that the model can learn on its own how to synthesize a scattering distribution and noise texture appropriate to the geometric structure of the second preliminary radiographic image.
[0120] Various deep learning algorithms can be used to implement a scattering simulation machine learning model. Representative examples include the U-Net structure, which is specialized for image-to-image translation. Through an encoder-decoder structure and skip connections, U-Net can effectively learn texture information, such as low-frequency components like scattering or noise, while maintaining high-resolution structural information of the input image (second pre-radiography image). Additionally, by applying a Generative Adversarial Network (GAN), a more realistic scattering effect can be simulated through competitive learning, where a generator creates an image that is difficult to distinguish from the actual simulation result and a discriminator evaluates it.
[0121] Subsequently, the simulation device (100) can acquire an output radiation image based on a second preliminary radiation image without using a Monte Carlo method that consumes a lot of processing resources. That is, by using the steps below instead of the steps of FIG. 3, the simulation device (100) can acquire an output radiation image of the same quality as FIG. 3 quickly while using fewer processing resources. The simulation device (100) can perform a step (301) of acquiring information of a virtual subject including at least one of the material, shape, and component information of the subject. The simulation device (100) can perform a step (302) of acquiring radiation source information related to at least one of the location of the virtual radiation source or the intensity of the radiation. The simulation device (100) can perform a step (303) of acquiring detector information for a detector that receives radiation from the virtual radiation source. The simulation device (100) can perform the step (320) of acquiring a second preliminary radiation image based on Forward Projection at a predetermined second resolution, based on information of the subject, radiation source information, and detector information. The simulation device (100) can perform the step of acquiring an output radiation image by applying the second preliminary radiation image to a scattering simulation machine learning model.
[0122] By using a scattering simulation machine learning model trained in this manner, it is possible to immediately acquire high-precision scattering images using only high-speed forward projection simulation (generation of a second preliminary radiation image) without performing computationally expensive Monte Carlo simulations. While the Monte Carlo method is accurate, it has the disadvantage of being time-consuming; however, the present disclosure utilizes it only for generating training data and uses the trained AI model in the actual inference stage, thereby drastically reducing the overall simulation speed to a real-time level.
[0123] Furthermore, rather than simply adding noise, the AI is trained using physical simulation results (output radiographic images) as ground truths, which enables it to accurately simulate nonlinear scattering behaviors based on the thickness or density of the subject. This maximizes the efficiency of medical image analysis research and equipment development by providing images with artifact and noise characteristics that are highly realistic, even in cases involving complex human structures or metal implants.
[0124] The simulation device (100) can acquire an output radiation image through step (370). Additionally, the simulation device (100) can acquire an output radiation image through a scattering simulation machine learning model. Based on the output radiation image acquired in this way, the simulation device (100) can acquire a three-dimensional radiation image using a sparse view.
[0125] To acquire CT (Computed Tomography) or CBCT (Cone Beam CT) images, multiple projection images are required by rotating the radiation source around the subject. Generally, to reconstruct a high-quality 3D image, hundreds to thousands of projection images are required across the entire 360-degree range at very fine angular intervals (e.g., 0.5 or 1 degree). However, as the number of scans increases, the radiation exposure received by the patient increases proportionally; therefore, reducing the number of scans while maintaining the minimum image quality necessary for diagnosis is a critical challenge in the field of medical imaging. Theoretically, 3D reconstruction is possible if a certain level of projection data is secured within an angular range of 180 degrees or more; however, if data is insufficient (undersampling), severe streak artifacts may occur in the image.
[0126] In conventional CT or CBCT image acquisition methods, projected images had to be mechanically and precisely acquired at predetermined angles. These predetermined angles can be set in various ways depending on the equipment specifications or imaging protocols, for example, at 1-degree intervals, 2-degree intervals, or wider 5-degree intervals. However, as the interval widens—that is, as the condition becomes sparse-view—a problem arises in which the quality of the reconstructed image deteriorates rapidly.
[0127] The present disclosure provides a technology to restore projection images of insufficient angles using artificial intelligence by utilizing an image interpolation machine learning model to solve these problems. A simulation device (100) can input a projection image captured at a first angle (e.g., 1 degree) and a projection image captured at a second angle (e.g., 3 degrees) into an image interpolation machine learning model to virtually generate and restore a projection image corresponding to a third angle (e.g., 2 degrees), which is an angle between them. Here, the first angle and the second angle may be adjacent shooting angles. Through this, actual shooting is performed sparsely to drastically reduce the amount of radiation exposure to the patient, while at the data processing stage, a sufficient amount of projection data can be secured as if it were shot densely.
[0128] The simulation device (100) can reconstruct a 3D CT image or CBCT image by combining all projection images restored through the image interpolation machine learning model and actual (or simulated) projection images. Since a sufficient number of projection data is secured, clear, high-quality 3D volume data without artifacts can be obtained even when applying the Feltkamp (FDK) algorithm or iterative reconstruction technique. This serves as a key technical means to overcome the limitations of low-dose imaging through software.
[0129] As training data for creating an image interpolation machine learning model, output radiation images obtained in the previously described step (370) or high-quality output radiation images obtained through a scattering simulation machine learning model may be used. Although it is difficult to obtain dense data at 1-degree intervals from actual patients as it may cause harm to the patients, the simulation device (100) of the present disclosure can generate high-quality projection images indefinitely at desired angles (e.g., in 0.1-degree units) in a virtual environment. Therefore, the simulation device (100) can train the model by generating a large number of pairs of '1-degree and 3-degree images (input)' and '2-degree images (correct answers)' through simulation, and the generated images contain precise scattering and noise information based on Monte Carlo, thereby ensuring high accuracy even when applied in actual clinical settings.
[0130] The training process of an image interpolation machine learning model is carried out through forward propagation and backpropagation. During forward propagation, the model infers and generates an image of an intermediate angle from input images of adjacent angles, and calculates the pixel-level loss by comparing it with the ground truth. In the backpropagation process, the weights within the neural network are updated using optimization algorithms, such as Gradient Descent, to minimize this loss. By repeating this process, the model identifies geometric changes and texture flow between two adjacent images and acquires the ability to precisely synthesize images of missing intermediate viewpoints.
[0131] Deep learning algorithms suitable for learning the temporal or spatial continuity of images can be applied to image interpolation machine learning models. For example, image generation models based on U-Net structures or Generative Adversarial Networks (GANs) may be used, and optical flow-based neural networks or deformable convolutional networks used in video frame interpolation techniques may also be utilized. These algorithms do not simply average pixel values but predict structural changes caused by the movement or rotation of the subject to fill in empty spaces, thereby enabling the restoration of clear intermediate images without blur.
[0132] By utilizing image interpolation machine learning models, the number of projection images required for CT or CBCT scans can be drastically reduced, effectively minimizing patient radiation exposure. Furthermore, by using sophisticated training data generated through simulation, the problem of insufficient actual clinical data can be resolved, and high-performance interpolation models can be constructed. As a result, high-resolution 3D images comparable to those produced by high-dose imaging can be reconstructed even under low-dose imaging conditions, thereby ensuring both diagnostic accuracy and safety.
[0133] FIG. 9 is a flowchart illustrating the operation method of a simulation device according to one embodiment of the present disclosure.
[0134] A virtual radiation source may be rotatable around a predetermined axis of rotation. For example, the virtual radiation source may be included in a CT or CBCT device. The virtual radiation source may be used to acquire CT or CBCT images.
[0135] The simulation device (100) can perform the step (910) of acquiring multiple output radiation images according to the position of a virtual radiation source. The simulation device (100) can perform the step (910) of acquiring multiple output radiation images according to the position change of the virtual radiation source as basic data for reconstructing a three-dimensional volume image. In this step, the simulation device (100) can simulate the rotational movement of actual CT (Computed Tomography) or CBCT (Cone Beam CT) equipment in a virtual environment. Specifically, the virtual radiation source can rotate and move at constant angular intervals (e.g., 1 degree, 0.5 degrees) around a predetermined rotation axis (Isocenter), and the simulation device (100) can sequentially generate output radiation images, which are two-dimensional projection images, by calculating the interaction between the radiation source, the subject, the grid, and the detector for each rotation angle (View Angle).
[0136] The simulation device (100) can repeatedly perform the Monte Carlo and forward projection hybrid simulation techniques described above to acquire output radiation images at each angle. In order to increase computational efficiency, data that does not change, such as material property information of the subject or grid specifications, can be cached in memory (120) and reused even if the position of the radiation source changes. The number of multiple output radiation images acquired can be set from hundreds to thousands depending on the quality target of the 3D image to be reconstructed, and in the case of sparse-view CT simulation, only a smaller number of images may be selectively acquired.
[0137] By virtually acquiring projected images from various angles, this effectively replaces the cumbersome and dangerous process of rotating a gantry to image actual patients with computer simulation. This not only fundamentally prevents patient radiation exposure but also provides the freedom to generate images with desired temporal and angular resolutions, regardless of the equipment's mechanical rotation speed or frame rate. Furthermore, since independent simulations are possible for each angle, the overall data acquisition time can be drastically reduced through parallel computing.
[0138] The simulation device (100) can perform the step (920) of acquiring a three-dimensional output radiation image based on a plurality of output radiation images. The simulation device (100) can perform the step (920) of acquiring a three-dimensional output radiation image that visualizes the internal structure of a subject in three dimensions based on the acquired plurality of output radiation images. This step is an image reconstruction process that converts two-dimensional projection data into three-dimensional volume data by back-projecting. The simulation device (100) can synthesize a plurality of output radiation images into a single integrated three-dimensional output radiation image by applying various CT reconstruction techniques, such as the Feldkamp-Davis-Kress (FDK) algorithm or the Iterative Reconstruction algorithm.
[0139] The 3D output radiation image is stored in voxel units, and each voxel value can represent the X-ray attenuation coefficient (Hounsfield Unit, HU) of the material at the corresponding location. Based on this 3D data, the simulation device (100) can extract various cross-sectional images such as coronal, sagittal, and axial planes, or display the three-dimensional shape of the subject on the display through volume rendering. In particular, by using a high-quality output radiation image with scattering removal and noise correction applied in the preceding step for reconstruction, a clean image quality with minimized cupping artifacts or streak artifacts can be obtained even in the 3D image.
[0140] By extending the results of 2D simulations into 3D, it is possible to provide information on the three-dimensional structure and depth of the subject, which is difficult to perceive in simple planar images. This enables the clinical value provided by CT or CBCT in actual medical settings (e.g., accurate identification of tumor locations, verification of complex bone fracture patterns, etc.) to be realized in a virtual environment. Furthermore, through the 3D reconstruction process, minute noise remaining in the 2D image can be averaged out and removed, or conversely, artifacts visible only from specific angles can be analyzed as manifesting in 3D space, thereby allowing for more precise grid performance evaluation and image quality optimization.
[0141] The simulation device (100) may perform the step of receiving a 3D preliminary radiation image and generating a machine-learned 3D scattering simulation machine learning model to output a 3D output radiation image. The simulation device (100) may perform the step of receiving a 3D preliminary radiation image and generating a machine-learned 3D scattering simulation machine learning model to output a 3D output radiation image that reflects scattering and physical phenomena. In this step, the simulation device (100) may use 3D volume data, rather than 2D planar images, as the unit of learning to model complex interactions between X-rays and matter occurring in three-dimensional space. The simulation device (100) may reconstruct a plurality of second preliminary radiation images acquired based on a forward projection method to generate a 3D preliminary radiation image that does not contain scattering information, and utilize this as input data for the model.
[0142] A 3D scattering simulation machine learning model can perform the function of analyzing the geometric structure and density information of an input 3D preliminary radiographic image, inferring the 3D scattering distribution predicted to occur in the structure, and synthesizing it onto the original. The 3D output radiographic image that serves as the ground truth of the training data may be a high-precision 3D image generated based on Monte Carlo simulation, or real data captured by actual CT equipment. The simulation device (100) can learn the non-linear correlation between the input geometric volume data and the physical volume data that serves as the ground truth, and update weight parameters that reproduce 3D artifacts caused by beam hardening or Compton scattering.
[0143] Deep learning algorithms capable of processing 3D spatial information can be applied to the creation of a 3D scatter simulation machine learning model. For example, V-Net, a U-Net structure based on a 3D Convolutional Neural Network (3D CNN), can be used. This model receives 3D volume data as input, undergoes encoding and decoding processes, and can predict scattering information by analyzing spatial associations between voxels. Additionally, by utilizing a 3D Generative Adversarial Network (3D GAN), more sophisticated and realistic 3D textures can be realized by training to compete with a discriminator so that the virtual 3D scattering image created by the generator is indistinguishable from the actual Monte Carlo result.
[0144] The simulation device (100) can acquire a 3D output radiation image without using the Monte Carlo method by utilizing a 3D scattering simulation machine learning model. Therefore, the simulation device (100) can use fewer processing resources and can output a 3D output radiation image very quickly. More specifically, the simulation device (100) can perform a step (301) of acquiring information of a virtual subject including at least one of the material, shape, and component information of the subject. The simulation device (100) can perform a step (302) of acquiring radiation source information related to at least one of the position of a virtual radiation source or the intensity of radiation that can be rotated around a predetermined rotation axis. The simulation device (100) can perform a step (303) of acquiring detector information for a detector that receives radiation from the virtual radiation source. The simulation device (100) can perform the step (320) of acquiring a second preliminary radiation image based on Forward Projection at a predetermined second resolution, based on information of the subject, radiation source information, and detector information. The simulation device (100) can perform the step (1010) of acquiring a plurality of second preliminary radiation images according to the position of a rotating virtual radiation source. The simulation device can perform the step of acquiring an output radiation image by applying the plurality of second preliminary radiation images to a 3D scattering simulation machine learning model.
[0145] By creating a 3D scattering simulation machine learning model that directly learns 3D data in this way, it is possible to precisely simulate three-dimensional scattering effects and artifacts that are difficult to realize with 2D image processing. X-ray scattering is not limited to a specific cross-section but affects adjacent cross-sections or the entire volume; since the present disclosure learns information about the entire 3D space in an integrated manner, it can accurately reproduce even cone-beam artifacts and scattering interference phenomena between slices.
[0146] In addition, once the model is created, it has the effect of immediately obtaining high-quality 3D output radiographic images using only high-speed forward projection results (3D preliminary radiographic images) without having to perform time-consuming Monte Carlo simulations every time. This drastically reduces the time required to build large-scale 3D datasets during the development of medical imaging systems and provides near-real-time simulation performance, enabling efficient virtual testing for various clinical conditions.
[0147] FIG. 10 is a flowchart illustrating the operation method of a simulation device according to one embodiment of the present disclosure.
[0148] The simulation device (100) can perform the step (1010) of acquiring a plurality of second preliminary radiation images according to the position of a virtual radiation source. The simulation device (100) can perform the step (1010) of acquiring a plurality of second preliminary radiation images according to the position change of a virtual radiation source as basic data for generating three-dimensional volume data. In this step, the simulation device (100) can sequentially generate high-resolution projection images captured by rotating the virtual radiation source around the subject using a forward projection method. The second preliminary radiation image may be an ideal projection image that is geometrically clear and noise-free, and does not contain scattering information from Monte Carlo simulation.
[0149] The simulation device (100) can acquire hundreds to thousands of second preliminary radiation images by moving the position of the radiation source at preset angle intervals (e.g., 0.5 degrees, 1 degree). At this time, each image is a two-dimensional image in which the internal structure of the subject, such as bones, organs, and blood vessels, is superimposed, and is used to restore three-dimensional information through subsequent three-dimensional reconstruction. This process follows the same geometric principles as actual CT scanning, but differs in that it extracts only pure signals from which physical scattering or electronic noise is excluded.
[0150] By obtaining clean second preliminary radiographic images free of scattering or noise from various angles, it has the effect of generating high-quality reference data that can be used as 'Label' or 'Target' data when training artificial intelligence models. This plays a key role in helping subsequent noise removal models clearly distinguish between noise components to be removed and signal components to be preserved in actual images.
[0151] The simulation device (100) can perform the step (1020) of acquiring a three-dimensional preliminary radiation image based on a plurality of second preliminary radiation images. The simulation device (100) can perform the step (1020) of acquiring a three-dimensional preliminary radiation image based on the acquired plurality of second preliminary radiation images. This step is a process of reconstructing two-dimensional projection data into volume data in three-dimensional space through a back-projection algorithm, etc. The three-dimensional preliminary radiation image is voxel data that expresses the internal structure of the subject in three dimensions, and each voxel value can represent an ideal X-ray attenuation coefficient that is free from the influence of scattering or artifacts.
[0152] The simulation device (100) can generate a three-dimensional preliminary radiographic image by applying the Feltkamp algorithm or an iterative reconstruction technique. Unlike actual CT images, this image contains perfect anatomical structural information with no beam hardening, cupping caused by scattered radiation, or streak artifacts. Therefore, this image serves as a reference model that defines the 'best image quality state' that a noise removal model should reach.
[0153] By acquiring 3D preliminary radiographic images from which noise and artifacts have been completely removed, it is possible to maximize the learning efficiency and performance of 3D noise removal models. The clearer and cleaner the ground truth data of the training data, the faster and more accurately the AI model can learn noise removal patterns; consequently, when applied to actual clinical images, it can provide a high-performance filtering effect that selectively removes only noise without damaging minute tissues.
[0154] The simulation device (100) can perform the step (1030) of generating a machine-learned 3D noise removal machine learning model to receive a 3D output radiation image and output a 3D preliminary radiation image. Since the method for acquiring the 3D output radiation image has already been explained in step (920), a redundant explanation is omitted.
[0155] The simulation device (100) can perform the step (1030) of generating a 3D noise removal machine learning model using the acquired 3D output radiation image of realistic quality as input and the 3D preliminary radiation image of ideal quality generated only by forward projection as the correct answer (Label). This step is a process that enables the artificial intelligence model to have the ability to distinguish between noise and signals through a supervised learning method. The simulation device (100) can optimize model parameters to produce an output similar to the correct answer image by selectively removing only the blurry noise components caused by shadows due to beam hardening or scattering while maintaining structural information of bones or organs in the input 3D volume data.
[0156] The training of a 3D noise removal machine learning model can be performed through an iterative process of forward propagation and backpropagation. First, in the forward propagation process, the simulation device (100) can inject a 3D output radiation image containing noise into the input layer of the model. The input data passes through a plurality of 3D convolutional layers, activation functions, and pooling layers, where features are extracted and transformed to output a predicted 3D image with noise removed. At this time, the simulation device (100) can calculate a loss value by calculating the difference between the predicted image output by the model and the ground truth data, the 3D preliminary radiation image, using the Mean Squared Error (MSE) or L1 loss function.
[0157] Subsequently, during the backpropagation process, the simulation device (100) can update the weights and biases within the neural network in a direction that minimizes the calculated loss value. The simulation device (100) can propagate the gradient of the loss from the output layer to the input layer using the chain rule and update the parameters using an optimization algorithm (Optimizer) such as Gradient Descent, Adam, or RMSprop. These forward and backpropagation processes are repeated until the loss value converges below a preset threshold or reaches a specified number of learning epochs, through which the model learns generalized noise removal performance for various anatomical structures and noise patterns.
[0158] As an available algorithm, the 3D U-Net (or V-Net) architecture, specialized for 3D medical image processing, can be applied. 3D U-Net has a structure that captures contextual information of the 3D volume in the contracting path and restores precise localization information in the expanding path. By directly transmitting high-resolution structural information of the input image to the output through skip connections, it can effectively remove only noise without loss of anatomical detail. Additionally, by utilizing a 3D conditional generative adversarial network (3D cGAN) to perform competitive learning where a generator creates a noise-removed image and a discriminator evaluates the quality of that image, it is possible to restore 3D images with sharper and more realistic textures without blur.
[0159] By creating and utilizing a 3D noise removal machine learning model, it is possible to dramatically improve the image quality of CT images acquired in low-dose imaging or environments with severe scattering. In particular, by generating a 'noise-free, perfect ground truth image (3D preliminary radiographic image)' through simulation—which is impossible to obtain with actual imaging data alone—and utilizing it for training, the problem of the absence of ground truth data, which was a limitation of conventional supervised learning, can be fundamentally resolved.
[0160] In addition, by performing learning in 3D volume units rather than 2D cross-sectional units, it is possible to achieve consistent noise removal that considers the continuity between slices. This enables the effective suppression of three-dimensional artifacts (e.g., cone-beam artifacts) that occur during the 3D reconstruction process, and consequently contributes to providing more accurate 3D diagnostic information to medical professionals.
[0161] The simulation device (100) may perform a step of receiving an output radiation image and generating a machine-learned noise-removing machine learning model to output a second preliminary radiation image. In this step, the simulation device (100) may perform supervised learning by utilizing the two types of images generated through the preceding hybrid simulation process as training data pairs. Specifically, the output radiation image is a realistic image that reflects scattering information from Monte Carlo simulation and quantum noise based on the actual number of photons, and serves as input data for the noise-removing machine learning model. On the other hand, the second preliminary radiation image is an ideal image that is generated solely by the forward projection method and has no scattering or noise at all, and serves as correct answer data for the noise-removing machine learning model.
[0162] A noise removal machine learning model can learn to analyze an output radiation image, which is a two-dimensional planar image, and to identify and remove only scattered radiation, beam hardening artifacts, and particle noise components that degrade image quality while maintaining the anatomical structure information of the subject. The learning process can be performed using a Convolutional Neural Network (CNN)-based U-Net structure or an Autoencoder structure. A simulation device (100) can generate a prediction image by passing the input output radiation image through a noise removal machine learning model, and update weights by performing backpropagation in a direction that minimizes the pixel-unit error (Loss) between this prediction image and the correct answer, the second preliminary radiation image.
[0163] In particular, this noise removal machine learning model can be distinguished from a 3D noise removal machine learning model that processes 3D volume data after reconstruction is complete, in that it processes 2D projection data prior to reconstruction. Since the noise removal machine learning model learns 2D patterns at each projection angle, it can be specialized in removing high-frequency noise such as fog phenomena that spread widely across the entire image due to scattered rays, detector pixel defects, and quantum noise. The simulation device (100) can use tens of thousands or more data pairs generated under various dose conditions (changes in the number of photons) or various subject thickness conditions for learning, thereby ensuring that the noise removal machine learning model does not overfit to specific conditions and possesses general-purpose noise removal performance.
[0164] By generating a noise reduction machine learning model using simulation data, it is possible to utilize a second preliminary radiographic image, which is a perfect ground truth image that cannot be obtained from actual clinical data, for training. Since scattering and noise cannot be physically completely eliminated from images of actual patients even when high doses are used, a model trained using these as ground truth images is bound to have limitations in noise reduction performance. However, the present disclosure provides a mathematically perfect, defect-free image as ground truth through forward projection, thereby enabling the noise reduction machine learning model to learn how to perfectly separate noise and signals.
[0165] In addition, by removing noise through a noise removal machine learning model in the pre-reconstruction stage, it is possible to fundamentally prevent noise from being amplified or deformed into structural artifacts during the 3D reconstruction process. Fine noise in a 2D projected image can cause streak artifacts that spread throughout the entire 3D space during the back-projection process, but the present disclosure can significantly improve the quality of the final 3D image by blocking this in advance. Since the output radiation image generated by the simulation device (100) closely mimics the noise characteristics of the actual equipment, the noise removal machine learning model learned through this can exhibit excellent performance without any sense of incongruity when applied to actual clinical images.
[0166] FIG. 11 is a drawing showing a method of operation of a simulation device according to one embodiment of the present disclosure.
[0167] The simulation device (100) can perform the step (1110) of acquiring an actual radiation image obtained by irradiating an actual subject with radiation. In the step (1110), the simulation device (100) can acquire image data taken using X-ray or CT equipment on an actual patient or subject at a medical site. The actual radiation image may contain a lot of noise due to low-dose imaging, or may be an image in which artifacts have occurred due to the patient's movement or metallic foreign objects.
[0168] The simulation device (100) can receive images directly from a hospital's PACS (Picture Archiving and Communication System) system or medical imaging device. The images acquired at this time may be 2D projected images or 3D cross-sectional images that have already been reconstructed. The simulation device (100) can convert the format of the images for the subsequent application of a noise removal model, or perform a task to normalize the brightness or contrast of the images through a preprocessing process.
[0169] The simulation device (100) can perform a step (1120) of obtaining an improved image with reduced noise by applying a noise removal machine learning model to an actual radiation image. The simulation device (100) can perform a step (1120) of obtaining an improved image with reduced noise and improved image quality by applying a previously trained noise removal machine learning model to the acquired actual radiation image. Since the process of generating the noise removal machine learning model used in step (1120) has already been explained, a redundant explanation is omitted. The simulation device (100) can input the actual radiation image into this model and perform an inference process to identify and remove scattered radiation, beam hardening artifacts, quantum noise, etc. present in the image.
[0170] The model can selectively filter out only the factors that degrade image quality based on learned patterns while preserving the anatomical structures (bones, organs, etc.) of the input images. For example, it can brighten beam-hardening shadows occurring around bones or smooth out the rough graininess characteristic of low-dose images. The final enhanced images possess clarity and contrast comparable to high-dose images, providing medical professionals with a state in which lesions can be identified more clearly.
[0171] By applying a high-performance noise removal machine learning model trained on simulation data to actual images, it is possible to obtain high-quality diagnostic images without increasing the patient's radiation exposure. Previously, it was necessary to increase the radiation dose to improve image quality, but the present disclosure effectively removes noise through software-based post-processing alone, thereby ensuring patient safety while improving diagnostic accuracy. Furthermore, since existing low-quality images can be restored without re-imaging, it contributes to reducing medical costs and improving clinical efficiency.
[0172] We have examined various embodiments so far. Those skilled in the art will understand that the present invention may be embodied in modified forms without departing from the essential characteristics of the invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the invention is defined by the claims, not by the foregoing description, and all variations within the scope of equivalents should be interpreted as being included in the invention.
[0173] Meanwhile, the embodiments of the present invention described above can be written as a program that can be executed on a computer and can be implemented in a general-purpose digital computer that operates the program using a computer-readable recording medium. The computer-readable recording medium includes storage media such as magnetic storage media (e.g., ROM, floppy disk, hard disk, etc.) and optical reading media (e.g., CD-ROM, DVD, etc.).
Claims
Claim 1 A method of operation of a simulation device including a processor and memory, comprising: a step of acquiring information of a virtual subject including at least one of the material, shape, and component information of the subject; a step of acquiring radiation source information related to at least one of the location or intensity of a virtual radiation source; a step of acquiring detector information for a detector receiving radiation from the virtual radiation source; a step of acquiring a first preliminary radiation image based on Monte Carlo at a predetermined first resolution based on the information of the subject, the radiation source information, and the detector information; a step of acquiring a second preliminary radiation image based on Forward Projection at a predetermined second resolution of a higher resolution than the first resolution based on the information of the subject, the radiation source information, and the detector information; a step of acquiring a third preliminary radiation image by converting the second preliminary radiation image to a resolution lower than the second resolution; and a step of resizing the image converted to a lower resolution back to the second resolution to acquire a fourth preliminary radiation image A method of operation of a simulation device comprising: a step of acquiring; a step of acquiring a residual image including high-frequency components based on the fourth preliminary radiation image and the second preliminary radiation image; a step of acquiring a weighted average image by weighting the third preliminary radiation image and the first preliminary radiation image based on predetermined weights; and a step of acquiring an output radiation image of a second resolution based on the weighted average image and the residual image. Claim 2 A method of operation of a simulation device according to claim 1, further comprising: a step of acquiring an actual radiation image by taking an actual radiation image based on information of the subject and information of the detector; a step of acquiring actual noise level information of the actual radiation image; a step of acquiring virtual noise level information of the output radiation image; and a step of modifying the radiation source information so that the virtual noise level information becomes identical to the actual noise level information. Claim 3 A method of operation of a simulation device according to claim 1, further comprising the step of generating a machine-learned scattering simulation machine learning model to receive the second preliminary radiation image and output the output radiation image. Claim 4 A method of operation of a simulation device according to claim 1, wherein the virtual radiation source is rotatable about a predetermined axis of rotation and further comprising the step of acquiring a plurality of output radiation images according to the position of the virtual radiation source; and the step of acquiring a three-dimensional output radiation image based on the plurality of output radiation images. Claim 5 A method of operation of a simulation device according to claim 4, further comprising the step of generating a machine-learned 3D scattering simulation machine learning model to receive a 3D preliminary radiation image and output the 3D output radiation image. Claim 6 A method of operation of a simulation device according to claim 4, further comprising: a step of acquiring a plurality of second preliminary radiation images according to the position of the virtual radiation source; a step of acquiring a three-dimensional preliminary radiation image based on the plurality of second preliminary radiation images; and a step of generating a machine-learned three-dimensional noise removal machine learning model to receive the three-dimensional output radiation image and output the three-dimensional preliminary radiation image. Claim 7 A method of operation of a simulation device according to claim 1, further comprising the step of generating a machine-learned noise removal machine learning model to receive the output radiation image and output the second preliminary radiation image. Claim 8 A method of operation of a simulation device according to claim 7, further comprising: a step of acquiring an actual radiation image obtained by irradiating an actual subject with radiation; and a step of acquiring an improved image with reduced noise by applying the noise removal machine learning model to the actual radiation image.
Citation Information
Patent Citations
System and method for determining a trained neural network model for scattering correction
US20210042888A1
X-ray image apparatus and method for obtaining medicalimage thereof
KR1020180059329A
X-ray Image Processing Method and Computer Readable Storage Medium
KR1020210004085A
Method of providing user interface for selecting an optimized grid specifications
KR102333005B1
