Image processing method, training method, trained model, radiological image processing module, radiological image processing program, and radiological image processing system

Abstract

This image processing method comprises: an image acquiring step for irradiating an object F with an energy beam to acquire an image in which an energy beam transmitted through the object F is captured; a spatial blurring map generating step for generating a spatial blurring map indicating a spatial distribution of blurring of noise on the basis of the image; and a processing step for inputting the image and the spatial blurring map into a trained model 207 built in advance by machine learning, and performing image processing to remove the noise from the image.

Description

Image processing method, training method, trained model, radiation image processing module, radiation image processing program, and radiation image processing system

[0001] One aspect of the embodiment relates to an image processing method, a training method, a trained model, a radiation image processing module, a radiation image processing program, and a radiation image processing system.

[0002] Techniques for removing noise from images using machine learning have been known for some time. For example, Patent Document 1 below discloses an image processing method for removing noise from a radiographic image. In this image processing method, an evaluation value is derived from the pixel values ​​of each pixel of the radiographic image based on relational data representing the relationship between pixel values ​​(brightness values) and an evaluation value that evaluates the spread of noise. A noise map is generated as data in which the derived evaluation values ​​are associated with each pixel of the radiographic image, and the noise map and the radiographic image are input into a trained model.

[0003] International Publication No. 2022 / 172506

[0004] In the image processing method described above, noise in an image is removed by taking into consideration the spread of noise in the luminance direction. For example, noise in an image is removed by taking into consideration the relationship between the luminance value and the standard deviation of pixel values. However, if the blur of noise, which is the spread of noise in the spatial direction, changes in the image, sufficient noise removal effect may not be achieved.

[0005] Therefore, one aspect of the embodiment has been made in consideration of such problems, and aims to provide an image processing method, a training method, a trained model, a radiological image processing module, a radiological image processing program, and a radiological image processing system that can respond to changes in noise blur in an image.

[0006] An image processing method according to one aspect of an embodiment includes an image acquisition step of irradiating an object with an energy beam and acquiring an image of the energy beam that has passed through the object; a spatial blur map generation step of generating a spatial blur map that indicates the distribution of spatial blur of noise based on the image; and a processing step of inputting the image and the spatial blur map into a trained model that has been constructed in advance by machine learning, and performing image processing to remove noise from the image.

[0007] Alternatively, a radiation image processing module according to another aspect of the embodiment includes an image acquisition unit that acquires a radiation image obtained by irradiating an object with radiation and capturing the radiation that has passed through the object; a spatial blur map generation unit that generates a spatial blur map that indicates the distribution of spatial blur of noise based on the radiation image; and a processing unit that inputs the radiation image and the spatial blur map into a trained model that has been constructed in advance by machine learning, and performs image processing to remove noise from the radiation image.

[0008] Alternatively, a radiological image processing program according to another aspect of the embodiment causes a processor to function as an image acquisition unit that acquires a radiological image in which radiation is irradiated onto an object and the radiation that has passed through the object is captured; a spatial blur map generation unit that generates a spatial blur map that indicates the distribution of spatial blur of noise based on the radiological image; and a processing unit that inputs the radiological image and the spatial blur map into a trained model that has been constructed in advance by machine learning, and performs image processing to remove noise from the radiological image.

[0009] Alternatively, a radiation image processing system according to another aspect of the embodiment includes the above-described radiation image processing module, a radiation source that irradiates an object with radiation, and an imaging device that captures radiation that has passed through the object to obtain a radiation image.

[0010] According to one aspect of the present invention, it is possible to provide an image processing method, a training method, a trained model, a radiological image processing module, a radiological image processing program, and a radiological image processing system that can respond to changes in noise blur in an image.

[0011] 1 is a schematic configuration diagram of an image acquisition device 1 according to an embodiment. FIG. 1 is a block diagram showing an example of the hardware configuration of a control device 20 of FIG. 1. FIG. 2 is a block diagram showing the functional configuration of the control device 20 of FIG. 1. FIG. 3 is a diagram showing an example of an X-ray image acquired by the image acquisition unit 201 of FIG. 3. FIG. 4 is a diagram showing an example of generation of a noise map by the noise map generation unit 202 of FIG. 3. FIG. 5 is a diagram showing an example of generation of a spatial blur map by the spatial blur map generation unit 203 of FIG. 3. FIG. 6 is a diagram for explaining spatial blur evaluated by blur evaluation information in the spatial blur map of FIG. 6. FIG. 7 is a diagram for explaining a method of deriving second relationship data G70 showing the correspondence between pixel values ​​and sigma values ​​of FIG. 6. FIG. 8 is a graph showing a relationship referenced when deriving second relationship data G70 showing the correspondence between pixel values ​​and sigma values ​​of FIG. 6. FIG. 9 is a diagram for explaining a method of deriving second relationship data G70 showing the correspondence between pixel values ​​and sigma values ​​of FIG. 6. FIG. 10 is a diagram showing an example of input / output data of a trained model 207 of FIG. 3. FIG. 11 is a diagram showing an example of generation of training data by the training data generation unit 206 of FIG. 3. FIG. 12 is a flowchart showing the procedure of observation processing by the image acquisition device 1. 3A and 3B are diagrams showing examples of X-ray images acquired by the image acquisition device 1 before and after noise removal processing.

[0012] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description, the same elements or elements having the same functions will be denoted by the same reference numerals, and redundant description will be omitted.

[0013] FIG. 1 is a configuration diagram of an image acquisition device 1, which is a radiation image processing system according to an embodiment. As shown in FIG. 1 , the image acquisition device 1 irradiates an object F conveyed in a conveying direction TD with X-rays (energy rays) (radiation) and acquires an X-ray image (image) (radiation image) of the object F captured based on the X-rays transmitted through the object F. The image acquisition device 1 uses the X-ray image to perform foreign body inspection, weight inspection, and product inspection of the object F. Examples of applications include food inspection, baggage inspection, circuit board inspection, battery inspection, and material inspection. The image acquisition device 1 includes a belt conveyor (conveying means) 60, an X-ray irradiator (generation source) 50, an X-ray detection camera (imaging device) 10, a control device (radiation image processing module) 20, a display device 30, and an input device 40 for various inputs. Note that, in the embodiment of the present disclosure, the image may be an image captured using energy rays transmitted through the object F. The energy ray may be, for example, radiation such as X-rays and gamma rays, or any of visible light, infrared rays, ultraviolet rays, and electron beams. In this embodiment, the image is a radiation image such as an X-ray image, but may be other images.

[0014] The belt conveyor 60 has a belt portion on which the object F is placed. By moving the belt portion in the conveying direction TD, the object F is conveyed in the conveying direction TD at a predetermined conveying speed. The conveying speed of the object F is, for example, 48 m / min. The belt conveyor 60 can change the conveying speed, for example, to 24 m / min, 96 m / min, or other speeds as needed. The belt conveyor 60 can also change the height position of the belt portion as needed to change the distance between the X-ray irradiator 50 and the object F. Note that examples of the object F conveyed by the belt conveyor 60 include various items, such as meat, seafood, agricultural products, confectionery, and other foods; rubber products such as tires; resin products, metal products, and mineral resources; waste; and electronic components and electronic circuit boards. The X-ray irradiator 50 is a device that irradiates (outputs) X-rays onto the object F as an X-ray source. The X-ray irradiator 50 is a point light source that irradiates diffused X-rays in a fixed irradiation direction over a predetermined angular range. The X-ray irradiator 50 is disposed above the belt conveyor 60 at a predetermined distance from the belt conveyor 60 so that the X-ray irradiation direction is directed toward the belt conveyor 60 and the diffused X-rays cover the entire width direction of the object F (a direction intersecting the conveying direction TD). The irradiation range of the X-ray irradiator 50 is a predetermined divided range in the length direction of the object F (a direction parallel to the conveying direction TD). As the object F is conveyed in the conveying direction TD on the belt conveyor 60, X-rays are irradiated over the entire length direction of the object F. The tube voltage and tube current of the X-ray irradiator 50 are set by the control device 20. The X-ray irradiator 50 irradiates the belt conveyor 60 with X-rays of a predetermined energy and a predetermined radiation dose according to the set tube voltage and tube current. A filter 51 that transmits X-rays in a predetermined wavelength range is provided near the belt conveyor 60 side of the X-ray irradiator 50. The filter 51 is not necessarily required and may be omitted as appropriate.

[0015] The X-ray detection camera 10 detects X-rays that have passed through the object F among the X-rays irradiated onto the object F by the X-ray irradiator 50 and outputs a signal based on the X-rays. The X-ray detection camera 10 is a dual-line X-ray camera in which two sets of X-ray detection components are arranged. In the image acquisition device 1 according to this embodiment, each X-ray image is generated based on the X-rays detected on each line (first line and second line) of the dual-line X-ray camera. Then, by performing an averaging process, an addition process, or the like on the two generated X-ray images, it is possible to acquire a clear (bright) image with a smaller X-ray dose than when an X-ray image is generated based on X-rays detected on a single line. Note that the X-ray detection camera 10 may be configured as a single-line X-ray camera in which one set of X-ray detection components is arranged, a multi-line X-ray camera in which two or more sets of X-ray detection components are arranged, or two or more single-line X-ray cameras.

[0016] The X-ray detection camera 10 includes a filter 19, scintillators 11a and 11b, line scan cameras 12a and 12b, a sensor control unit 13, amplifiers 14a and 14b, AD converters 15a and 15b, correction circuits 16a and 16b, output interfaces 17a and 17b, and an amplifier control unit 18. The scintillator 11a, line scan camera 12a, amplifier 14a, AD converter 15a, correction circuit 16a, and output interface 17a are electrically connected to each other and constitute a first line. The scintillator 11b, line scan camera 12b, amplifier 14b, AD converter 15b, correction circuit 16b, and output interface 17b are electrically connected to each other and constitute a second line. The line scan camera 12a of the first line and the line scan camera 12b of the second line are arranged side by side along the transport direction TD. In the following, the configuration common to the first line and the second line will be described by taking the configuration of the first line as a representative.

[0017] The scintillator 11a is fixed on the line scan camera 12a by adhesive or the like, and converts X-rays that have passed through the object F into scintillation light. The scintillator 11a outputs the scintillation light to the line scan camera 12a. The filter 19 transmits X-rays in a predetermined wavelength range toward the scintillator 11a. The filter 19 is not necessarily required and may be omitted as appropriate.

[0018] The line scan camera 12a detects scintillation light from the scintillator 11a, converts it into an electric charge, and outputs the detected light as a detection signal (electrical signal) to the amplifier 14a. The line scan camera 12a has a plurality of line sensors arranged in parallel in a direction intersecting the transport direction TD. The line sensors are, for example, charge coupled device (CCD) image sensors or complementary metal-oxide semiconductor (CMOS) image sensors, and include a plurality of photodiodes.

[0019] The sensor control unit 13 controls the line scan cameras 12a and 12b to repeatedly capture images at a predetermined detection period so that the line scan cameras 12a and 12b can capture images of X-rays that have passed through the same region of the object F. The predetermined detection period may be set common to the line scan cameras 12a and 12b based on, for example, the distance between the line scan cameras 12a and 12b, the speed of the belt conveyor 60, the distance between the X-ray irradiator 50 and the object F on the belt conveyor 60 (FOD (Focus Object Distance)), and the distance between the X-ray irradiator 50 and the line scan cameras 12a and 12b (FDD (Focus Detector Distance)). Alternatively, the predetermined period may be set individually for each of the line scan cameras 12a and 12b based on the pixel width of the photodiode in a direction perpendicular to the pixel array direction of the line sensor of each of the line scan cameras 12a and 12b. In this case, the difference (delay time) in the detection cycles between the line scan cameras 12a and 12b may be determined based on the distance between the line scan cameras 12a and 12b, the speed of the belt conveyor 60, the distance (FOD) between the X-ray irradiator 50 and the object F on the belt conveyor 60, and the distance (FDD) between the X-ray irradiator 50 and the line scan cameras 12a and 12b, and individual cycles may be set for each. The amplifier 14a amplifies the detection signal at a predetermined set amplification factor to generate an amplified signal and outputs the amplified signal to the AD converter 15a. The set amplification factor is set by the amplifier control unit 18. The amplifier control unit 18 sets the set amplification factors of the amplifiers 14a and 14b based on predetermined imaging conditions.

[0020] The AD converter 15a converts the amplified signal (voltage signal) output by the amplifier 14a into a digital signal and outputs it to the correction circuit 16a. The correction circuit 16a performs predetermined corrections, such as signal amplification, on the digital signal and outputs the corrected digital signal to the output interface 17a. The output interface 17a outputs the digital signal to the outside of the X-ray detection camera 10. In FIG. 1, the AD converter, the correction circuit, and the output interface are each shown as separate components, but they may also be integrated into one.

[0021] The control device 20 is a computer such as a PC (Personal Computer). The control device 20 generates an X-ray image based on digital signals (amplified signals) output from the X-ray detection camera 10 (more specifically, the output interfaces 17a and 17b). The control device 20 generates one X-ray image by averaging or adding the two digital signals output from the output interfaces 17a and 17b. The generated X-ray image is subjected to a noise removal process (described later) and then output to the display device 30, where it is displayed. The control device 20 also controls the X-ray irradiator 50, the amplifier control unit 18, and the sensor control unit 13. Note that the control device 20 in this embodiment is an independent device provided outside the X-ray detection camera 10, but may be integrated into the X-ray detection camera 10.

[0022] FIG. 2 shows the hardware configuration of the control device 20. As shown in FIG. 2, the control device 20 is physically a computer or the like including processors such as a CPU (Central Processing Unit) 101 and a GPU (Graphic Processing Unit) 105, storage media such as a RAM (Random Access Memory) 102 and a ROM (Read Only Memory) 103, a communication module 104, and an input / output module 106, all of which are electrically connected to one another. The control device 20 may include, as the input device 40 and the display device 30, a display, a keyboard, a mouse, a touch panel display, or the like, or may include a data recording device such as a hard disk drive or semiconductor memory. The control device 20 may also be composed of multiple computers.

[0023] FIG. 3 is a block diagram showing the functional configuration of the control device 20. The control device 20 includes an image acquisition unit 201, a noise map generation unit 202, a spatial blur map generation unit 203, a processing unit 204, a construction unit 205, and a training data generation unit 206. The functional units of the control device 20 shown in FIG. 3 are realized by loading a program (the radiation image processing program of this embodiment) onto hardware such as the CPU 101, GPU 105, and RAM 102, and operating the communication module 104 and input / output module 106 under the control of the CPU 101 and GPU 105, and reading and writing data from and to the RAM 102. The CPU 101 and GPU 105 of the control device 20 execute the computer program to cause the control device 20 to function as the functional units shown in FIG. 3 and sequentially execute processes corresponding to the image processing method described below. The CPU 101 and GPU 105 may be standalone hardware, or either one of them may be used. The CPU 101 and the GPU 105 may be implemented in a programmable logic device such as an FPGA, like a software processor. The RAM and ROM may be standalone hardware devices or may be embedded in a programmable logic device such as an FPGA. All of the various data required to execute the computer program and the various data generated by the execution of the computer program are stored in an internal memory such as the ROM 103 or the RAM 102, or in a storage medium such as a hard disk drive. A trained model 207 (described below) is pre-stored in the internal memory or storage medium of the control device 20. The trained model 207 is read by the CPU 101 and the GPU 105 to cause the CPU 101 and the GPU 105 to perform noise reduction processing on an image.

[0024] The functions of each functional unit of the control device 20 will be described in detail below.

[0025] The image acquisition unit 201 acquires an image in which an energy beam is irradiated onto the object F and the energy beam that has passed through the object F is captured. Specifically, the image acquisition unit 201 generates an X-ray image based on a digital signal (amplified signal) output from the X-ray detection camera 10 (more specifically, the output interfaces 17a and 17b). The image acquisition unit 201 generates one X-ray image by averaging or adding the two digital signals output from the output interfaces 17a and 17b. FIG. 4 is a diagram showing an example of an X-ray image acquired by the image acquisition unit 201. The image acquisition unit 201 cuts out the image into multiple partial images (multiple first partial images). For example, as shown in FIG. 5, the image acquisition unit 201 divides an image (radiographic image) G1 into multiple partial images G2.

[0026] The noise map generation unit 202 derives a noise evaluation value from the pixel value of each pixel of the image based on first relationship data indicating the relationship between the pixel value and a noise evaluation value that evaluates the spread of the noise value. The noise map generation unit 202 generates a noise map, which is data that associates the derived noise evaluation value with each pixel of the image. Specifically, the noise map generation unit 202 acquires a plurality of partial images generated by the image acquisition unit 201. The noise map generation unit 202 derives a standard deviation of pixel values ​​from the pixel values ​​of each pixel of the partial images based on the first relationship data. The noise map generation unit 202 generates a noise map, which is data that associates the derived standard deviation of pixel values ​​with each pixel of the partial images. At this time, the noise map generation unit 202 derives the noise evaluation value from the average energy of the energy beam that has passed through the object F and the pixel value of each pixel of the partial images.

[0027] First, the noise map generation unit 202 calculates the average energy of the energy rays that have passed through the object F. As an example, the noise map generation unit 202 calculates the average energy of the X-rays (radiation) that have passed through the object F based on condition information input via the input device 40 or the like. The condition information is information that indicates either the conditions of the energy ray generation source or the imaging conditions when irradiating the object F with energy rays to image the object F. Note that the noise map generation unit 202 may accept the input of the condition information as a direct input of information such as numerical values, or as a selective input of information such as numerical values ​​that have been set in advance in an internal memory. The noise map generation unit 202 accepts the input of the above condition information from the user, but may also acquire some of the condition information (such as the tube voltage) according to the detection result of the control state by the control device 20.

[0028] The condition information is, for example, information indicating the operating conditions of the X-ray irradiator (generation source) 50 when capturing an X-ray image of the object F, or the imaging conditions of the X-ray detection camera 10. Examples of the operating conditions include all or some of the type of X-ray source, tube voltage, tube current, target angle, target material, etc. The condition information indicating the imaging conditions may include all or some of the following: the material, thickness, and density of the filters 51, 19 arranged between the X-ray irradiator 50 and the X-ray detection camera 10; the distance (FDD) between the X-ray irradiator 50 and the X-ray detection camera 10; the type and thickness of the window material of the X-ray detection camera 10; information on the scintillators 11a, 11b of the X-ray detection camera 10 (e.g., thickness, material, density, multiplication factor, surface reflectance, diffusion coefficient, or absorption coefficient); X-ray detection camera information (e.g., gain setting value, circuit noise value, saturated charge amount, conversion coefficient value (e- / count), camera line rate (Hz) or line speed (m / min)); information on the object F (measured substance, thickness, density);

[0029] For example, the noise map generation unit 202 calculates the spectrum of X-rays detected by the X-ray detection camera 10 using, for example, a known approximation formula such as Tucker's, based on information included in the condition information, such as the tube voltage, target angle, target material, material and thickness of the filters 51 and 19, presence or absence of the filters 51 and 19, type of window material of the X-ray detection camera 10 and presence or absence thereof, and material and thickness of the scintillators 11 a and 11 b of the X-ray detection camera 10. The noise map generation unit 202 further calculates a spectral intensity integral and a photon number integral from the X-ray spectrum, and calculates the value of the average energy of the X-rays by dividing the spectral intensity integral by the photon number integral. Note that the calculation of the X-ray spectrum may also use known approximation formulas such as those by Kramers or Birch et al.

[0030] Next, the noise map generating unit 202 derives the standard deviation of the pixel values ​​from the calculated average energy of the energy beam and the pixel values ​​of each pixel of the partial image using first relationship data that indicates the relationship between the pixel values ​​and the standard deviation of the pixel values ​​(noise evaluation value that evaluates the spread of noise values). The noise map generating unit 202 generates a noise map by associating each pixel of the partial image with the derived standard deviation of the pixel values.

[0031] For example, the noise map generating unit 202 derives, by simulation, first relational data indicating the relationship between pixel values ​​and noise evaluation values ​​that evaluate the spread of noise values. The relational expressions between pixel values, average X-ray energy, and standard deviation of pixel values ​​used by the noise map generating unit 202 are expressed by the following expressions (1) to (4).

[0032] In the above formulas (1) to (3), the variable Noise is the standard deviation of pixel values, the variable Signal is the signal value of a pixel (pixel value), the constant F is the noise factor, and the variable E ｍ is the average energy of the X-ray, the constant M is the multiplication factor by the scintillator, and the constant coeff Mis a coefficient for adjusting the multiplication factor of the scintillator, constant C is information indicating the coupling efficiency between the line scan camera 12a and the scintillator 11a or between the line scan camera 12b and the scintillator 11b in the X-ray detection camera 10, and constant Q is information indicating the quantum efficiency of the line scan camera 12a or the line scan camera 12b.

[0033] In the above formula (1), the constant cf is a conversion coefficient for converting pixel signal values ​​into electric charges in the line scan camera 12 a or 12 b, and the constant R is information indicating the readout noise in the line scan camera 12 a or 12 b. The conversion coefficient cf and the readout noise R are determined by the gain setting in the line scan camera 12 a or 12 b.

[0034] In the above formula (1), the constant M Ｓｉ is the multiplication factor of the line scan camera (silicon) when X-rays incident on the scintillator 11a or 11b are incident on the line scan camera 12a or 12b without being converted into visible light, and is a constant rate si is the silicon direct incidence rate, which indicates the probability that X-rays incident on the scintillator 11a or scintillator 11b will be incident on the line scan camera 12a or line scan camera 12b without being converted into visible light, and is the constant C Ｓ is information indicating a shading correction value, and the constant offset is information indicating a camera offset indicating the offset value of the line scan cameras 12a and 12b. In the above formula (1), the constant N is information indicating the number of sensors. The constant N may be information indicating, for example, the number of line scan cameras (number of lines), or may be information indicating the binning setting of the line scan camera 12a or the line scan camera 12b. In the above formula (1), coeff noise is a coefficient for adjusting noise. In the above equations (2) and (4), the constant F ｐ is the coefficient indicating blur, and the constant coeff Fpare pieces of information indicating a coefficient for adjusting the coefficient indicating blur.

[0035] When the above formulas (1) to (4) are used, the noise map generating unit 202 substitutes the pixel value of each pixel of the X-ray image acquired by the image acquiring unit 201 into the variable Signal, and the variable E ｍ is substituted with the value of the average energy calculated by the noise map generation unit 202. Then, the noise map generation unit 202 obtains the variable Noise calculated using the above formulas (1) to (4) as the value of the standard deviation of the pixel values. Note that other parameters including the average energy may be acquired by receiving input by the noise map generation unit 202, or may be set in advance.

[0036] Furthermore, for example, the noise map generating unit 202 may derive the first relational data indicating the relationship between pixel values ​​and the standard deviation of pixel values ​​based on an image obtained by actually capturing an image. As an example, the noise map generating unit 202 may acquire an X-ray image captured by irradiating a jig with X-rays. The jig may be a flat plate-like member or the like whose thickness and material are known. The noise map generating unit 202 may derive the relational data indicating the relationship between pixel values ​​and the standard deviation of pixel values ​​from the acquired X-ray image.

[0037] The following describes how the noise map generator 202 derives first relational data representing the relationship between pixel values ​​and the standard deviation of pixel values ​​from an X-ray image of a jig. For example, the jig may be a member whose thickness changes stepwise in one direction. First, the noise map generator 202 derives pixel values ​​(hereinafter referred to as true pixel values) for each step of the jig in the X-ray image of the jig, assuming that there is no noise, and derives the standard deviation of the pixel values ​​based on the true pixel values. Specifically, the noise map generator 202 derives the average value of pixel values ​​for a given step of the jig. The noise map generator 202 then sets the derived average value of pixel values ​​as the true pixel value for that step. The noise map generator 202 derives the difference between each pixel value and the true pixel value for that step as a noise value. The noise map generator 202 derives the standard deviation of pixel values ​​from the derived noise value for each pixel value.

[0038] The noise map generating unit 202 then derives first relationship data based on the relationship between the true pixel values ​​and the standard deviation of the pixel values. Specifically, the noise map generating unit 202 derives the true pixel values ​​and the standard deviation of the pixel values ​​for each jig step. The noise map generating unit 202 plots the derived relationship between the true pixel values ​​and the standard deviation of the pixel values ​​on a graph and draws an approximation curve, thereby deriving a relationship graph that represents the relationship between the pixel values ​​and the standard deviation of the pixel values. Note that the approximation curve may be an exponential approximation, a linear approximation, a logarithmic approximation, a polynomial approximation, a power approximation, or the like.

[0039] FIG. 5 is a diagram showing an example of a noise map generated by the noise map generating unit 202. The noise map generating unit 202 derives a relationship graph G4 (first relationship data) representing the correspondence between pixel values ​​and the standard deviation of pixel values ​​in an image captured by an energy beam. The noise map generating unit 202 acquires multiple partial images G2 generated by the image acquiring unit 201. The noise map generating unit 202 then derives relationship data G3 representing the correspondence between each pixel position and pixel value from the partial images G2. Furthermore, the noise map generating unit 202 derives the standard deviation of pixel values ​​corresponding to each pixel position in the partial images by applying the correspondence represented by the relationship graph G4 to each pixel value in the relationship data G3. As a result, the noise map generating unit 202 associates the derived standard deviation of pixel values ​​with each pixel position and derives relationship data G5 representing the correspondence between each pixel position and the standard deviation of pixel values. The noise map generating unit 202 then generates a noise map G6 based on the derived relationship data G5.

[0040] The spatial blur map generating unit 203 generates a spatial blur map indicating the distribution of spatial blur of noise based on the image. Specifically, the spatial blur map generating unit 203 generates a spatial blur map for each of a plurality of partial images.

[0041] More specifically, first, the spatial blur map generation unit 203 acquires a plurality of partial images generated by the image acquisition unit 201. Fig. 6 is a diagram showing an example of a spatial blur map generated by the spatial blur map generation unit 203. In the example shown in Fig. 6, the spatial blur map generation unit 203 acquires a partial image G2 generated by the image acquisition unit 201 by cutting out an image G1.

[0042] Next, the spatial blur map generator 203 derives blur evaluation information from the pixel values ​​of each pixel of the partial image based on second relationship data indicating the relationship between pixel values ​​and blur evaluation information that evaluates the spatial blur of noise. In the example shown in Fig. 6, the spatial blur map generator 203 derives second relationship data G70 indicating the correspondence relationship between pixel values ​​in an image in which energy rays are captured and sigma values. The spatial blur map generator 203 applies the correspondence relationship indicated by the second relationship data G70 to each pixel value in the partial image G2, thereby deriving blur evaluation information corresponding to the pixel at each pixel position in the partial image G2.

[0043] The blur evaluation information is an index for evaluating the spatial blur of noise (the magnitude of noise blur), and is, for example, a sigma value. The sigma value is an index that indicates the spatial spread in a Gaussian distribution or the like. The blur evaluation information may be, for example, at least one of a half-width, a point spread function (PSF), a contrast transfer function (CTF), and a modulation transfer function (MTF). Note that when the blur evaluation information is a contrast transfer function or a modulation transfer function, it has multiple parameters. In this case, the spatial blur map is expressed using multiple channels.

[0044] 7(a) and (b) are diagrams showing the spread of luminance (spatial spread) from a given pixel. In FIG. 7(b), the vertical axis represents the relative luminance value, and the horizontal axis represents the pixel value. As shown in FIG. 7, the spatial blur of noise refers to the spread of the luminance (pixel value) at a given pixel toward the periphery of that pixel. A change in the spatial blur of noise refers to a change in the luminance spread (noise frequency). For example, as the luminance spread in FIG. 7(a) increases (as the noise frequency decreases), the half-width of the graph in FIG. 7(b) increases. Furthermore, as the luminance spread in FIG. 7(a) decreases (as the noise frequency increases), the half-width of the graph in FIG. 7(b) decreases. Note that the spatial blur of noise varies depending on the operating conditions of the energy ray source, the type and thickness of the scintillator, etc.

[0045] The spatial blur map generator 203 generates second relationship data indicating the relationship between pixel values ​​and blur evaluation information that evaluates spatial blur of noise through simulation. Specifically, the spatial blur map generator 203 first acquires condition information. Next, as shown in FIG. 8 , the spatial blur map generator 203 generates a point spread function (PSF) for each transmittance in the X-ray image by executing a Monte Carlo simulation or the like based on parameters included in the condition information. Here, the transmittance in the X-ray image may be, for example, a value obtained by dividing the pixel value of each pixel by background luminance or a value obtained by dividing the pixel value of each pixel by a preset pixel value. The spatial blur map generator 203 approximates the generated point spread function with a Gaussian function, as shown in the following formula (5), to derive a sigma value σ for each point spread function as blur evaluation information. Finally, the spatial blur map generation unit 203 derives the correspondence between the transmittance and the sigma value, and generates second relationship data indicating the relationship between the pixel value and the sigma value by multiplying the transmittance by the background luminance or a preset pixel value. In the above formula (5), the variable σ is the sigma value, and the function PSF simis information indicating a point spread function generated by simulation, a variable x is the x coordinate in the point spread function and Gaussian function, and a variable y is the y coordinate in the point spread function and Gaussian function.

[0046] In the example shown in FIG. 8 , the spatial blur map generator 203 generates point spread functions F1 to F4 for each transmittance in the X-ray image by performing a Monte Carlo simulation based on the condition information. The spatial blur map generator 203 approximates the point spread functions F1 to F4 with Gaussian functions to derive sigma values ​​σ of 0.36, 0.358, 0.348, and 0.322, respectively. The spatial blur map generator 203 derives a correspondence relationship G71 between transmittance and sigma value shown in FIG. 9 by associating the derived sigma values ​​with transmittance. In FIG. 9 , the horizontal axis represents transmittance, and the vertical axis represents sigma value. Finally, the second relationship data G70 indicating the relationship between pixel values ​​and sigma values ​​shown in FIG. 6 is derived by multiplying the transmittance of the correspondence relationship G71 by background luminance or a predetermined luminance.

[0047] The spatial blur map generator 203 may generate second relationship data indicating the relationship between pixel values ​​and blur evaluation information that evaluates the spatial blur of noise, based on an image obtained by actually capturing an image. Specifically, the spatial blur map generator 203 may generate the second relationship data based on an image obtained by actually capturing an image of a jig. Fig. 10 shows an example of the structure of the jig. For example, a jig P1, which is a member whose thickness changes stepwise in one direction, is used as the jig. A resolution chart P2 is provided for each step of the jig P1.

[0048] For example, the spatial blur map generator 203 acquires an image capturing energy rays transmitted through the jig P1 and the resolution chart P2. The spatial blur map generator 203 derives a modulation transfer function (MTF) or a contrast transfer function (CTF) for each step of the jig P1 based on the resolution chart P2 captured in the acquired image. The spatial blur map generator 203 derives a point spread function (PSF) for each step of the jig P1 from the derived modulation transfer function or contrast transfer function. As shown in the above formula (5), the spatial blur map generator 203 derives a sigma value σ for each step of the jig P1 by approximating the generated point spread function with a Gaussian function. The spatial blur map generator 203 derives a correspondence between the transmittance and the sigma value by associating the transmittance derived for each step of the jig P1 with the sigma value for each step of the jig P1. Finally, the spatial blur map generator 203 generates second relationship data indicating the relationship between the pixel value and the sigma value by multiplying the transmittance by the background luminance or a preset pixel value. Note that the derivation of the point spread function from the above-mentioned modulation transfer function can be realized by various methods and configurations utilizing publicly known techniques.

[0049] Finally, the spatial blur map generation unit 203 generates data in which the derived blur evaluation information is associated with each pixel of the partial image as a spatial blur map. In the example shown in Fig. 6, the spatial blur map generation unit 203 associates the derived blur evaluation information with each pixel position of the partial image G2, and generates a spatial blur map G8 that indicates the distribution of spatial blur of noise.

[0050] The processing unit 204 inputs the image, noise map, and spatial blur map into a trained model 207 constructed in advance by machine learning, and performs image processing to remove noise from the image. Specifically, as shown in FIG. 11 , the processing unit 204 acquires a trained model 207 (described later) constructed by the construction unit 205 from an internal memory or storage medium in the control device 20. The processing unit 204 inputs the partial image G2 generated by the image acquisition unit 201, the noise map G6 generated by the noise map generation unit 202, and the spatial blur map G8 generated by the spatial blur map generation unit 203 into the trained model 207. The processing unit 204 uses the noise map G6 as a noise weight and the spatial blur map G8 as a noise frequency weight, respectively. The processing unit 204 then uses the trained model 207 to perform image processing to remove noise from each of the multiple partial images G2, thereby generating multiple noise-removed images G9 from which noise has been removed. The processing unit 204 then integrates the multiple noise-removed images G9 to generate an image G10 from which noise has been removed, and outputs the image G10 to the display device 30 or the like.

[0051] The construction unit 205 uses, as training data, a plurality of partial images (second partial images) generated as training images by cutting out an image (hereinafter referred to as the “whole image”), a noise map and a spatial blur map generated from the partial images, and a noise image obtained by adding noise to the partial images. The construction unit 205 constructs a trained model 207 by machine learning, using the noise map, spatial blur map, and noise image as inputs, and outputs a denoised image obtained by removing noise from the noise image so that the denoised image resembles the partial image. The construction unit 205 stores the constructed trained model 207 in an internal memory or a storage medium within the control device 20. Machine learning includes supervised learning, unsupervised learning, and reinforcement learning, and includes deep learning and neural network learning. In this embodiment, a two-dimensional convolutional neural network described in the paper “Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising” by Kai Zhang et al. is used as an example of a deep learning algorithm. In addition to being constructed by the construction unit 205 , the trained model 207 may also be generated by an external computer or the like and downloaded to the control device 20 .

[0052] The training data generation unit 206 generates training data used by the construction unit 205 to construct the trained model 207. Specifically, the training data generation unit 206 first acquires an entire image and cuts out multiple partial images from the entire image as training images. The entire image used for machine learning may be, for example, an image generated by simulation or a standard image available via an internet connection, etc. In the example shown in FIG. 12 , the training data generation unit 206 cuts out multiple partial images G12 from the acquired entire image G11.

[0053] Next, the training data generation unit 206 sets a noise evaluation value that evaluates the spread of noise as a noise setting value for each partial image. The training data generation unit 206 generates a noise map, which is data in which a noise evaluation value is associated with each pixel of a partial image based on the noise setting value, for each partial image. In the example shown in Fig. 12, the training data generation unit 206 generates a noise map G13 from a plurality of partial images G12.

[0054] Specifically, the training data generation unit 206 sets a different noise evaluation value as the noise setting value for each partial image using one of the following three methods: The training data generation unit 206 generates data in which a noise setting value is associated with each pixel of the partial image and set as a noise map. In the first method, the training data generation unit 206 sets an arbitrary noise setting value for one partial image. That is, the noise setting value is set uniformly in one partial image, and at the same time, the noise setting value is set randomly in the entire image. In the second method, the training data generation unit 206 further cuts out one partial image into multiple regions. The training data generation unit 206 sets an arbitrary noise setting value uniformly in one region. That is, the noise setting value is set randomly in one partial image. Note that further cutting out (dividing) a partial image into multiple regions may be achieved, for example, by Otsu's binarization method, or by various methods and configurations utilizing publicly known technologies. In the third method, the training data generation unit 206 sets, as a noise setting value for each pixel of the partial image, a noise evaluation value derived from the pixel value of each pixel of the partial image based on first relationship data indicating the relationship between the pixel value and the noise evaluation value that evaluates the spread of the noise value, in the same way as the noise map generation unit 202. In this case, the training data generation unit 206 may derive the first relationship data by simulation, as described above, or may derive the first relationship data based on an image obtained by actually capturing an image (actual measurement).

[0055] Next, the training data generation unit 206 sets blur evaluation information that evaluates the spatial blur of noise as blur setting information for each partial image. The training data generation unit 206 generates a spatial blur map that indicates the distribution of spatial blur of noise for the multiple partial images based on the blur setting information.

[0056] Specifically, the training data generation unit 206 sets different blur evaluation information as blur setting information for each partial image using one of the following three methods: The training data generation unit 206 generates data in which blur setting information is associated with each pixel of a partial image and set as a spatial blur map; In a fourth method, the training data generation unit 206 sets arbitrary blur setting information for one partial image; that is, the blur setting information is set uniformly in one partial image, and at the same time, the blur setting information is set randomly in the entire image; In a fifth method, the training data generation unit 206 further cuts out one partial image into multiple regions; The training data generation unit 206 sets arbitrary blur evaluation information uniformly in one region; that is, the blur evaluation information is set randomly in one partial image. In the sixth method, the training data generation unit 206 sets the blur evaluation information derived from the pixel values ​​of each pixel of the partial image as blur setting information for each pixel of the partial image based on the second relationship data indicating the relationship between the pixel values ​​and the blur evaluation information, in the same way as the spatial blur map generation unit 203. At this time, the training data generation unit 206 generates the second relationship data based on a simulation or an actually captured image (actual measurement), as described above.

[0057] Next, the training data generation unit 206 generates a noise image by adding noise to the partial image based on the noise map and the spatial blur map. In the example shown in Fig. 12, the training data generation unit 206 generates noise data G15 based on the noise map G13 and the spatial blur map G14, and adds the generated noise data G15 to the partial image G12 to generate a noise image G16.

[0058] The training data generation unit 206 generates a noise distribution having the same size as the partial image based on the noise map. Specifically, the training data generation unit 206 determines a noise value for each pixel of the partial image based on the noise map. The training data generation unit 206 generates the noise distribution by associating the determined noise value with each pixel of the partial image.

[0059] The training data generation unit 206 blurs the noise distribution based on the spatial blur map to generate noise data. Specifically, the training data generation unit 206 determines in advance a correspondence relationship between blur evaluation information and a point spread function. The training data generation unit 206 determines a point spread function for each pixel by applying the determined correspondence relationship to the spatial blur map. The training data generation unit 206 blurs the noise distribution based on the point spread function determined for each pixel to generate noise data. In the example shown in FIG. 12 , the training data generation unit 206 blurs the noise distribution based on the spatial blur map G14 to generate noise data G15. Note that when the blur evaluation information is a point spread function, the training data generation unit 206 generates noise data using the spatial blur map because the point spread function has already been determined for each pixel in the spatial blur map.

[0060] More specifically, when the blur evaluation information is a sigma value, the training data generation unit 206 determines in advance the correspondence between the sigma value and the point spread function by one of the following three methods. In the seventh method, the training data generation unit 206 regards a Gaussian distribution corresponding to the sigma value as the point spread function. In this case, the closer the Gaussian distribution is to the point spread function of an actual scintillator, the more effectively the trained model 207 that can be constructed can achieve noise reduction.

[0061] In the eighth method, a point spread function is calculated for each scintillator by simulation, taking into account the type and thickness of the scintillator. The simulation is, for example, a photon Monte Carlo simulation within the scintillator. The training data generation unit 206 determines the sigma value of a Gaussian distribution that most closely approximates the calculated point spread function as the sigma value corresponding to the point spread function. In this case, the correspondence between the sigma value and the point spread function can be determined taking into account changes in the point spread function (light spread function) (blur) due to differences in the type and thickness of the scintillator. This allows noise images to be generated as training data using a point spread function that more closely approximates the blur in an actually captured image. As a result, more effective noise removal can be achieved in the trained model 207 constructed using the training data.

[0062] In an eighth method, the training data generation unit 206 may determine the sigma value of the point spread function by further considering the X-ray energy spectrum. In this case, the training data generation unit 206 acquires the operating conditions (tube voltage, filter, and radiation source information) of the X-ray irradiator 50 included in the condition information and calculates the X-ray energy spectrum reaching the scintillators 11a and 11b based on the operating conditions. The training data generation unit 206 determines the sigma value of the point spread function by using the above-described eighth method by taking the X-ray energy spectrum into consideration. Here, for example, even if the same scintillator is used, when the X-ray irradiator 50 is operating at a low tube voltage (low energy), a high proportion of light is emitted from the surface side of the scintillator, which increases the distance that light travels to the line scan cameras 12a and 12b, resulting in a large point spread function. Furthermore, for example, even with the same scintillator, when the tube voltage (high energy) is high, a higher proportion of light is emitted from the back side of the scintillator compared to when the energy is low, and the distance the light travels to the line scan cameras 12a and 12b is shorter, resulting in a smaller point spread function. In the eighth method, the point spread function is calculated taking into account the fact that the spread of light in the scintillator differs depending on the X-ray energy. This allows noise images to be generated as training data using a point spread function that more closely resembles the blur in an actually captured image. Therefore, more effective noise removal can be achieved in the trained model 207 constructed using the training data.

[0063] In the ninth method, the training data generation unit 206 acquires an actually measured point spread function. The training data generation unit 206 sets the sigma value of the Gaussian distribution that most closely approximates the acquired point spread function as the sigma value corresponding to the point spread function. In this case, the point spread function may be measured for each case where the scintillator type and thickness are different or for each case where the X-ray energy is different.

[0064] As shown in FIG. 12, the training data generating unit 206 generates a noise image G16 by adding noise to the partial image G12 based on the generated noise data.

[0065] Next, the procedure for observing an energy beam transmission image of the object F using the image acquisition device 1 according to this embodiment, i.e., the flow of the training method and image processing method according to this embodiment, will be described. Fig. 13 is a flowchart showing the procedure for observing using the image acquisition device 1. Hereinafter, steps S100 and S101 correspond to the training method in this embodiment. Steps S102 to S106 correspond to the image processing method in this embodiment.

[0066] First, training data is generated by the control device 20 (Step S100: training data generation step). Specifically, first, the control device 20 acquires an entire image and cuts out the entire image into a plurality of partial images.

[0067] Next, the control device 20 sets a noise evaluation value that evaluates the spread of noise as a noise setting value for each partial image. At this time, the control device 20 sets a different noise evaluation value as the noise setting value for each partial image. For example, the control device 20 uses the third method described above to set a noise evaluation value derived from the pixel value of each pixel of the partial image as the noise setting value for each pixel of the partial image based on first relationship data that indicates the relationship between the pixel value and the noise evaluation value that evaluates the spread of the noise value. In addition, the control device 20 generates a noise map, which is data that associates the noise evaluation value with each pixel of the image, for multiple partial images based on the noise setting value. Specifically, the control device 20 generates a noise map, which is data that associates the noise evaluation value with each pixel of the image.

[0068] Next, the control device 20 sets blur evaluation information that evaluates the spatial blur of noise as blur setting information for each partial image. At this time, the control device 20 sets different blur evaluation information for each partial image as blur setting information. For example, the control device 20 uses the sixth method described above to set blur evaluation information derived from the pixel values ​​of each pixel of the partial image as blur setting information for each pixel of the partial image based on second relationship data that indicates the relationship between the pixel value and the blur evaluation information that evaluates the spatial blur of noise. In addition, the control device 20 generates a spatial blur map that indicates the distribution of spatial blur of noise for multiple partial images based on the blur setting information. Specifically, the control device 20 generates data that associates the blur setting information with each pixel of the partial image and sets it as the spatial blur map.

[0069] Finally, a noise image is generated by adding noise to the partial image based on the generated noise map and spatial blur map by the control device 20. In this manner, the noise map, spatial blur map, and noise image are generated by the control device 20.

[0070] Next, the control device 20 constructs a trained model 207 (step S101: construction step). Specifically, the control device 20 uses the partial image, the spatial blur map, and the noise image as training data, and constructs the trained model 207 by machine learning, for inputting the spatial blur map and the noise image, and outputting a denoised image obtained by removing noise from the noise image so that the denoised image approximates the partial image.

[0071] Next, the object F is set in the image acquisition device 1, an image of the object F is captured, and the control device 20 acquires an image of the object F (step S102: image acquisition step). Next, the control device 20 generates a noise map (step S103: noise map generation step). Specifically, the control device 20 cuts out the acquired image into a plurality of partial images. The control device 20 derives a noise evaluation value from the pixel values ​​of each pixel of the partial images based on first relationship data indicating the relationship between the pixel values ​​and the noise evaluation value that evaluates the spread of the noise values, and generates a noise map, which is data that associates the derived noise evaluation value with each pixel of the partial images.

[0072] Next, the control device 20 generates a spatial blur map indicating the distribution of spatial blur of noise based on the acquired image (step S104: spatial blur map generation step). Specifically, the control device 20 derives blur evaluation information from the pixel value of each pixel of the partial image based on second relationship data indicating the relationship between the pixel value and blur evaluation information that evaluates the spatial blur of noise, and generates data in which the derived blur evaluation information is associated with each pixel of the partial image as a spatial blur map.

[0073] Next, the control device 20 executes image processing to remove noise from the image (step S105: processing step). Specifically, the control device 20 inputs the partial images, noise map, and spatial blur map to the trained model 207, and executes image processing to remove noise from the partial images. The multiple partial images from which noise has been removed are integrated to generate an image from which noise has been removed.

[0074] Finally, the processing unit 204 outputs the output image, which is the image that has been subjected to the noise removal processing, to the display device 30 (step S106).

[0075] According to the image acquisition device 1 described above, a spatial blur map indicating the distribution of spatial blur of noise is generated based on an image captured of an energy beam transmitted through an object F, and the image and the spatial blur map are input to a trained model 207 constructed in advance by machine learning, and image processing is performed to remove noise from the image. According to this configuration, noise in the image is removed by machine learning, taking into account changes in noise blur. For example, even if the noise blur changes due to changes in the measurement conditions of the energy beam, noise in the image is effectively removed. This allows noise removal that corresponds to changes in noise blur in the image to be achieved using the trained model 207. As a result, noise in the image can be effectively removed.

[0076] The above-mentioned effects will be specifically described. In conventional image acquisition devices, noise is removed from an image by taking into account the spread of noise in the luminance direction. For example, noise is removed from an image by taking into account the relationship between luminance values ​​and the standard deviation of pixel values. However, if the noise blur, which is the spread of noise in the spatial direction, changes in the image, sufficient noise removal effect may not be achieved. For example, in an X-ray camera, if the noise frequency changes due to a difference in the type of scintillator, noise may not be removed appropriately. As an example, in conventional image processing methods, if an image contains high-frequency noise and low-frequency noise, a trained model constructed to remove high-frequency noise cannot remove the low-frequency noise. In contrast, the image acquisition device 1 according to this embodiment removes noise from an image by taking into account changes in noise frequency (changes in noise blur). This allows optimal noise removal from an image even in situations where the noise frequency varies. For example, multiple types of noise with different frequencies can be effectively removed.

[0077] Furthermore, the control device 20 of this embodiment derives a noise evaluation value from the pixel value of each pixel of the image based on first relationship data indicating the relationship between pixel values ​​and a noise evaluation value that evaluates the spread of noise values, and generates a noise map, which is data associating each pixel of the image with the derived noise evaluation value. The control device 20 inputs the image, the spatial blur map, and the noise map into the trained model 207 and performs image processing to remove noise from the image. With this configuration, the spread of noise values ​​evaluated from the pixel values ​​of each pixel of the image is further taken into consideration, and noise in each pixel of the image is removed by machine learning. This makes it possible to use the trained model 207 to achieve noise removal that is more responsive to the relationship between pixel values ​​and the spread of noise in the image. As a result, noise in the image can be removed more effectively.

[0078] The above-mentioned effects will be specifically explained. FIG. 14( a) shows an image before noise removal. FIG. 14( b) shows an image after noise removal using a conventional image processing method. FIG. 14( c) shows an image after noise removal using the image processing method according to this embodiment. As shown in FIGS. 14( a) and 14(b), in the conventional image processing method, when a fine structure of a subject is captured in an image, the fine structure is determined to be noise and removed from the image, resulting in the fine structure of the subject being blurred in the image. In contrast, according to the image processing method according to this embodiment, the fine structure of the subject is not determined to be noise, as shown in FIG. 14(c). As a result, noise can be effectively removed from the image without blurring the fine structure, enabling appropriate noise removal.

[0079] Furthermore, the control device 20 of this embodiment generates a spatial blur map for each of a plurality of partial images cut out from an image, inputs the plurality of partial images into the trained model 207 and executes image processing to remove noise from the plurality of partial images, and integrates the plurality of partial images from which noise has been removed to generate an image from which noise has been removed. With this configuration, noise in an image can be removed by simple processing by executing image processing to remove noise for each of a plurality of partial images cut out from an image.

[0080] Furthermore, the control device 20 of this embodiment derives blur evaluation information from the pixel values ​​of each pixel of the image based on second relationship data indicating the relationship between pixel values ​​and blur evaluation information that evaluates the spatial blur of noise, and generates data in which the derived blur evaluation information is associated with each pixel of the image as a spatial blur map. With this configuration, the spatial blur of noise evaluated from the pixel values ​​of each pixel of the image is taken into consideration, and noise in each pixel of the image is removed by machine learning. This allows noise in the image to be removed more effectively.

[0081] Furthermore, the control device 20 of this embodiment extracts multiple partial images from the overall image as training images, sets blur evaluation information evaluating the spatial blur of noise as blur setting information, and generates a spatial blur map indicating the distribution of spatial blur of noise for the multiple partial images based on the blur setting information. The control device 20 generates a noise image by adding noise to the partial image based on the generated spatial blur map. The control device 20 uses the partial images, the spatial blur map, and the noise image as training data, and constructs a trained model 207 by machine learning. The trained model 207 inputs the spatial blur map and the noise image, and outputs a denoised image obtained by removing noise from the noise image so that the denoised image approximates the partial image. With this configuration, the trained model 207 used for denoising an image is constructed by machine learning using training data. As a result, when an image and a spatial blur map generated from the image are input to the trained model 207, noise removal corresponding to changes in spatial blur of noise can be achieved. As a result, noise in the image can be effectively removed.

[0082] Furthermore, the control device 20 of this embodiment sets the blur setting information as different blur evaluation information for each partial image. This configuration enables noise removal that can accommodate various changes in noise blur, resulting in more effective noise removal in the image.

[0083] Furthermore, the control device 20 of this embodiment sets blur evaluation information derived from the pixel values ​​of each pixel of the partial image as blur setting information for each pixel of the partial image based on second relationship data indicating the relationship between pixel values ​​and their standard deviations, and generates data in which the blur setting information is associated with each pixel of the partial image as a spatial blur map. This configuration makes it possible to achieve noise removal that takes into account the spatial blur of noise evaluated from the pixel values ​​of each pixel of the image. As a result, noise in the image can be removed more effectively.

[0084] Furthermore, the control device 20 of this embodiment generates the second relationship data through simulation. With this configuration, second relationship data that is more suitable for removing noise from an image can be obtained. As a result, noise in an image can be removed more effectively.

[0085] Furthermore, the control device 20 of this embodiment generates the second relationship data based on an image obtained by actually capturing the image. With this configuration, the second relationship data that is more suitable for removing noise from the image can be obtained. As a result, noise in the image can be removed more effectively.

[0086] In the control device 20 of this embodiment, the blur evaluation information is at least one of a sigma value, a half-width, a point spread function, a contrast transfer function, and a modulation transfer function. With this configuration, noise in an image can be more effectively removed based on information that more specifically evaluates the spatial blur of noise.

[0087] Furthermore, in the control device 20 of this embodiment, the trained model 207 is a model constructed as described above, and causes the processor to execute image processing to remove noise from an image captured of an energy beam that has passed through the object F. This makes it possible to realize noise removal corresponding to changes in the blur of noise in the image using the trained model 207. As a result, noise in the image can be effectively removed.

[0088] The above describes an embodiment of the present invention, but the present invention is not limited to the above embodiment, and may be modified or applied to other things within the scope that does not change the gist described in each claim.

[0089] The processing unit 204 according to the above embodiment inputs a partial image, a spatial blur map, and a noise map into the trained model 207 and performs image processing to remove noise from the partial image, but the noise map does not have to be input into the trained model 207.

[0090] For example, the training data generation unit 206 may not generate a noise map as training data. In this case, the training data generation unit 206 may generate a spatial blur map as in the above embodiment, and may also generate a noise image by adding noise to a partial image based on the generated spatial blur map. More specifically, the training data generation unit 206 may determine a noise value for each pixel of the partial image and generate a noise distribution of the same size as the partial image. Each noise value in the noise distribution may be determined based on a predetermined standard deviation or may be determined arbitrarily by a method other than the above. The training data generation unit 206 may then use the partial image, the spatial blur map, and the noise image as training data to construct, by machine learning, a trained model 207 that receives the spatial blur map and the noise image as inputs and removes noise from the noise image to output a denoised image such that the denoised image resembles the partial image.

[0091] Furthermore, for example, the control device 20 may not include the noise map generation unit 202. In this case, the processing unit 204 may input the partial image and the spatial blur map to the trained model 207 and perform image processing to remove noise from the partial image.

[0092] In the flow of the image processing method shown in FIG. 13 , the control device 20 may generate a noise image by adding noise to a partial image based on the generated spatial blur map in step S100. In step S101, the control device 20 may use the partial image, the spatial blur map, and the noise image as training data to construct a trained model 207 by machine learning, with the spatial blur map and the noise image as input, to output a denoised image obtained by removing noise from the noise image so that the denoised image resembles the partial image. The control device 20 may not execute step S103. In step S105, the control device 20 may input the partial image and the spatial blur map to the trained model 207 and execute image processing to remove noise from the partial image.

[0093] The image acquisition unit 201 according to the above embodiment does not need to cut out an image into multiple partial images. For example, the noise map generation unit 202 may generate a noise map for each image. As an example, as shown in FIG. 5 , the noise map generation unit 202 may derive relationship data G3 representing the correspondence between each pixel position and pixel value from image G1 instead of partial image G2. Furthermore, the noise map generation unit 202 may derive the standard deviation of pixel values ​​corresponding to pixels at each pixel position in image G1 by applying the correspondence indicated by the relationship graph G4 to each pixel value in the relationship data G3. As a result, the noise map generation unit 202 may associate the derived standard deviation of pixel values ​​with each pixel position and derive relationship data G5 indicating the correspondence between each pixel position and the standard deviation of pixel values. Then, the noise map generation unit 202 may generate a noise map G6 based on the derived relationship data G5.

[0094] For example, the spatial blur map generation unit 203 may generate a spatial blur map for each image. As an example, as shown in Fig. 6, the spatial blur map generation unit 203 may derive blur evaluation information corresponding to pixels at each pixel position in the image G1 by applying the correspondence relationship indicated by the second relationship data G70 to each pixel value in the image G1. The spatial blur map generation unit 203 may associate the derived blur evaluation information with each pixel position in the image G1 and generate a spatial blur map G8 that indicates the distribution of spatial blur of noise.

[0095] For example, the processing unit 204 may input the image, the spatial blur map, and the noise map to the trained model 207 and perform image processing to remove noise from the image. As an example, as shown in Fig. 11 , the control device 20 may input the image G1, the noise map G6, and the spatial blur map G8 instead of the partial image G2 to the trained model 207 and perform image processing to remove noise from the image G1. In this case, the control device 20 generates a plurality of noise-removed images G9 from which noise has been removed, and outputs the noise-removed images G9 to the display device 30 or the like.

[0096] In the control device 20 according to the above embodiment, the overall image used in machine learning may be, but is not limited to, an image generated by simulation or a standard image available via an internet connection or the like. For example, the overall image used in machine learning may be a high-output image. Furthermore, for example, the overall image used in machine learning may be a Noise2Noise image. In this case, the control device 20 may generate a noise map and a sigma map from the Noise2Noise image. The control device 20 may construct the trained model 207 by unsupervised learning or the like using the Noise2Noise image, the noise map, and the sigma map as training data.

[0097] According to the above embodiment, a spatial blur map showing the distribution of spatial blur of noise is generated based on an image captured of an energy beam transmitted through an object, and the image and the spatial blur map are input into a trained model previously constructed by machine learning, and image processing is performed to remove noise from the image. With this configuration, noise in the image is removed by machine learning, taking into account changes in noise blur. For example, even if the noise blur changes due to changes in measurement conditions of the energy beam, noise in the image is effectively removed. This allows noise removal that corresponds to changes in noise blur in the image to be achieved using the trained model. As a result, noise in the image can be effectively removed.

[0098] Preferably, the above-described embodiment further includes a noise map generation step of deriving a noise evaluation value from the pixel values ​​of each pixel of the image based on first relationship data indicating the relationship between pixel values ​​and a noise evaluation value that evaluates the spread of noise values, and generating a noise map that is data associating each pixel of the image with the derived noise evaluation value, and the processing step of inputting the image, the spatial blur map, and the noise map into a trained model and performing image processing to remove noise from the image. Preferably, the above-described embodiment further includes a noise map generation unit that derives a noise evaluation value from the pixel values ​​of each pixel of the radiographic image based on the first relationship data indicating the relationship between pixel values ​​and a noise evaluation value that evaluates the spread of noise values, and generates a noise map that is data associating each pixel of the radiographic image with the derived noise evaluation value, and the processing unit inputs the radiographic image, the spatial blur map, and the noise map into the trained model and performs image processing to remove noise from the radiographic image. With this configuration, noise in each pixel of the image is removed by machine learning, further taking into account the spread of noise values ​​evaluated from the pixel values ​​of each pixel of the image. This allows for the use of a trained model to achieve noise removal that better corresponds to the relationship between pixel values ​​and the extent of noise in an image, resulting in more effective noise removal in an image.

[0099] In the above embodiment, it is preferable that the spatial blur map generating step generates a spatial blur map for each of a plurality of first partial images cut out from the image, and that the processing step inputs the plurality of first partial images instead of the image into a trained model, performs image processing to remove noise from the plurality of first partial images, and integrates the plurality of noise-removed first partial images to generate a noise-removed image. Also, in the above embodiment, it is preferable that the spatial blur map generating unit generates a spatial blur map for each of a plurality of first partial images cut out from the radiographic image, and that the processing unit inputs the plurality of first partial images instead of the radiographic image into a trained model, performs image processing to remove noise from the plurality of first partial images, and integrates the plurality of noise-removed first partial images to generate a noise-removed radiographic image. According to this configuration, noise in the image can be removed by simple processing by performing image processing to remove noise for each of a plurality of first partial images cut out from the image.

[0100] In the above embodiment, it is preferable that the spatial blur map generating step derives blur evaluation information from the pixel values ​​of each pixel of the image based on second relationship data indicating the relationship between pixel values ​​and blur evaluation information evaluating the spatial blur of noise, and generates data corresponding to the derived blur evaluation information for each pixel of the image as a spatial blur map. In the above embodiment, it is preferable that the spatial blur map generating unit derives blur evaluation information from the pixel values ​​of each pixel of the radiographic image based on the second relationship data indicating the relationship between pixel values ​​and blur evaluation information evaluating the spatial blur of noise, and generates data corresponding to the derived blur evaluation information for each pixel of the radiographic image as a spatial blur map. According to this configuration, noise in each pixel of the image is removed by machine learning, taking into account the spatial blur of noise evaluated from the pixel values ​​of each pixel of the image. This allows noise in the image to be removed more effectively.

[0101] Preferably, the training method according to the above embodiment comprises a training data generation step of extracting a plurality of second partial images from an overall image as training images, setting blur evaluation information that evaluates the spatial blur of noise as blur setting information for each second partial image, generating a spatial blur map that indicates the distribution of the spatial blur of noise for the plurality of second partial images based on the blur setting information, and generating a noise image by adding noise to the second partial image based on the generated spatial blur map; and a construction step of constructing by machine learning a trained model using the second partial images, the spatial blur map, and the noise image as training data, and using the spatial blur map and the noise image as input to output a noise-removed image in which noise has been removed from the noise image so that the noise-removed image approximates the second partial image. In the above embodiment, it is preferable to further include a training data generation unit that extracts multiple second partial images from the entire image as training images, sets blur evaluation information evaluating the spatial blur of noise as blur setting information for each second partial image, generates a spatial blur map indicating the distribution of the spatial blur of noise for the multiple second partial images based on the blur setting information, and generates a noise image by adding noise to the second partial image based on the generated spatial blur map; and a construction unit that constructs a trained model by machine learning using the second partial images, the spatial blur map, and the noise image as training data, and inputs the spatial blur map and the noise image to output a denoised image by removing noise from the denoised image so that the denoised image resembles the second partial image. According to this configuration, the trained model used for denoising the image is constructed by machine learning using the training data. As a result, when an image and a spatial blur map generated from the image are input to the trained model, noise removal corresponding to changes in the spatial blur of noise can be achieved. As a result, noise in the image can be effectively removed.

[0102] In the above embodiment, it is preferable that in the training data generating step, the blur setting information is set as blur evaluation information that differs from one another for each second partial image. Also, in the above embodiment, it is preferable that the training data generating unit sets the blur setting information as blur evaluation information that differs from one second partial image to another. With this configuration, noise removal that can respond to various changes in noise blur can be realized. As a result, noise in the image can be removed more effectively.

[0103] In the above embodiment, it is preferable that the training data generating step sets blur evaluation information derived from the pixel values ​​of each pixel of the second partial image as blur setting information for each pixel of the second partial image based on second relationship data indicating the relationship between pixel values ​​and blur evaluation information evaluating the spatial blur of noise, and generates data in which the blur setting information is associated with each pixel of the second partial image as a spatial blur map. In the above embodiment, the training data generating unit sets blur evaluation information derived from the pixel values ​​of each pixel of the second partial image as blur setting information for each pixel of the second partial image based on the second relationship data indicating the relationship between pixel values ​​and blur evaluation information evaluating the spatial blur of noise, and generates data in which the blur setting information is associated with each pixel of the second partial image as a spatial blur map. This configuration enables noise removal that takes into account the spatial blur of noise evaluated from the pixel values ​​of each pixel of the image. As a result, noise in the image can be removed more effectively.

[0104] In the above embodiment, it is preferable that the training data generating step generates the second relationship data by simulation. Also, in the above embodiment, it is preferable that the training data generating unit generates the second relationship data by simulation. With this configuration, second relationship data that is more suitable for removing noise from images can be obtained. As a result, noise in images can be removed more effectively.

[0105] In the above embodiment, it is preferable that the training data generating step generates the second relationship data based on images obtained by actually capturing images. Also, in the above embodiment, it is preferable that the training data generating unit generates the second relationship data based on radiographic images obtained by actually capturing images. With this configuration, second relationship data that is more suitable for removing noise from images can be obtained. As a result, noise in images can be removed more effectively.

[0106] In the above embodiment, it is preferable that the blur evaluation information is at least one of a sigma value, a half-width, a point spread function, a contrast transfer function, and a modulation transfer function. With this configuration, noise in an image can be more effectively removed based on information that more specifically evaluates the spatial blur of noise.

[0107] Alternatively, the trained model of the embodiment is preferably a trained model constructed by the training data generating step and the construction step described above, and the processor is caused to perform image processing to remove noise from an image captured of an energy beam that has passed through an object. This allows noise removal to be achieved using the trained model in response to changes in the blur of noise in the image. As a result, noise in the image can be effectively removed.

[0108] The image processing method of the embodiment is [1] "an image processing method comprising: an image acquisition step of irradiating an object with an energy beam and acquiring an image of the energy beam that has passed through the object; a spatial blur map generation step of generating a spatial blur map that indicates the distribution of spatial blur of noise based on the image; and a processing step of inputting the image and the spatial blur map into a trained model that has been constructed in advance by machine learning, and performing image processing to remove noise from the image."

[0109] The image processing method of the embodiment may be [2] "the image processing method described in the above [1], further comprising a noise map generation step of deriving a noise evaluation value from the pixel value of each pixel of the image based on first relationship data indicating the relationship between the pixel value and the noise evaluation value that evaluates the spread of the noise value, and generating a noise map that is data in which the derived noise evaluation value is associated with each pixel of the image, and in the processing step, the image, the spatial blur map, and the noise map are input into a trained model, and image processing is performed to remove noise from the image."

[0110] The image processing method of the embodiment may be [3] "the image processing method described in [1] or [2] above, in which in the spatial blur map generation step, a spatial blur map is generated for each of a plurality of first partial images cut out from an image, and in the processing step, the plurality of first partial images are input into a trained model instead of an image, and image processing is performed to remove noise from the plurality of first partial images, and the plurality of first partial images from which noise has been removed are integrated to generate an image from which noise has been removed."

[0111] The image processing method of the embodiment may be [4] "the image processing method according to any one of the above [1] to [3], wherein in the spatial blur map generation step, blur evaluation information is derived from the pixel value of each pixel of the image based on second relationship data indicating the relationship between the pixel value and blur evaluation information that evaluates the spatial blur of noise, and data in which the derived blur evaluation information is associated with each pixel of the image is generated as a spatial blur map."

[0112] The training method of the embodiment is [5] a training method comprising: a training data generation step of cutting out a plurality of second partial images from an entire image as training images, setting blur evaluation information that evaluates the spatial blur of noise as blur setting information for each second partial image, generating a spatial blur map that indicates the distribution of the spatial blur of noise for the plurality of second partial images based on the blur setting information, and generating a noise image by adding noise to the second partial image based on the generated spatial blur map; and a construction step of constructing a trained model by machine learning using the second partial images, the spatial blur map, and the noise image as training data, and using the spatial blur map and the noise image as input to output a noise-removed image by removing noise from the noise image so that the noise-removed image approximates the second partial image.

[0113] The training method of the embodiment may be [6] "the training method described in [5] above, in which in the training data generation step, the blur setting information is set as blur evaluation information that differs from each other for each second partial image."

[0114] The training method of the embodiment may be [7] "the training method described in either of [5] or [6] above, in which in the training data generation step, blur evaluation information derived from the pixel values ​​of each pixel of the second partial image is set as blur setting information for each pixel of the second partial image based on second relationship data indicating the relationship between pixel values ​​and blur evaluation information that evaluates the spatial blur of noise, and data set by associating the blur setting information with each pixel of the second partial image is generated as a spatial blur map."

[0115] The training method of the embodiment may be [8] "the training method described in [7] above, in which, in the training data generation step, the second relational data is generated by simulation."

[0116] The training method of the embodiment may be [9] "the training method described in [7] above, in which, in the training data generation step, second relational data is generated based on images obtained by actually capturing images."

[0117] The training method of the embodiment may be

[10] "the training method described in any one of [5] to [9] above, in which the blur evaluation information is at least one of a sigma value, a half-width, a point spread function, a contrast transfer function, and a modulation transfer function."

[0118] The trained model of the embodiment may be

[11] "a trained model constructed by the training method described in any one of [5] to

[10] above, which causes a processor to perform image processing to remove noise from an image captured of an energy beam that has passed through an object."

[0119] The radiation image processing module of the embodiment is

[12] "a radiation image processing module including: an image acquisition unit that acquires a radiation image obtained by irradiating an object with radiation and capturing the radiation that has passed through the object; a spatial blur map generation unit that generates a spatial blur map that indicates the distribution of spatial blur of noise based on the radiation image; and a processing unit that inputs the radiation image and the spatial blur map into a trained model that has been constructed in advance by machine learning, and performs image processing to remove noise from the radiation image."

[0120] The radiological image processing module of the embodiment may be

[13] "the radiological image processing module described in

[12] above, further including a noise map generation unit that derives a noise evaluation value from the pixel values ​​of each pixel of the radiological image based on first relationship data indicating the relationship between pixel values ​​and a noise evaluation value that evaluates the spread of noise values, and generates a noise map that is data that associates each pixel of the radiological image with the derived noise evaluation value, and the processing unit inputs the radiological image, the spatial blur map, and the noise map into a trained model, and performs image processing to remove noise from the radiological image."

[0121] The radiation image processing module of the embodiment may be

[14] "the radiation image processing module described in

[12] or

[13] above, wherein the spatial blur map generation unit generates a spatial blur map for each of a plurality of first partial images cut out from the radiation image, and the processing unit inputs the plurality of first partial images instead of the radiation image into a trained model and performs image processing to remove noise from the plurality of first partial images, and integrates the plurality of first partial images from which noise has been removed to generate a radiation image from which noise has been removed."

[0122] The radiation image processing module of the embodiment may be

[15] "a radiation image processing module according to any one of

[12] to

[14] above, wherein the spatial blur map generation unit derives blur evaluation information from the pixel value of each pixel of the radiation image based on second relationship data indicating the relationship between the pixel value and blur evaluation information that evaluates the spatial blur of noise, and generates data in which the derived blur evaluation information is associated with each pixel of the radiation image as a spatial blur map."

[0123] The radiation image processing module of the embodiment is

[16] "a radiation image processing module according to any one of the above

[12] to

[15] , further comprising: a training data generation unit that cuts out a plurality of second partial images from an entire image as training images, sets blur evaluation information that evaluates the spatial blur of noise as blur setting information for each second partial image, generates a spatial blur map that indicates the distribution of the spatial blur of noise for the plurality of second partial images based on the blur setting information, and generates a noise image by adding noise to the second partial image based on the generated spatial blur map; and a construction unit that uses the second partial images, the spatial blur map, and the noise image as training data, and constructs by machine learning a trained model for inputting the spatial blur map and the noise image and removing noise from the noise image, and outputting the noise-removed image so that the noise-removed image approximates the second partial image."

[0124] The radiation image processing module of the embodiment may be

[17] "a radiation image processing module described in

[16] above, in which the training data generation unit sets the blur setting information as blur evaluation information that differs from each other for each second partial image."

[0125] The radiation image processing module of the embodiment may be

[18] "a radiation image processing module described in either

[16] or

[17] above, wherein the training data generation unit sets blur evaluation information derived from the pixel values ​​of each pixel of the second partial image as blur setting information for each pixel of the second partial image based on second relationship data indicating the relationship between pixel values ​​and blur evaluation information that evaluates the spatial blur of noise, and generates data set by associating the blur setting information with each pixel of the second partial image as a spatial blur map."

[0126] The radiation image processing module of the embodiment may be

[19] "the radiation image processing module according to the above

[18] , in which the training data generation unit generates the second relational data by simulation."

[0127] The radiation image processing module of the embodiment may be

[20] "a radiation image processing module described in

[18] above, in which the training data generation unit generates second relational data based on radiation images obtained by actually capturing the radiation image."

[0128] The radiation image processing module of the embodiment may be

[21] "a radiation image processing module according to any one of

[16] to

[20] above, in which the blur evaluation information is at least one of a sigma value, a half-width, a point spread function, a contrast transfer function, and a modulation transfer function."

[0129] The radiation image processing program of the embodiment may be

[22] "a radiation image processing program that causes a processor to function as an image acquisition unit that acquires a radiation image in which radiation is irradiated onto an object and the radiation that has passed through the object is captured, a spatial blur map generation unit that generates a spatial blur map that indicates the distribution of spatial blur of noise based on the radiation image, and a processing unit that inputs the radiation image and the spatial blur map into a trained model that has been constructed in advance by machine learning, and performs image processing to remove noise from the radiation image."

[0130] The radiation image processing system of the embodiment may be

[23] "a radiation image processing system comprising the radiation image processing module described in

[12] to

[21] above, a radiation source that irradiates an object with radiation, and an imaging device that captures the radiation that has passed through the object to obtain a radiation image."

[0131] 10...X-ray detection camera (imaging device), 20...control device (radiation image processing module), 50...X-ray irradiator (source), 201...image acquisition unit, 202...noise map generation unit, 203...spatial blur map generation unit, 204...processing unit, 205...construction unit, 206...training data generation unit, 207...trained model, F...object, G1, G10...image (radiation image), G2...partial image (first partial image), G12...partial image (second partial image), G4...relationship graph (first relationship data), G6, G13...noise map, G70...second relationship data, G8, G14...spatial blur map, G9...noise-removed image, G11...entire image, G16...noise image.

Claims

1. an image acquisition step of irradiating an object with an energy beam and acquiring an image of the energy beam that has passed through the object; a spatial blur map generation step of generating a spatial blur map indicating a distribution of spatial blur of noise based on the image; a processing step of inputting the image and the spatial blur map into a trained model previously constructed by machine learning, and performing image processing to remove noise from the image; An image processing method comprising:

2. a noise map generating step of deriving a noise evaluation value from the pixel value of each pixel of the image based on first relationship data indicating a relationship between the pixel value and a noise evaluation value obtained by evaluating a spread of noise values, and generating a noise map as data associating the derived noise evaluation value with each pixel of the image; In the processing step, the image, the spatial blur map, and the noise map are input to the trained model, and image processing is performed to remove noise from the image. The image processing method according to claim 1 .

3. In the spatial blur map generating step, the spatial blur map is generated for each of a plurality of first partial images obtained by cutting out the image; In the processing step, the plurality of first partial images are input to the trained model instead of the image, and image processing is performed to remove noise from the plurality of first partial images, and the plurality of first partial images from which noise has been removed are integrated to generate the image from which noise has been removed.

3. The image processing method according to claim 1.

4. In the spatial blur map generating step, the blur evaluation information is derived from pixel values ​​of each pixel of the image based on second relationship data indicating a relationship between pixel values ​​and blur evaluation information that evaluates spatial blur of noise, and data in which the derived blur evaluation information is associated with each pixel of the image is generated as the spatial blur map.

3. The image processing method according to claim 1.

5. a training data generation step of extracting a plurality of second partial images from the entire image as training images, setting blur evaluation information that evaluates the spatial blur of noise as blur setting information for each of the second partial images, generating a spatial blur map that indicates the distribution of the spatial blur of noise for the plurality of second partial images based on the blur setting information, and generating a noise image by adding noise to the second partial images based on the generated spatial blur map; a construction step of constructing a trained model by machine learning using the second partial image, the spatial blur map, and the noise image as training data, and using the spatial blur map and the noise image as inputs to output a denoised image obtained by removing noise from the noise image so that the denoised image approximates the second partial image; A training method that includes:

6. In the training data generating step, the blur setting information is set as the blur evaluation information that differs from one another for each of the second partial images. The training method according to claim 5.

7. In the training data generation step, the blur evaluation information derived from the pixel values ​​of the pixels of the second partial image is set as the blur setting information for each pixel of the second partial image based on second relationship data indicating a relationship between pixel values ​​and blur evaluation information that evaluates spatial blur of noise, and data set by associating the blur setting information with each pixel of the second partial image is generated as the spatial blur map. The training method according to claim 5.

8. In the training data generating step, the second relational data is generated by simulation. The training method according to claim 7.

9. In the training data generating step, the second relational data is generated based on an image obtained by actually capturing an image. The training method according to claim 7.

10. In the training data generation step, the blur evaluation information is at least one of a sigma value, a half width, a point spread function, a contrast transfer function, and a modulation transfer function. The training method according to claim 5.

11. A trained model constructed by the training method according to any one of claims 5 to 10, A trained model that causes a processor to perform image processing to remove noise from images of energy rays passing through an object.

12. an image acquisition unit that acquires a radiological image obtained by irradiating an object with radiation and capturing the radiation that has passed through the object; a spatial blur map generating unit that generates a spatial blur map indicating a distribution of spatial blur of noise based on the radiation image; a processing unit that inputs the radiographic image and the spatial blur map into a trained model that has been constructed in advance by machine learning, and executes image processing to remove noise from the radiographic image; A radiation image processing module comprising:

13. a noise map generator that derives a noise evaluation value from the pixel values ​​of each pixel of the radiographic image based on first relationship data indicating the relationship between pixel values ​​and a noise evaluation value that evaluates the spread of noise values, and generates a noise map that is data that associates the derived noise evaluation value with each pixel of the radiographic image; the processing unit inputs the radiographic image, the spatial blur map, and the noise map into the trained model, and performs image processing to remove noise from the radiographic image. The radiation image processing module according to claim 12 .

14. the spatial blur map generation unit generates the spatial blur map for each of a plurality of first partial images extracted from the radiographic image; the processing unit inputs the plurality of first partial images into the trained model instead of the radiographic image, performs image processing to remove noise from the plurality of first partial images, and integrates the plurality of first partial images from which noise has been removed to generate the radiographic image from which noise has been removed. The radiation image processing module according to claim 12 or 13.

15. the spatial blur map generation unit derives the blur evaluation information from the pixel values ​​of each pixel of the radiographic image based on second relationship data indicating a relationship between pixel values ​​and blur evaluation information that evaluates spatial blur of noise, and generates data in which the derived blur evaluation information is associated with each pixel of the radiographic image as the spatial blur map. The radiation image processing module according to claim 12 or 13.

16. a training data generation unit that cuts out a plurality of second partial images from the entire image as training images, sets blur evaluation information that evaluates spatial blur of noise as blur setting information for each of the second partial images, generates a spatial blur map that indicates the distribution of spatial blur of noise for the plurality of second partial images based on the blur setting information, and generates a noise image by adding noise to the second partial images based on the generated spatial blur map; a construction unit that constructs, by machine learning, a trained model using the second partial image, the spatial blur map, and the noise image as training data, and inputs the spatial blur map and the noise image to output a denoised image obtained by removing noise from the noise image so that the denoised image approximates the second partial image. The radiation image processing module according to claim 12 or 13.

17. the training data generation unit sets the blur setting information as the blur evaluation information that differs for each of the second partial images. The radiation image processing module according to claim 16.

18. the training data generation unit sets the blur evaluation information derived from the pixel value of each pixel of the second partial image as the blur setting information for each pixel of the second partial image based on second relationship data indicating a relationship between pixel values ​​and blur evaluation information that evaluates spatial blur of noise, and generates data that associates the blur setting information with each pixel of the second partial image and sets the blur evaluation information as the spatial blur map. The radiation image processing module according to claim 16.

19. the training data generation unit generates the second relational data through a simulation.

19. The radiation image processing module according to claim 18.

20. the training data generation unit generates the second relational data based on radiographic images obtained by actually capturing the radiographic images.

19. The radiation image processing module according to claim 18.

21. The blur evaluation information is at least one of a sigma value, a half-width, a point spread function, a contrast transfer function, and a modulation transfer function. The radiation image processing module according to claim 16.

22. The processor, an image acquisition unit that acquires a radiological image obtained by irradiating an object with radiation and capturing the radiation that has passed through the object; a spatial blur map generator that generates a spatial blur map indicating a distribution of spatial blur of noise based on the radiation image; and a radiological image processing program that inputs the radiological image and the spatial blur map into a trained model that has been constructed in advance by machine learning, and causes the program to function as a processing unit that executes image processing to remove noise from the radiological image.

23. a radiation image processing module according to claim 12 or 13; a source for irradiating the object with radiation; an imaging device that captures an image of radiation that has passed through the object to obtain the radiation image; A radiation image processing system comprising: