Radiation image processing method, machine learning method, trained model, preprocessing method for machine learning, radiation image processing module, radiation image processing program, and radiation image processing system

Abstract

According to the present invention, a control device 20 comprises: an image acquisition unit 203 which acquires a radiographic image obtained by irradiating a subject F with radiation and image-capturing the radiation that has penetrated through the subject F; a noise map generation unit 204 which derives an evaluation value from a pixel value of each pixel of the radiographic image on the basis of relationship data that indicates the relationship between the pixel values and the evaluation values obtained by evaluating the spread of noise values, and generates a noise map that is data obtained by associating each pixel of the radiographic image with the derived evaluation value; and a processing unit 205 which inputs the radiographic image and the noise map to a trained model 207 pre-constructed by machine learning, and executes image processing for removing noise from the radiographic image.

Description

Technical Field

[0001] One aspect of the embodiments relates to a radiation image processing method, a machine learning method, a learned model, a preprocessing method for machine learning, a radiation image processing module, a radiation image processing program, and a radiation image processing system.

Background Art

[0002] Conventionally, a method of performing noise removal using a learned model by machine learning on image data has been known (see, for example, Patent Document 1 below). According to this method, noise from the image data is automatically removed, so that the object can be observed accurately.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the conventional method as described above, when a radiation image generated by transmitting radiation such as X-rays through an object is targeted, noise removal may not be sufficient. For example, due to conditions such as the conditions of a radiation source such as an X-ray source and the type of filter used, the relationship between the pixel values and noise in the image tends to vary, and noise tends not to be effectively removed.

[0005] Therefore, one aspect of the embodiments has been made in view of such problems, and an object thereof is to provide a radiation image processing method, a machine learning method, a learned model, a preprocessing method for machine learning, a radiation image processing module, a radiation image processing program, and a radiation image processing system that can effectively remove noise in a radiation image.

Means for Solving the Problems

[0006] A radiation image processing method according to an aspect of an embodiment includes an image acquisition step of acquiring a radiation image in which an object is irradiated with radiation and the radiation transmitted through the object is imaged, a noise map generation step of deriving an evaluation value from the pixel value of each pixel of the radiation image based on relationship data representing the relationship between the evaluation value obtained by evaluating the spread of the pixel value and the noise value, and generating a noise map which is data associating the derived evaluation value with each pixel of the radiation image, and a processing step of inputting the radiation image and the noise map into a pre-constructed learned model by machine learning and performing image processing for removing noise from the radiation 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 in which an object is irradiated with radiation and the radiation transmitted through the object is imaged, a noise map generation unit that derives an evaluation value from the pixel value of each pixel of the radiation image based on relationship data representing the relationship between the evaluation value obtained by evaluating the spread of the pixel value and the noise value, and generates a noise map which is data associating the derived evaluation value with each pixel of the radiation image, and a processing unit that inputs the radiation image and the noise map into a pre-constructed learned model by machine learning and performs image processing for removing noise from the radiation image.

[0008] Alternatively, a radiation image processing program according to another aspect of the embodiment causes a processor to function as an image acquisition unit that acquires a radiation image in which an object is irradiated with radiation and the radiation transmitted through the object is imaged, a noise map generation unit that derives an evaluation value from the pixel value of each pixel of the radiation image based on relationship data representing the relationship between the evaluation value obtained by evaluating the spread of the pixel value and the noise value, and generates a noise map which is data associating the derived evaluation value with each pixel of the radiation image, and a processing unit that inputs the radiation image and the noise map into a pre-constructed learned model by machine learning and performs image processing for removing noise from the radiation image.

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

[0010] According to either the one aspect or the other aspect described above, an evaluation value is derived from the pixel value of each image of the radiation image based on the relationship data representing the relationship between the evaluation value obtained by evaluating the spread of the pixel value and the noise value, and a noise map is generated, which is data in which the derived evaluation value is associated with each pixel of the radiation image. Then, the radiation image and the noise map are input into a pre-constructed learned model by machine learning, and image processing for removing noise from the radiation image is executed. According to such a configuration, the spread of the noise value evaluated from the pixel value of each pixel of the radiation image is considered, and the noise at each pixel of the radiation image is removed by machine learning. Thereby, noise removal corresponding to the relationship between the pixel value and the spread of noise in the radiation image can be realized using the learned model. As a result, the noise in the radiation image can be effectively removed.

Advantages of the Invention

[0011] According to one aspect of the present disclosure, noise in a radiation image of a subject can be effectively removed.

Brief Description of the Drawings

[0012]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

Figure 18

Figure 19

Figure 20

Figure 21

Figure 22

Figure 23

Figure 24

Figure 25

Figure 26

Figure 27

Figure 28

Figure 29

Figure 30

Figure 31

Figure 32

Figure 33

Embodiments for Carrying Out the Invention

[0013] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the description, the same reference numerals will be used for the same elements or elements having the same function, and redundant descriptions will be omitted. [First Embodiment]

[0014] FIG. 1 is a configuration diagram of an image acquisition device 1 which is a radiation image processing system according to the first embodiment. As shown in FIG. 1, the image acquisition device 1 is a device that irradiates an object F conveyed in the conveyance direction TD with X-rays (radiation), and acquires an X-ray image (radiation image) obtained by imaging the object F based on the X-rays transmitted through the object F. The image acquisition device 1 uses the X-ray image to perform foreign object inspection, weight inspection, product inspection, etc. on the object F, and examples of its applications include food inspection, baggage inspection, substrate inspection, battery inspection, material inspection, etc. The image acquisition device 1 includes a belt conveyor (conveyance means) 60, an X-ray irradiator (radiation 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 performing various inputs. Note that the radiation image in the embodiments of the present disclosure includes not only X-ray images but also images obtained by radiation other than X-rays such as γ-rays.

[0015] The belt conveyor 60 has a belt portion on which the object F is placed, and by moving the belt portion in the transport direction TD, the object F is transported in the transport direction TD at a predetermined transport speed. The transport speed of the object F is, for example, 48 m / min. The belt conveyor 60 can change the transport speed to, for example, a transport speed of 24 m / min, 96 m / min, etc., as necessary. Also, the belt conveyor 60 can appropriately change the height position of the belt portion and change the distance between the X-ray irradiator 50 and the object F. Note that examples of the object F transported by the belt conveyor 60 include various articles such as foods like meat, seafood, agricultural products, confectionery, rubber products like tires, resin products, metal products, resource materials like minerals, waste, and electronic components and electronic substrates. The X-ray irradiator 50 is a device that irradiates (outputs) X-rays as an X-ray source to the object F. The X-ray irradiator 50 is a point light source and irradiates X-rays while diffusing them within a predetermined angular range in a certain irradiation direction. The X-ray irradiator 50 is arranged above the belt conveyor 60 at a predetermined distance from the belt conveyor 60 such that the irradiation direction of the X-rays is directed toward the belt conveyor 60 and the diffused X-rays cover the entire width direction (the direction intersecting the transport direction TD) of the object F. Also, the irradiation range of the X-ray irradiator 50 in the length direction (the direction parallel to the transport direction TD) of the object F is a predetermined divided range in the length direction, and as the object F is transported in the transport direction TD by the belt conveyor 60, the 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 X-rays with a predetermined energy and radiation dose corresponding to the set tube voltage and tube current toward the belt conveyor 60. Also, a filter 51 that transmits a predetermined wavelength range of X-rays is provided in the vicinity of the X-ray irradiator 50 on the side of the belt conveyor 60. The filter 51 is not necessarily required and may be omitted as appropriate.

[0016] The X-ray detection camera 10 detects the X-rays that have passed through the object F among the X-rays irradiated on 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 configurations for detecting X-rays are arranged. In the image acquisition device 1 according to the first embodiment, X-ray images are respectively generated based on the X-rays detected by each line (the first line and the second line) of the dual-line X-ray camera. Then, for the two generated X-ray images, by performing average processing, addition processing, etc., compared with the case of generating an X-ray image based on the X-rays detected by one line, an image that is clear (has a large luminance) can be obtained with a smaller X-ray dose.

[0017] 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, the line scan camera 12a, the amplifier 14a, the AD converter 15a, the correction circuit 16a, and the output interface 17a are electrically connected to each other and are the configuration related to the first line. Also, the scintillator 11b, the line scan camera 12b, the amplifier 14b, the AD converter 15b, the correction circuit 16b, and the output interface 17b are electrically connected to each other and are the configuration related to the 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 conveyance direction TD. Hereinafter, for the configurations common to the first line and the second line, the configuration of the first line will be described representatively.

[0018] The scintillator 11a is fixed onto the line scan camera 12a by adhesion or the like, and converts X-rays transmitted through the object F into scintillation light. The scintillator 11a outputs the scintillation light to the line scan camera 12a. The filter 19 transmits a predetermined wavelength range of X-rays toward the scintillator 11a. The filter 19 is not necessarily required and may be omitted as appropriate.

[0019] The line scan camera 12a detects the scintillation light from the scintillator 11a, converts it into charges, and outputs it 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 conveyance direction TD. The line sensor is, for example, a CCD (Charge Coupled Device) image sensor, a CMOS (Complementary Metal-Oxide Semiconductor) image sensor, or the like, and includes a plurality of photodiodes.

[0020] 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 X-rays transmitted through the same region of the object F. The predetermined detection period may be set to a common period for 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)). Also, the predetermined period may be set individually based on the pixel width of the photodiodes in the direction orthogonal to the pixel array direction of the line sensors of the line scan cameras 12a and 12b, respectively. In this case, the detection period deviation (delay time) between the line scan cameras 12a and 12b is specified according to 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), and the distance between the X-ray irradiator 50 and the line scan cameras 12a and 12b (FDD), and individual periods may be set. 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 the amplification factor 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.

[0021] 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 outside the X-ray detection camera 10. In FIG. 1, the AD converter, the correction circuit, and the output interface exist individually, but they may also be integrated into one.

[0022] The control device 20 is a computer such as a PC (Personal Computer). The control device 20 generates an X-ray image based on the digital signals (amplified signals) output from the X-ray detection cameras 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 output to the display device 30 after being subjected to noise removal processing described later, and is displayed by the display device 30. Further, the control device 20 controls the X-ray irradiator 50, the amplifier control unit 18, and the sensor control unit 13. Note that the control device 20 of the first embodiment is a device provided independently outside the X-ray detection camera 10, but may be integrated inside the X-ray detection camera 10.

[0023] Figure 2 shows the hardware configuration of the control device 20. As shown in Figure 2, physically, the control device 20 is a computer including a CPU (Central Processing Unit) 101 and a GPU 105 (Graphic Processing Unit) which are processors, a RAM (Random Access Memory) 102 and a ROM (Read Only Memory) 103 which are recording media, a communication module 104, and an input / output module 106, etc., and each is electrically connected. Note that the control device 20 may include a display, a keyboard, a mouse, a touch panel display, etc. as the input device 40 and the display device 30, or may include a data recording device such as a hard disk drive or a semiconductor memory. Further, the control device 20 may be composed of a plurality of computers.

[0024] FIG. 3 is a block diagram showing the functional configuration of the control device 20. The control device 20 includes an input unit 201, a calculation unit 202, an image acquisition unit 203, a noise map generation unit 204, a processing unit 205, and a construction unit 206. Each functional unit of the control device 20 shown in FIG. 3 is realized by loading a program (the radiation image processing program of the first embodiment) onto hardware such as the CPU 101, GPU 105, and RAM 102, and operating the communication module 104, the input / output module 106, etc. under the control of the CPU 101 and GPU 105, and reading and writing data in the RAM 102. The CPU 101 and GPU 105 of the control device 20 function as each functional unit in FIG. 3 by executing this computer program, and sequentially execute processes corresponding to the radiation image processing method described later. Note that the CPU 101 and GPU 105 may be a single piece of hardware, or only one of them may be used. Also, the CPU 101 and GPU 105 may be implemented in programmable logic such as an FPGA like a soft processor. The RAM and ROM may also be a single piece of hardware, or may be built into programmable logic such as an FPGA. All various data necessary for the execution of this computer program, and various data generated by the execution of this computer program are stored in a built-in memory such as the ROM 103 and RAM 102, or a storage medium such as a hard disk drive. Also, a learned model 207 (described later) for causing the CPU 101 and GPU 105 to execute noise removal processing on an X-ray image is stored in advance in the built-in memory or storage medium in the control device 20 so that it can be read by the CPU 101 and GPU 105.

[0025] Hereinafter, the details of the functions of each functional unit of the control device 20 will be described.

[0026] The input unit 201 receives an input of condition information indicating either the conditions of the radiation source or the imaging conditions when irradiating the object F with radiation to image it. Specifically, the input unit 201 receives from the user of the image acquisition device 1 an input of condition information indicating the operating conditions of the X-ray irradiator (radiation source) 50 when imaging the X-ray image of the object F, or the imaging conditions by the X-ray detection camera 10, etc. Examples of the operating conditions include all or part of the tube voltage, target angle, target material, etc. Examples of the condition information indicating the imaging conditions include the material and thickness of the filters 51 and 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 of window material of the X-ray detection camera 10, and information regarding the material and thickness of the scintillators 11a and 11b of the X-ray detection camera 10, X-ray detection camera information (e.g., gain setting value, circuit noise value, saturation charge amount, conversion coefficient value (e- / count), line rate (Hz) or line speed (m / min) of the camera), information of the object F, etc., all or part of which are included. The input unit 201 may receive the input of the condition information as a direct input of information such as numerical values, or as a selection input for information such as numerical values set in the internal memory in advance. The input unit 201 receives the above-mentioned input of the condition information from the user, but may also acquire some of the condition information (such as tube voltage, etc.) according to the detection result of the control state by the control device 20.

[0027] The calculation unit 202 calculates the average energy of the X-rays (radiation) that have passed through the object F based on the condition information. The condition information includes at least any one of the tube voltage of the radiation source, information about the object F, information about the filter provided in the camera used for imaging the object F, information about the scintillator provided in the camera, and information about the filter provided in the X-ray radiation source. Specifically, the calculation unit 202 calculates the value of the average energy of the X-rays that pass through the object F using the image acquisition device 1 and are detected by the X-ray detection camera 10 based on the condition information received by the input unit 201. For example, the calculation unit 202 calculates the spectrum of the X-rays detected by the X-ray detection camera 10 based on information such as the tube voltage, target angle, target material, material and thickness and presence or absence of the filters 51, 19, type and presence or absence of the window material of the X-ray detection camera 10, and material and thickness of the scintillators 11a, 11b of the X-ray detection camera 10, using, for example, an approximation formula such as the well-known Tucker. Then, the calculation unit 202 further calculates the spectral intensity integral value and the photon number integral value from the X-ray spectrum, and calculates the value of the average energy of the X-rays by dividing the spectral intensity integral value by the photon number integral value.

[0028] A calculation method using the well-known Tucker approximation formula will be described. For example, when the calculation unit 202 specifies the target as tungsten and the target angle as 25°, it can determine Em: the kinetic energy at the time of electron-target collision, T: the electron kinetic energy in the target, A: the proportionality constant determined by the atomic number of the target material, ρ: the density of the target, μ(E): the linear attenuation coefficient of the target material, B: a function of Z and T that changes gently, C: the Thomson-Whiddington constant, θ: the target angle, and c: the speed of light in a vacuum. Further, the calculation unit 202 can calculate the irradiated X-ray spectrum by calculating the following formula (1) based on these.

[0029]

Equation

[0030] Next, the calculation unit 202 can calculate the X-ray energy spectrum that passes through the filter and the object F and is absorbed by the scintillator using the X-ray attenuation formula of the following formula (2).

[0031]

Equation

[0032] The image acquisition unit 203 irradiates the object F with radiation and acquires a radiation image obtained by imaging the radiation that has passed through the object F. Specifically, the image acquisition unit 203 generates an X-ray image based on the digital signal (amplified signal) output from the X-ray detection camera 10 (more specifically, the output interfaces 17a, 17b). The image acquisition unit 203 generates one X-ray image by averaging or adding the two digital signals output from the output interfaces 17a, 17b. FIG. 4 is a diagram showing an example of the X-ray image acquired by the image acquisition unit 203.

[0033] The noise map generation unit 204 derives an evaluation value from the pixel value of each pixel of the radiation image based on the relational data representing the relationship between the pixel value and the evaluation value obtained by evaluating the spread of the noise value, and generates a noise map which is data associating the derived evaluation value with each pixel of the radiation image. At this time, the noise map generation unit 204 derives the evaluation value from the average energy of the radiation that has passed through the object F and the pixel value of each pixel of the radiation image. Specifically, the noise map generation unit 204 uses the relational expression (relational data) between the pixel value and the standard deviation of the noise value (the evaluation value obtained by evaluating the spread of the noise value) to derive the standard deviation of the noise value from the average energy of the X-rays calculated by the calculation unit 202 and the pixel value of each pixel of the X-ray image (radiation image) acquired by the image acquisition unit 203. The noise map generation unit 204 generates a noise standard deviation map (noise map) by associating the derived standard deviation of the noise value with each pixel of the X-ray image.

[0034] The relational expression between the pixel value and the average energy used by the noise map generation unit 204 and the standard deviation of the noise value is represented by the following formula (4).

[0035]

Equation

[0036] FIG. 5 is a diagram showing an example of generating a noise standard deviation map by the noise map generation unit 204. The noise map generation unit 204 uses the relational expression (4) between the pixel value and the standard deviation of the noise value, substitutes various pixel values into the variable Signal, and obtains the correspondence between the pixel value and the variable Noise, thereby deriving a relational graph G3 representing the correspondence between the pixel value and the standard deviation of the noise value. Then, the noise map generation unit 204 derives relational data G2 representing the correspondence between each pixel position and the pixel value from the X-ray image G1 acquired by the image acquisition unit 203. Further, the noise map generation unit 204 applies the correspondence shown in the relational graph G3 to each pixel value in the relational data G2, thereby deriving the standard deviation of the noise value corresponding to the pixel at each pixel position in the X-ray image. As a result, the noise map generation unit 204 associates the derived standard deviation of the noise with each pixel position, and derives relational data G4 showing the correspondence between each pixel position and the standard deviation of the noise. Then, the noise map generation unit 204 generates a noise standard deviation map G5 based on the derived relational data G4.

[0037] The processing unit 205 inputs the radiation image and the noise map into a learned model 207 constructed in advance by machine learning, and executes image processing for removing noise from the radiation image. That is, as shown in FIG. 6, the processing unit 205 acquires the learned model 207 (described later) constructed by the construction unit 206 from the built-in memory or the storage medium in the control device 20. The processing unit 205 inputs the X-ray image G1 acquired by the image acquisition unit 203 and the noise standard deviation map G5 generated by the noise map generation unit 204 into the learned model 207. Thereby, the processing unit 205 generates an output image G6 by executing image processing for removing noise from the X-ray image G1 using the learned model 207. Then, the processing unit 205 outputs the generated output image G6 to the display device 30 or the like.

[0038] The construction unit 206 constructs, by machine learning, a learned model 207 that outputs noise-removed image data based on a training image that is a radiation image, a noise map generated from the training image based on a relational expression between a pixel value and a standard deviation of a noise value, and noise-removed image data that is data with noise removed from the training image, using the training image and the noise map as training data. The construction unit 206 stores the constructed learned model 207 in a built-in memory or a storage medium within the control device 20. Machine learning includes supervised learning, unsupervised learning, and reinforcement learning, and among these learnings, there are deep learning (deep neural network learning), neural network learning, and the like. In the first embodiment, as an example of a deep learning algorithm, 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 adopted. Note that the learned model 207 may be generated by an external computer or the like and downloaded to the control device 20, in addition to being constructed by the construction unit 206. Note that the radiation image used for machine learning includes a radiation image obtained by imaging a known structure or an image reproduced from the radiation image.

[0039] FIG. 7 is an example of a training image that is one of the training data used for constructing the learned model 207. As the training image, X-ray images of patterns with various thicknesses, various materials, and various resolutions can be used as imaging targets. The example shown in FIG. 7 is a training image G7 generated for chicken. The training image G7 may be an X-ray image actually generated for a plurality of types of known structures using the image acquisition device 1, or an image generated by simulation calculation. The X-ray image may be obtained using a device different from the image acquisition device 1.

[0040] The construction unit 206 derives an evaluation value from the pixel value of each pixel of the radiation image based on relationship data representing the relationship between the evaluation value obtained by evaluating the spread of the pixel value and the noise value, and generates a noise map which is data associating the derived evaluation value with each pixel of the radiation image, as preprocessing for performing machine learning. Specifically, when constructing the learned model 207, the construction unit 206 acquires a training image generated by actual imaging, simulation calculation, or the like from the image acquisition unit 203 or the like. Then, the construction unit 206 sets, for example, the operating conditions of the X-ray irradiator 50 of the image acquisition apparatus 1 or the imaging conditions of the image acquisition apparatus 1. Alternatively, the construction unit 206 sets the operating conditions or imaging conditions of the X-ray irradiator 50 during the simulation calculation. The construction unit 206 calculates the average energy of the X-ray based on the above-described operating conditions or imaging conditions using the same method as the calculation unit 202. Further, the construction unit 206 generates a noise standard deviation map based on the average energy of the X-ray and the training image using the same method as the method by the noise map generation unit 204 as shown in FIG. 5. That is, the preprocessing method of the machine learning method includes a noise map generation step of generating a noise map which is data associating the evaluation value derived from the pixel value of each pixel of the radiation image with the evaluation value based on the relationship data representing the relationship between the evaluation value obtained by evaluating the spread of the pixel value and the noise value.

[0041] The construction unit 206 constructs a learned model 207 by machine learning using a training image, a noise map generated from the training image, and noise-removed image data which is data obtained by removing noise from the training image in advance as training data. Specifically, the construction unit 206 acquires in advance noise-removed image data obtained by removing noise from the training image. When the training image is an X-ray image generated by simulation calculation, the construction unit 206 uses, as the noise-removed image data, an image before noise is added in the process of generating the training image. On the other hand, when the training image is an X-ray image actually generated by the image acquisition device 1 for a plurality of known structures, the construction unit 206 uses, as the noise-removed image data, an image obtained by removing noise from the X-ray image using image processing such as an average value filter, a median filter, a bilateral filter, or an NLM filter. The construction unit 206 constructs a learned model 207 that outputs noise-removed image data based on the training image and the noise standard deviation map by performing training by machine learning.

[0042] Next, a procedure for observing an X-ray transmission image of the object F using the image acquisition device 1 according to the first embodiment, that is, a flow of the radiation image processing method according to the first embodiment will be described. FIG. 8 is a flowchart showing the procedure of the observation process by the image acquisition device 1.

[0043] First, the construction unit 206 constructs, by machine learning, a learned model 207 that outputs noise-removed image data based on a training image and a noise standard deviation map using, as training data, the training image, the noise standard deviation map generated from the training image based on a relational expression, and the noise-removed image data (step S100). Next, the input unit 201 receives an input of condition information indicating operation conditions of the X-ray irradiator 50 or imaging conditions by the X-ray detection camera 10, etc. from an operator (user) of the image acquisition device 1 (step S101). Then, the calculation unit 202 calculates a value of the average energy of the X-rays detected by the X-ray detection camera 10 based on the condition information (step S102).

[0044] Subsequently, in the image acquisition device 1, the object F is set and the object F is imaged, and an X-ray image of the object F is acquired by the control device 20 (step S103). Further, based on the relational expression between the pixel value and the standard deviation of the noise value by the control device 20, the standard deviation of the noise value is derived from the average energy of the X-ray and the pixel value of each pixel of the X-ray image, and the derived standard deviation of the noise is associated with each pixel value, whereby a noise standard deviation map is generated (step S104).

[0045] Next, the X-ray image of the object F and the noise standard deviation map are input to the pre-constructed and stored learned model 207 by the processing unit 205, and noise removal processing is executed on the X-ray image (step S105). Further, an output image, which is the X-ray image subjected to the noise removal processing, is output to the display device 30 by the processing unit 205. (Step S106).

[0046] According to the image acquisition device 1 described above, using the relational expression between the pixel value and the standard deviation of the noise value, the standard deviation of the noise value is derived from the pixel value of each image of the X-ray image, and a noise standard deviation map, which is data associating the derived standard deviation of the noise value with each pixel of the X-ray image, is generated. Then, the X-ray image and the noise standard deviation map are input to the learned model 207 pre-constructed by machine learning, and image processing for removing noise from the X-ray image is executed. According to such a configuration, the standard deviation of the noise value derived from the pixel value of each pixel of the X-ray image is considered, and the noise at each pixel of the X-ray image is removed by machine learning. Thereby, using the learned model 207, noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise value in the X-ray image can be realized. As a result, the noise in the X-ray image can be effectively removed.

[0047] In particular, in X-ray images, the noise pattern changes depending on differences in tube voltage, filter, scintillator, conditions of the X-ray detection camera (gain setting value, circuit noise value, saturation charge amount, conversion coefficient value (e− / count), line rate of the camera), the object, and the like. Therefore, when attempting to achieve noise removal by machine learning, it is conceivable to prepare trained models trained under various conditions. That is, as a comparative example, a method of constructing a plurality of trained models according to the conditions at the time of measuring an X-ray image, selecting a trained model for each condition, and executing noise removal processing can also be adopted. In such a comparative example, for example, trained models must be constructed for each noise condition such as the average energy of X-rays, the gain of the X-ray detection camera, and the type of X-ray camera, and it is necessary to generate a huge number of trained models, which may require a lot of time for construction. As an example, when there are 10 types of average energies of X-rays, 8 types of gains of X-ray detection cameras, and 3 types of product types, 240 trained models are required. If it takes one day to construct one trained model, it will take 240 days for machine learning. In this regard, according to the present embodiment, by generating a noise map from an X-ray image and using the noise map as input data for machine learning, the noise conditions that require the generation of a trained model can be reduced, and the learning time for constructing the trained model 207 is greatly reduced.

[0048] FIGS. 9 and 10 each show an example of an X-ray image before and after noise removal processing acquired by the image acquisition device 1. FIG. 9 is an X-ray image acquired under the same conditions as the training data used for constructing the trained model 207 for the operating conditions of the X-ray irradiator 50 and the imaging conditions by the X-ray detection camera 10. The X-ray image G8 is a measured image, the X-ray image G9 is an image subjected to noise removal processing using a trained model learned under the same conditions as the imaging conditions of the comparative example, and the X-ray image G10 is an image subjected to noise removal processing using the noise standard deviation map by the control device 20 in the first embodiment. The standard deviations of the noise values in the X-ray images G8, G9, and G10 are 14.3, 3.4, and 3.7, respectively.

[0049] FIG. 10 shows X-ray images obtained when the operating conditions of the X-ray irradiator 50 or the imaging conditions by the X-ray detection camera 10 are different from the training data used for constructing the learned model 207. The X-ray image G11 is an actually measured image, the X-ray image G12 is an image subjected to noise removal processing using a learned model trained under conditions different from the imaging conditions of the comparative example, and the X-ray image G13 is the X-ray image subjected to noise removal processing using the noise standard deviation map by the control device 20 in the first embodiment. The standard deviations of the noise values in the X-ray images G11, G12, and G13 are 3.5, 2.0, and 0.9, respectively.

[0050] In the comparative example, as shown in FIG. 9, when the operating conditions of the X-ray irradiator 50 or the imaging conditions by the X-ray detection camera 10 are the same as the training data used for constructing the learned model 207, in the X-ray image G9, the standard deviation of the noise value is sufficiently reduced as compared with the X-ray image G8 before the noise removal processing. The learned model in the comparative example can output an X-ray image with sufficiently removed noise. However, as shown in FIG. 10, when the operating conditions of the X-ray irradiator 50 or the imaging conditions by the X-ray detection camera 10 are different from the training data used for constructing the learned model 207, in the X-ray image G12 after the noise processing, it is not sufficiently reduced as compared with the X-ray image G11 before the noise removal processing. Therefore, the learned model in the comparative example cannot output an X-ray image with sufficiently removed noise when the conditions are different between the training time and the imaging time.

[0051] In contrast, according to the first embodiment, the learned model 207 is constructed in consideration of changes in the operating conditions of the X-ray irradiator 50 or the imaging conditions by the X-ray detection camera 10 during measurement of the X-ray image. As a result, as shown in FIGS. 9 and 10, in the X-ray images G10 and 13, the standard deviation of the noise values is sufficiently reduced as compared with the X-ray images G8 and 11 before the respective noise removal processes. Therefore, according to the first embodiment, sufficient noise removal corresponding to changes in the operating conditions of the X-ray irradiator 50 or the imaging conditions by the X-ray detection camera 10 is realized. Thereby, noise in the X-ray image can be effectively removed using a single learned model 207.

[0052] Generally, X-ray images contain noise resulting from X-ray generation. Although it is conceivable to increase the X-ray dose to improve the signal-to-noise ratio (SN ratio) of the X-ray image, in that case, there are problems such as an increase in the radiation dose to the sensor and a shortening of the sensor's lifespan, as well as a shortening of the lifespan of the X-ray source, making it difficult to achieve both an improvement in the SN ratio and a longer lifespan. In addition, since the amount of heat generated also increases as the X-ray dose is increased, it may be necessary to take measures to dissipate the increased heat. In the first embodiment, since there is no need to increase the X-ray dose, it is possible to achieve both an improvement in the SN ratio and a longer lifespan, and to omit heat dissipation measures.

[0053] In addition, the control device 20 of the first embodiment has a function of deriving the standard deviation of the noise value from the average energy of the X-rays transmitted through the object F and the pixel values of each pixel in the X-ray image. Here, in the comparative example, for example, when the average energy changes, the relationship between the pixel value and the noise in the X-ray image fluctuates, and the noise cannot be sufficiently removed even using a learned model. On the other hand, in the present embodiment, since the average energy of the X-rays transmitted through the object F is considered and the standard deviation of the noise value in the pixel value of each pixel in the X-ray image is derived, noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise value in the X-ray image can be realized. As a result, the noise in the X-ray image can be removed more effectively. Also, in the comparative example, it was necessary to construct a different learned model for each average energy. On the other hand, according to the first embodiment, since the difference in average energy is reflected in the noise standard deviation map and the noise standard deviation map is input to the learned model, the number of learned models that need to be constructed is one. Thereby, the learning time for constructing the learned model 207 is greatly reduced.

[0054] Further, the control device 20 of the first embodiment has a function of receiving an input of condition information indicating either the operating conditions of the X-ray irradiator 50 or the imaging conditions by the X-ray detection camera 10, and calculating the average energy based on the condition information. Furthermore, the condition information includes at least any one of the tube voltage of the X-ray irradiator 50, information about the object F, information about the filter provided in the X-ray irradiator 50, information about the filter provided in the X-ray detection camera 10, and information about the scintillator provided in the X-ray detection camera 10. According to such a configuration, since the average energy of the X-rays transmitted through the object F is accurately calculated, noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise value can be realized. As a result, the noise in the X-ray image can be removed even more effectively.

[0055] Further, in the control device 20 of the first embodiment, the spread of the noise value is evaluated as the standard deviation of the noise value. As a result, the spread of the noise value in the pixel value of each pixel of the X-ray image is evaluated more precisely, so that noise removal corresponding to the relationship between the pixel value and the noise can be realized. As a result, the noise in the X-ray image can be removed more effectively.

[0056] Further, the control device 20 of the first embodiment has a function of constructing, by machine learning, a learned model 207 that outputs noise removal image data based on a training image that is an X-ray image, a noise standard deviation map generated from the training image based on the relational expression between the pixel value and the standard deviation of the noise value, and noise removal image data that is data with noise removed from the training image. According to such a configuration, the learned model 207 used for noise removal of the X-ray image is constructed by machine learning using the training data. As a result, when the training image and the noise standard deviation map generated from the training image are input to the learned model 207, noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise value can be realized. As a result, the noise in the X-ray image can be removed more effectively.

[0057] Further, in the control device 20 of the first embodiment, in order to generate a noise standard deviation map that is training data for machine learning, based on the relational expression between the pixel value and the standard deviation of the noise value, the standard deviation of the noise value is derived from the pixel value of each pixel of the training image, and a noise standard deviation map is generated, which is data associating the standard deviation of the noise derived for each pixel of the training image. According to such a configuration, the noise standard deviation map that is training data for machine learning corresponds to the relational expression between the pixel value and the standard deviation of the noise value. As a result, when the X-ray image and the noise standard deviation map generated from the X-ray image are input to the learned model 207, noise removal corresponding to the relational expression between the pixel value and the standard deviation of the noise value can be realized. As a result, the noise in the X-ray image can be removed more effectively. [Modification Example of the Control Device 20 of the First Embodiment]

[0058] FIG. 11 is a block diagram showing the functional configuration of the control device 20A in a modification of the first embodiment. Compared with the first embodiment described above, the control device 20A has a function of deriving the average energy of X-rays from the pixel values of the X-ray image in the calculation unit 202A, and a function of deriving a noise standard deviation map based on the pixel values of the X-ray image and the average energy of the X-rays derived from the X-ray image in the noise map generation unit 204A. FIG. 12 is a flowchart showing the procedure of the observation process by the image acquisition device 1 including the control device 20A of FIG. 11. As shown in FIG. 12, in the control device 20A, the process shown in step S103 of the control device 20 according to the first embodiment shown in FIG. 8 is performed immediately after step S100. Then, in the control device 20A, the processes shown in S102A and S104A are executed in place of the processes in steps S102 and S104 of the control device 20.

[0059] The calculation unit 202A calculates the average energy from the pixel values of each pixel of the radiation image (step S102A). Specifically, the calculation unit 202A derives in advance the relationship between the pixel value and the average energy for each condition information by simulation calculation of the X-ray spectrum or the like. The calculation unit 202A acquires condition information including at least the tube voltage acquired by the input unit 201 and the information of the scintillator provided in the X-ray detection camera 10. Then, the calculation unit 202A selects the relationship corresponding to the condition information from the relationships between the pixel value and the average energy derived in advance based on the condition information. Further, the calculation unit 202A derives the average energy for each pixel from the pixel values of each pixel of the X-ray image acquired by the image acquisition unit 203 based on the selected relationship.

[0060] Hereinafter, the derivation of the relationship between the pixel value and the average energy for each condition information by the calculation unit 202A will be described with reference to FIGS. 13 to 17.

[0061] First, the calculation unit 202A derives a graph G18 representing the relationship between the thickness of the object F and the X-ray transmittance, and a graph G19 representing the relationship between the thickness of the object F and the average energy of the X-rays, based on the condition information. Specifically, as shown in parts (a) to (d) of FIG. 13, the calculation unit 202A calculates, by simulation calculation, the energy spectra G14 to G17 of the X-rays transmitted when the thickness of the object F is variously changed, based on the condition information including at least the tube voltage and the information of the scintillator included in the X-ray detection camera 10. FIG. 13 is a graph showing an example of the simulation calculation result of the energy spectrum of the X-rays transmitted through the object F by the calculation unit 202A. Here, the energy spectra G14 to G17 of the transmitted X-rays are exemplified when the simulation calculation is performed by gradually increasing the thickness of the object F composed of water. Further, the calculation unit 202A calculates the average energy of the X-rays transmitted when the thickness of the object F is variously changed, based on the calculated energy spectra G14 to G17. Note that, in addition to the simulation calculation, the calculation unit 202A may obtain the relationship between the thickness and the average energy of the object F based on an X-ray image obtained by imaging a structure with a known thickness.

[0062] Furthermore, the calculation unit 202A also derives the relationship between the thickness of the object F and the X-ray transmittance based on the above simulation results. FIG. 14 is a chart showing an example of the relationship between the thickness of the object F, the average energy, and the transmittance derived by the calculation unit 202A. As shown in FIG. 14, the average energy of the transmitted X-rays and the X-ray transmittance are derived corresponding to each of the energy spectra G14 to G17 calculated for each thickness of the object F.

[0063] Subsequently, the calculation unit 202A derives a graph G18 showing the relationship between the thickness of the object F and the X-ray transmittance from the X-ray transmittances derived for objects F with various thicknesses. FIG. 15 is a graph showing the relationship between the thickness of the object F and the X-ray transmittance for the object F, derived by the calculation unit 202A. In addition, the calculation unit 202A derives a graph G19 showing the relationship between the thickness of the object F and the average energy of the X-rays from the average energies of the X-rays derived for objects F with various thicknesses. FIG. 16 is a graph showing the relationship between the thickness of the object F and the average energy of the X-rays transmitted through the object F, derived by the calculation unit 202A.

[0064] Then, based on the two graphs G18 and G19 derived for each of the various condition information, the calculation unit 202A derives a graph G20 showing the relationship between the pixel value and the average energy of the X-ray image, as shown in FIG. 17, for each of the various condition information. FIG. 17 is a graph showing the relationship between the pixel value and the average energy of the X-ray image, derived by the calculation unit 202A. Specifically, the calculation unit 202A derives the pixel value I0 of the X-ray transmission image when the object F does not exist, based on the condition information. Then, the calculation unit 202A sets the pixel value I of the X-ray image when the object F exists and calculates I / I0, which is the X-ray transmittance. Further, the calculation unit 202A derives the thickness of the object F from the calculated X-ray transmittance I / I0 based on the graph G18 showing the relationship between the thickness of the object F and the X-ray transmittance for the object F. Finally, the calculation unit 202A derives the average energy of the transmitted X-rays corresponding to that thickness, based on the derived thickness of the object F and the graph G19 showing the relationship between the thickness of the object F and the average energy of the transmitted X-rays. Subsequently, the calculation unit 202A performs the above derivation for each of the various condition information while varying the pixel value I of the X-ray image variously, thereby deriving a graph G20 showing the relationship between the pixel value of the X-ray image and the average energy of the transmitted X-rays, for each of the condition information.

[0065] Here, an example of deriving the average energy based on the pixel values by the calculation unit 202A will be described. For example, assume that the calculation unit 202A derives that the pixel value of the X-ray transmission image when the object F does not exist is I0 = 5000 based on the condition information, and sets the pixel value of the X-ray image when the object F exists to be I = 500. In this case, the calculation unit 202A calculates that the transmission rate of the X-ray is I / I0 = 0.1. Subsequently, the calculation unit 202A derives that the thickness corresponding to the X-ray transmission rate of 0.1 is 30 mm based on the graph G18 showing the relationship between the thickness of the object F and the X-ray transmission rate for the object F. Further, the calculation unit 202A derives that the average energy corresponding to the pixel value 500 is 27 keV based on the graph G19 showing the relationship between the thickness of the object F and the average energy of the transmitted X-ray. Finally, the calculation unit 202A repeats the derivation of the average energy of the X-ray for each pixel value and derives a graph G20 showing the relationship between the pixel values and the average energy of the X-ray image.

[0066] Furthermore, the calculation unit 202A selects the graph G20 corresponding to the condition information acquired by the input unit 201 from among the plurality of graphs G20 derived in advance by the above procedure. The calculation unit 202A derives the average energy of the transmitted X-ray corresponding to the pixel value of each pixel of the X-ray image acquired by the image acquisition unit 203 based on the selected graph G20.

[0067] Note that the calculation unit 202A may derive the average energy of X-rays from the condition information acquired by the input unit 201 and the pixel values of each pixel of the X-ray image, rather than deriving in advance the relationship between the pixel value and the average energy of X-rays for each condition information. Specifically, the calculation unit 202A derives the pixel value I0 of the X-ray image when the object does not exist based on the condition information. Then, the calculation unit 202A calculates the transmittance by obtaining the ratio of the pixel value I of each pixel of the X-ray image acquired by the image acquisition unit 203 to the pixel value I0. Further, the calculation unit 202A derives the thickness based on the graph G18 showing the relationship between the thickness and the transmittance of X-rays and the calculated transmittance. Then, the calculation unit 202A derives the average energy based on the graph G19 showing the relationship between the thickness and the average energy and the derived thickness, thereby deriving the average energy for each pixel value of each pixel of the X-ray image.

[0068] The noise map generation unit 204A generates a noise standard deviation map from the X-ray image acquired by the image acquisition unit 203 and the average energy of X-rays corresponding to each pixel of the X-ray image derived by the calculation unit 202A (step S104A). Specifically, the noise map generation unit 204A substitutes the pixel value of each pixel of the X-ray image acquired by the image acquisition unit 203 and the average energy derived for each pixel by the calculation unit 202A into the relational expression (4) to derive the standard deviation of the noise value for each pixel considering the thickness of the object. The noise map generation unit 204A generates the standard deviation of the noise value corresponding to each pixel of the X-ray image as a noise standard deviation map.

[0069] FIG. 18 is a graph showing an example of the relationship between the pixel value and the standard deviation of the noise value. This graph shows the relationship between the standard deviation of the noise value derived from the pixel value of the X-ray image and the pixel value of the X-ray image by the calculation unit 202A and the noise map generation unit 204A according to this modified example. In this modified example, since the standard deviation of the noise value is derived in consideration of the thickness of the object, the greater the pixel value, the smaller the thickness of the object, and the lower the average energy in the pixel. Therefore, as estimated from the relational expression (4), the change in the standard deviation of the noise value when the pixel value increases is different between the first embodiment and this modified example. In the example shown in FIG. 18, in the graph G22 of this modified example, the degree of increase in the standard deviation of the noise value when the pixel value increases is smaller than that of the graph G21 of the first embodiment.

[0070] In the control device 20A according to the modified example of the first embodiment, the average energy is calculated from the pixel value of each pixel of the X-ray image. Here, for example, when there are a plurality of objects with different thicknesses and materials in the X-ray image, the average energy varies greatly for each object, and noise cannot be sufficiently removed from the X-ray image. According to such a configuration, since the average energy of the X-rays transmitted through the object F is calculated for each pixel value of the X-ray image, for example, considering differences in thickness, material, etc., noise removal corresponding to the relationship between the pixel value and the noise of each pixel of the X-ray image can be realized. As a result, the noise in the X-ray image can be effectively removed.

[0071] Note that the control device 20A according to this modified example derives the average energy from the pixel values of the X-ray image using the graph G20 derived for each of various condition information. At this time, the average energy may be derived from the pixel values while ignoring the difference in the material of the object F. FIG. 19 is a graph showing the relationship between the pixel values of the X-ray image derived by the calculation unit 202A and the standard deviation of the noise values. Here, the change in the material of the object F is also considered as condition information and the relationship is derived. Graph G24 shows a derivation example when the material is aluminum, graph G23 shows a derivation example when the material is PET (Polyethylene terephthalate), and graph G25 shows a derivation example when the material is copper. Thus, even when the material of the object F changes, if the tube voltage of the X-ray irradiator 50 and the information of the scintillator provided in the X-ray detection camera 10 used for imaging the object F are the same, the relationship between the pixel value and the average energy of the transmitted X-ray does not change significantly, so the relationship between the pixel value and the standard deviation of the noise value also does not change significantly. Considering such a property, even if the control device 20A ignores the difference in the material of the object F as condition information, it can accurately derive the average energy from the pixel values of the X-ray image. Even in such a case, according to the control device 20A of this modified example, noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise can be realized. As a result, the noise in the X-ray image can be removed more effectively. [Second Embodiment]

[0072] FIG. 20 is a block diagram showing the functional configuration of the control device 20B in the second embodiment. The control device 20B is different from the control device 20 in the first embodiment described above in that the image acquisition unit 203B has a function of acquiring an X-ray image of the jig, and the noise map generation unit 204B has a function of deriving a graph showing the relationship between the pixel value and the standard deviation of the noise value from the X-ray image of the jig. FIG. 21 is a flowchart showing the procedure of the observation process by the image acquisition device 1 including the control device 20B in FIG. 20. As shown in FIG. 21, in the control device 20B according to the second embodiment, the processes shown in steps S201 and S202 are executed by replacing the processes in steps S101, S102, and S104 by the control device 20 according to the first embodiment shown in FIG. 8.

[0073] The image acquisition unit 203B acquires a radiation image of a jig irradiated with radiation and imaged after passing through the jig (step S201). Specifically, the image acquisition unit 203B acquires an X-ray image obtained by irradiating and imaging the jig and the object F with X-rays using the image acquisition device 1. As the jig, a flat plate member or the like with a known thickness and material is used. That is, the image acquisition unit 203B acquires an X-ray image of the jig imaged using the image acquisition device 1 prior to the observation process of the object F. Then, the image acquisition unit 203B acquires an X-ray image of the object F imaged using the image acquisition device 1. However, the acquisition timing of the X-ray images of the jig and the object F is not limited to the above, and they may be simultaneous or in the reverse timing (step S103). Also, the image acquisition unit 203B, similar to the image acquisition unit 203, acquires an X-ray image in which the object F is irradiated with X-rays and the X-rays passing through the object F are imaged.

[0074] In the image acquisition device 1, the jig is set and imaged, and the noise map generation unit 204B derives relationship data representing the relationship between the pixel value and the evaluation value obtained by evaluating the spread of the noise value from the resulting radiation image of the jig (step S202). Specifically, the noise map generation unit 204B derives a noise standard deviation map representing the relationship between the pixel value and the standard deviation of the noise value from the X-ray image of the jig.

[0075] FIG. 22 is a diagram showing an example of generation of a noise standard deviation map by the noise map generation unit 204B. The noise map generation unit 204B derives a relationship graph G27 representing the correspondence between the pixel value and the standard deviation of the noise value from the X-ray image G26 of the jig. Then, in the same manner as in the first embodiment, the noise map generation unit 204B derives relationship data G2 representing the correspondence between each pixel position and the pixel value from the X-ray image G1 acquired by the image acquisition unit 203B. Further, the noise map generation unit 204 applies the correspondence shown in the relationship graph G27 to each pixel in the relationship data G2, thereby deriving the standard deviation of the noise value corresponding to the pixel at each pixel position in the X-ray image. As a result, the noise map generation unit 204 associates the derived standard deviation of the noise with each pixel position, and derives relationship data G4 showing the correspondence between each pixel position and the standard deviation of the noise. Then, the noise map generation unit 204 generates a noise standard deviation map G5 based on the derived relationship data G4.

[0076] An explanation will be given of the derivation of the relationship graph G27 representing the relationship between the pixel value and the standard deviation of the noise value from the X-ray image G26 of the jig by the noise map generation unit 204B. FIG. 23 shows an example of the structure of the jig used for imaging in the second embodiment. For example, a member P1 whose thickness changes stepwise in one direction can be used for the jig. FIG. 24 shows an example of the X-ray image of the jig in FIG. 23. First, the noise map generation unit 204B derives the pixel value (hereinafter referred to as the true pixel value) when there is no noise for each step of the jig in the X-ray image G26 of the jig, and derives the standard deviation of the noise value based on the true pixel value. Specifically, the noise map generation unit 204B derives the average value of the pixel values in a certain step of the jig. Then, the noise map generation unit 204B sets the derived average value of the pixel values as the true pixel value in that step. The noise map generation unit 204B derives the difference between each pixel value and the true pixel value as the noise value in that step. The noise map generation unit 204B derives the standard deviation of the noise value from the derived noise values for each pixel value.

[0077] Then, the noise map generation unit 204B derives the relationship between the true pixel value and the standard deviation of the noise value as a relationship graph G27 of the pixel value and the standard deviation of the noise value. Specifically, the noise map generation unit 204B derives the true pixel value and the standard deviation of the noise value for each step of the jig. The noise map generation unit 204B plots the derived relationship between the true pixel value and the standard deviation of the noise value on a graph and draws an approximation curve to derive a relationship graph G27 representing the relationship between the pixel value and the standard deviation of the noise value. For the approximation curve, exponential approximation, linear approximation, logarithmic approximation, polynomial approximation, power approximation, etc. are used.

[0078] In the control device 20B of the second embodiment, the relationship data is generated based on the radiation image obtained by imaging the actual jig. As a result, the relationship data optimal for noise removal of the radiation image of the object F can be obtained. As a result, the noise in the radiation image can be removed more effectively.

[0079] Note that the noise map generation unit 204B may derive the relationship between the pixel value and the standard deviation of the noise value from the captured image when the tube current or the exposure time is changed in a state without an object without using a jig. According to such a configuration, since the relationship data is generated based on the radiation image actually captured and the noise map is generated, noise removal corresponding to the relationship between the pixel value and the spread of the noise can be realized. As a result, the noise in the radiation image can be removed more effectively.

[0080] Specifically, the image acquisition unit 203B acquires a plurality of radiation images captured in a state where there is no object (step S201), and the noise map generation unit 204B may derive the relationship between the pixel value and the standard deviation of the noise value from the radiation images acquired by the image acquisition unit 203B (step S202). The plurality of radiation images are a plurality of images in which at least one of the conditions of the radiation source or the imaging conditions is different from each other. As an example, the image acquisition unit 203B acquires a plurality of X-ray images captured using the image acquisition device 1 in a state where there is no object F prior to the observation process of the object F while the tube current or the exposure time is changed. Then, the noise map generation unit 204B derives the true pixel value for each X-ray image, and in the same manner as in the second embodiment, derives the standard deviation of the noise based on the true pixel value. Further, the noise map generation unit 204B plots the relationship between the true pixel value and the standard deviation of the noise on a graph and draws an approximation curve in the same manner as in the second embodiment, thereby deriving a relationship graph representing the relationship between the pixel value and the standard deviation of the noise value. Finally, the noise map generation unit 204B generates a noise standard deviation map from the X-ray images acquired by the image acquisition unit 203B based on the derived relationship graph in the same manner as in the first embodiment. [Third Embodiment]

[0081] FIG. 25 is a configuration diagram of an image acquisition device 1C which is a radiation image processing system according to the third embodiment. FIG. 26 is a block diagram showing an example of the functional configuration of the control device 20C according to the third embodiment. The image acquisition device 1C according to the third embodiment is different from the above-described first or second embodiment in that it includes an X-ray detection camera 10C (imaging device) having a two-dimensional sensor 12C or the like, and a control device 20C having a construction unit 206C and a learned model 207C, and in that it does not include a belt conveyor 60.

[0082] The image acquisition device 1C uses X-ray transmission images to perform foreign object inspection, weight inspection, product inspection, etc. on the object F. Applications include food inspection, baggage inspection, substrate inspection, battery inspection, material inspection, etc. Furthermore, applications of the image acquisition device 1C include medical applications, dental applications, industrial applications, etc. Medical applications include, for example, chest X-ray, mammography, CT (computed tomography), dual energy CT, tomosynthesis, etc. Dental applications include transmission, panorama, and CT, etc. Industrial applications include non-destructive inspection, security, and battery inspection, etc.

[0083] The image acquisition device 1C according to the third embodiment outputs an X-ray image obtained by imaging an X-ray transmission image based on X-rays that penetrate the stationary object F. However, the image acquisition device 1C may have a belt conveyor 60 as in the above-described image acquisition device 1 and may be configured to image the conveyed object F.

[0084] FIG. 27 is a block diagram showing the configuration of the X-ray detection camera 10C. As shown in FIGS. 25 and 27, the X-ray detection camera 10C includes a filter 19, a scintillator layer 11C, a two-dimensional sensor 12C, a sensor control unit 13C, and an output unit 14C. The sensor control unit 13C is electrically connected to the two-dimensional sensor 12C, the output unit 14C, and the control device 20C. The output unit 14C is also electrically connected to the two-dimensional sensor 12C and the control device 20C.

[0085] The scintillator layer 11C is fixed on the two-dimensional sensor 12C by adhesion or the like and converts X-rays that have passed through the object F into scintillation light (the detailed configuration will be described later). The scintillator layer 11C outputs the scintillation light to the two-dimensional sensor 12C. The filter 19 transmits a predetermined wavelength range of X-rays toward the scintillator layer 11C.

[0086] The two-dimensional sensor 12C detects scintillation light from the scintillator layer 11C, converts it into charges, and outputs it as a detection signal (electrical signal) to the output unit 14C. The two-dimensional sensor 12C is, for example, a line sensor or a flat panel sensor, and is disposed on the substrate 15C. The two-dimensional sensor 12C has M×N pixels P 1,1 ~P M,N which are two-dimensionally arranged in M rows and N columns. The M×N pixels P 1,1 ~P M,N are arranged at a constant pitch in both the row direction and the column direction. The pixel P m,n is located at the m-th row and the n-th column. The N pixels P m,1 ~P m,N in the m-th row are each connected to the sensor control unit 13C by the m-th row selection wiring L V,m . The output terminals of the M pixels P 1,n ~P M,n in the n-th column are each connected to the output unit 14C by the n-th column readout wiring L O,n . Note that M and N are each integers of 2 or more, m is each integer from 1 to M, and n is each integer from 1 to N.

[0087] The output unit 14C outputs a digital value generated based on the amount of charges input via the readout wiring L O,n . The output unit 14C includes N integration circuits 41(1) to 41(N), N hold circuits 42(1) to 42(N), an AD conversion unit 43, and a storage unit 44. Each integration circuit 41(n) has a common configuration. Also, each hold circuit 42(n) has a common configuration.

[0088] Each integration circuit 41(n) accumulates the charges input to the input terminal via any one of the column readout wirings L O,n . Each integration circuit 41(n) outputs a voltage value corresponding to the accumulated charge amount from the output terminal to the hold circuit 42(n). Each of the N integration circuits 41(1) to 41(N) is connected to the sensor control unit 13C by the reset wiring L R .

[0089] Each holding circuit 42(n) has an input terminal connected to the output terminal of the integrating circuit 41(n). Each holding circuit 42(n) holds the voltage value input to the input terminal and outputs the held voltage value from the output terminal to the AD conversion unit 43. Each of the N holding circuits 42(1) to 42(N) is connected to the sensor control unit 13C by a holding wiring L H and is also connected to the sensor control unit 13C by a wiring L for selecting the n-th column H,n .

[0090] The AD conversion unit 43 inputs the voltage values output from each of the N holding circuits 42(1) to 42(N) and performs AD conversion processing on the input voltage values (analog values). The AD conversion unit 43 outputs a digital value corresponding to the input voltage value to the storage unit 44. The storage unit 44 inputs and stores the digital values output from the AD conversion unit 43 and outputs the stored digital values in order.

[0091] The sensor control unit 13C outputs the m-th row selection control signal Vsel(m) to each of the N pixels P V,m in the m-th row via the m-th row selection wiring L m,1 ~P m,N . The sensor control unit 13C outputs the reset control signal Reset to each of the N integrating circuits 41(1) to 41(N) via the reset wiring L R . The sensor control unit 13C outputs the hold control signal Hold to each of the N holding circuits 42(1) to 42(N) via the hold wiring L H . The sensor control unit 13C outputs the n-th column selection control signal Hsel(n) to the holding circuit 42(n) via the n-th column selection wiring L H,n . Further, the sensor control unit 13C controls the AD conversion process in the AD conversion unit 43 and also controls the writing and reading of digital values in the storage unit 44.

[0092] The structure of the scintillator layer 11C disposed on the two-dimensional sensor 12C having the above structure will be described. FIG. 28 is a diagram showing the structure of the scintillator layer 11C, a cross-sectional view along the thickness direction of the substrate 15C upward, and a plan view showing the substrate 15C as viewed from the surface side where the two-dimensional sensor 12C is disposed downward. Thus, the scintillator layer 11C is separated and arranged corresponding to M×N pixels P 1,1 ~P M,N into K×L (K and L are integers of 1 or more) rectangular scintillator portions Q 1,1 ~Q K,L and a separation portion R located between these scintillator portions Q 1,1 ~Q K,L is formed. Note that the number L may be 1 or more and N or less, and the number K may be 1 or more and M or less. Further, the number L may be an integer of 1 or more and an integer obtained by dividing N by an integer, and the number K may be an integer of 1 or more and an integer obtained by dividing M by an integer. In this case, blurring due to the spread of light can be suppressed according to the interval of the separation portion R of the scintillator layer 11C. Also, the number L may be an integer larger than N, or the number K may be an integer larger than M. In this case, the interval of the separation portion R of the scintillator layer 11C becomes smaller than the interval between a plurality of pixels P 1,1 ~P M,N but the alignment between the scintillator layer 11C and a plurality of pixels P 1,1 ~P M,N becomes easy. In the present embodiment, for example, L = N and K = M, but it is not limited thereto.

[0093] The K×L scintillator portions Q 1,1 ~Q K,L are made of a scintillator material capable of converting incident X-rays into scintillation light, and are arranged so as to cover the entire pixels P 1,1 ~P M,N . As an example, M×N scintillator portions Q 1,1 ~Q M,N are arranged so as to cover the entire corresponding pixels P 1,1 ~P M,N . The separation portion R is the K×L scintillator portions Q1,1 ~Q K,L It is formed in a mesh shape so as to separate ~Q, and is composed of a material capable of shielding scintillation light. Further, the separation part R may contain a material that reflects scintillation light. Furthermore, the separation part R may be composed of a material capable of shielding radiation. As the material constituting such a scintillator layer 11C and the manufacturing method of the scintillator layer 11C, for example, the materials and manufacturing methods described in Japanese Patent Application Laid-Open No. 2001-99941 or Japanese Patent Application Laid-Open No. 2003-167060 can be used. However, the material and manufacturing method of the scintillator layer 11C are not limited to those described in the above documents.

[0094] The control device 20C generates an X-ray image based on the digital signal output from the X-ray detection camera 10C (more specifically, the storage unit 44 of the output unit 14C). The generated X-ray image is output to the display device 30 after being subjected to noise removal processing described later, and is displayed by the display device 30. Further, the control device 20C controls the X-ray irradiator 50 and the sensor control unit 13C. Note that the control device 20C of the third embodiment is a device provided independently outside the X-ray detection camera 10C, but may be integrated inside the X-ray detection camera 10C.

[0095] Here, the construction function of the learned model 207C by the construction unit 206C in the third embodiment will be described. FIG. 29 is a flowchart showing a procedure for creating image data which is teacher data (training images in the first and second embodiments) used for constructing the learned model 207C by the construction unit 206C.

[0096] The image data that is teacher data (also referred to as teacher image data) is created by a computer according to the following procedure. First, an image of a structure having a predetermined structure (structure image) is created (step S301). For example, an image of a structure (e.g., a jig) having a predetermined structure may be created by simulation calculation. Alternatively, an X-ray image of a structure such as a chart having a predetermined structure may be acquired to create a structure image. Next, for one pixel selected from among a plurality of pixels constituting the structure image, a sigma value that is the standard deviation of the pixel value is calculated (step S302). Then, based on the sigma value obtained in step S302, a noise distribution is set (step S303). This noise distribution is set so that the probability that the pixel value to which noise is added exceeds the original pixel value is higher compared to a normal distribution (Poisson distribution). In particular, the probability that the pixel value to which noise is added becomes 1.2 times or more the original pixel value is high (details will be described later). In this way, by setting the noise distribution based on the sigma value, teacher data under various noise conditions can be generated. Subsequently, a randomly set noise value is calculated along the noise distribution set based on the sigma value in step S303 (step S304). Further, by adding the noise value obtained in step S304 to the pixel value of one pixel, a pixel value constituting the image data that is teacher data is generated (step S305). The processes from step S302 to step S305 are performed for each of the plurality of pixels constituting the structure image (step S306), and teacher image data that is teacher data is generated (step S307). Also, when more teacher image data is required, it is determined whether to perform the processes from step S301 to step S307 for another structure image (step S308), and another teacher image data that is teacher data is generated. Note that the another structure image may be an image of a structure having the same structure or an image of a structure having a different structure.

[0097] FIG. 30 and FIG. 31 are diagrams showing examples of the noise distribution set in step S303 described above. The horizontal axes of FIGS. 30 and 31 indicate the pixel values after adding the noise values, with the pixel value before adding the noise value of the pixel being set to 100 (hereinafter referred to as the noise pixel value). The vertical axes of FIGS. 30 and 31 are the relative values of the frequencies of the noise pixel values. The relative value of the frequency of the noise pixel value is a value indicating the relative frequency of the pixel value after adding noise to the pixel value. When the relative value at each noise pixel value is divided by the total value of the relative values at each noise pixel value, it becomes a value indicating the probability of each noise pixel value when the pixel value is added with the noise value. The noise distributions G28 and G29 shown in FIGS. 30 and 31 are set in consideration of the fact that X-rays detected by the sensor in imaging using a scintillator appear as white dots in the X-ray image. For this reason, compared with the normal distribution (Poisson distribution), the noise distributions G28 and G29 have a higher probability that the noise pixel value exceeds the original pixel value. In particular, the probability that the noise pixel value is 120 or more (the noise pixel value corresponding to the case where a white dot occurs in the X-ray image) is high. Here, the case where the X-rays detected by the sensor appear as white dots in the X-ray image means that the X-rays are not absorbed by the scintillator and pass through the scintillator, and are directly converted into electrons by the sensor. When the X-rays are absorbed by the scintillator and converted into visible light, the sensor detects the visible light. On the other hand, when the X-rays are not absorbed by the scintillator, the X-rays are directly converted into electrons in the sensor. That is, the sensor that detects visible light detects not only the scintillation light (visible light) generated in the scintillator layer 11C but also the X-rays that have passed through the scintillator layer 11C. At this time, the number of electrons generated from the X-rays incident on the sensor is larger than when the X-rays are converted into visible light by the scintillator. Therefore, the X-rays that have passed through the scintillator layer 11C appear as white dots in the X-ray image. The occurrence of this white dot causes noise in the X-ray image. Therefore, by constructing the learned model 207C using the above-described noise distribution and performing noise removal using the constructed learned model 207C, the white dots appearing in the X-ray image can be removed as noise.Note that the case where white spots are likely to occur is, for example, when the tube voltage of the X-ray irradiator is high.

[0098] Note that a large number of image data, which is the teacher data used for constructing the learned model 207C, needs to be prepared. Also, for the structural image, an image with less noise is better, and ideally, an image without noise is better. Therefore, when generating a structural image by simulation calculation, a large number of images without noise can be generated, so it is effective to generate a structural image by simulation calculation.

[0099] The image acquisition device 1C of the third embodiment includes a two-dimensional sensor 12C as a flat panel sensor. And for each pixel P 1,1 ~P M,N of the two-dimensional sensor 12C, a scintillator part Q 1,1 ~Q M,N of the scintillator layer 11C and a separation part R are provided. Thereby, in the X-ray image acquired by the image acquisition device 1C, blurring of the image is reduced. As a result, the contrast in the X-ray image becomes higher and the intensity of noise becomes higher. Here, in the image acquisition device 1C of the third embodiment, noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise value in the X-ray image is performed using a learned model 207C constructed in advance by machine learning. Thereby, only the intensity of the noise is reduced in the above X-ray image. From the above, the image acquisition device 1C can acquire an X-ray image with reduced noise intensity and enhanced contrast.

[0100] Examples of the simulation results of the X-ray images acquired by the image acquisition device 1C are shown in FIGS. 32(a), 32(b), and 32(c). The X-ray image G30 shown in FIG. 32(a) is an X-ray image generated by simulation calculation based on the condition that imaging was performed using a scintillator made of CsI (cesium iodide) as the scintillator layer 11C. The scintillator made of CsI has, for example, a single-sheet shape extending along the pixels of the two-dimensional sensor 12C. The thickness of the scintillator made of CsI is set to 450 μm. The X-ray image G31 shown in FIG. 32(b) is an X-ray image generated by simulation calculation based on the condition that imaging was performed using the pixel scintillator having the structure shown in FIG. 28 as the scintillator layer 11C. The X-ray image G32 shown in FIG. 32(c) is an X-ray image generated by simulation calculation based on the condition that imaging was performed using the pixel scintillator as the scintillator layer 11C and noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise value was performed (the condition that the same noise removal as that of the image acquisition device 1C according to the third embodiment was performed). In this case, the pixel scintillators are provided for each pixel. The thickness of the pixel scintillator is set to, for example, 200 μm. The thickness along the direction in which the pixels of the partition walls (separation portions) of the pixel scintillators are arranged is set to 40 μm. Note that the X-ray image G32 is the X-ray image after noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise value has been executed in the X-ray image G31. Also, in each simulation, the pixels of the two-dimensional sensor are set to have a rectangular shape with a side length of 120 μm.

[0101] Hereinafter, in the X-ray images G30, G31, and G32, the value indicating the noise magnitude is the standard deviation value of the intensity in the background portion (the portion where no black dots are shown). The value indicating the contrast is the difference between the average value of the intensity in the background portion and the minimum value of the intensity in the portion where black dots are shown. Furthermore, the CN ratio (CNR: Contrast to Noise Ratio) in the X-ray images G30, G31, and G32 is the value obtained by dividing the value indicating the contrast by the value indicating the noise magnitude. In the X-ray images G30, G31, and G32, the values indicating the noise magnitude are 301.8, 1420.0, and 37.9 respectively, and the values indicating the contrast are 3808.1, 9670.9, and 8844.3 respectively. And in the X-ray images G30, G31, and G32, the CN ratios are 12.62, 6.81, and 233.16 respectively.

[0102] In the X-ray image G31 using the pixel scintillator, compared with the X-ray image G30, the contrast is higher and the noise is also larger. In other words, the CN ratio in the X-ray image G31 is 1 / 2 times the CN ratio of the X-ray image G30. That is, just using the pixel scintillator as the scintillator layer 11C cannot obtain an X-ray image with sufficiently removed noise. In contrast, according to the third embodiment, for the X-ray image obtained after using the pixel scintillator as the scintillator layer 11C, noise removal corresponding to the relationship between the pixel value and the standard deviation of the noise value in the X-ray image is performed using the pre-trained model 207C constructed by machine learning in advance. Thereby, in the X-ray image G32 according to the third embodiment, the contrast is higher and the noise is reduced compared with the X-ray image G30. And the CN ratio in the X-ray image G32 is 20 times the CN ratio of the X-ray image G30. That is, since the image acquisition device 1C according to the third embodiment has the same conditions as the simulation conditions of the X-ray image G32, an X-ray image with sufficiently removed noise can be obtained.

[0103] Also, in the control device 20C of the third embodiment, compared with the normal distribution, the noise distribution has a higher probability that the pixel value with noise added exceeds the original pixel value. Then, a pixel value with noise added is calculated along the noise distribution, and teacher image data is generated. Using the generated teacher image data, a learned model 207C is constructed. The X-ray image and the noise standard deviation map are input to the constructed learned model 207C, and image processing for removing noise from the X-ray image is executed. According to such a configuration, in consideration of the fact that the X-ray detected by the sensor appears as white dots in the X-ray image in imaging using the scintillator, image processing for removing noise from the X-ray image is executed. As a result, in the image acquisition device 1C using the scintillator layer 11C, an X-ray image with noise more effectively removed can be acquired. [Supplementary Explanation of the Construction Unit 206]

[0104] The construction of the learned model 207 by the construction unit 206 is executed in the same manner as the construction of the learned model 207C by the construction unit 206C. However, the noise distribution set in step S303 is not the noise distribution such as the above-described noise distributions G28 and G29, but the normal distribution. FIG. 33 is a diagram showing the normal distribution G33 used for generating teacher data. The horizontal axis of FIG. 33 shows the pixel value with noise added, with the pixel value before adding the noise value of the pixel being set to 100. The vertical axis of FIG. 33 is a relative value representing the frequency of the noise pixel value.

[0105] As described above, various embodiments of the present disclosure have been explained, but the embodiments of the present disclosure are not limited to the above embodiments. For example, the X-ray detection camera 10 is not limited to a dual-line X-ray camera, but may be a single-line X-ray camera, a dual-energy X-ray camera, a TDI (Time Delay Integration) scan X-ray camera, a multi-line X-ray camera having a plurality of lines of two or more lines, a two-dimensional X-ray camera, an X-ray flat panel sensor, an X-ray II, a direct conversion type X-ray camera that does not use a scintillator (a-Se, Si, CdTe, CdZnTe, TlBr, PbI2, etc.), a camera of an observation method using an optical lens by lens coupling for a scintillator, an imaging tube sensitive to radiation, or a point sensor sensitive to radiation. Further, the image acquisition device 1 is not limited to the above embodiment, and may be a radiation image processing system that images the object F in a stationary state, such as a CT (Computed Tomography) device. Furthermore, it may be a radiation image processing system that images while rotating the object F.

[0106] Also, in the above-described embodiment, in the noise map generation step, it is preferable to derive an evaluation value from the average energy of the radiation transmitted through the object and the pixel value of each pixel of the radiation image. Further, in the above embodiment, in the noise map generation unit, it is preferable to derive an evaluation value from the average energy of the radiation transmitted through the object and the pixel value of each pixel of the radiation image. Here, in the comparative example, for example, when the average energy changes, the relationship between the pixel value and the noise in the radiation image fluctuates, and the noise cannot be sufficiently removed even using a learned model. In this regard, by adopting the above configuration, the average energy of the radiation transmitted through the object is considered, and the spread of the noise value in the pixel value of each pixel of the radiation image is evaluated. Therefore, noise removal corresponding to the relationship between the pixel value and the spread of the noise in the radiation image can be realized. As a result, the noise in the radiation image can be removed more effectively.

[0107] In the above-described embodiment, it is also preferable to further include an input step of receiving an input of condition information indicating either the conditions of the radiation source or the imaging conditions when irradiating radiation to image an object, and a calculation step of calculating an average energy based on the condition information. In the above-described embodiment, it is also preferable to further include an input unit that receives an input of condition information indicating either the conditions of the radiation source or the imaging conditions when irradiating radiation to image an object, and a calculation unit that calculates an average energy based on the condition information. Further, it is also preferable that the condition information includes at least any one of the tube voltage of the radiation source, information about the object, information about the filter included in the camera used for imaging the object, information about the filter included in the radiation source, and information about the scintillator included in the camera used for imaging the object. According to such a configuration, since the average energy of the radiation passing through the object is accurately calculated, noise removal corresponding to the relationship between the pixel value and the spread of noise can be realized. As a result, noise in the radiation image can be more effectively removed.

[0108] In the above-described embodiment, it is preferable to further include a calculation step of calculating an average energy from the pixel values of each pixel of the radiation image. In the above-described embodiment, it is preferable to further include a calculation unit that calculates an average energy from the pixel values of each pixel of the radiation image. According to such a configuration, since the average energy of the radiation passing through the object is accurately calculated for each pixel value of each pixel of the radiation image, noise removal corresponding to the relationship between the pixel value of each pixel of the radiation image and the spread of noise can be realized. As a result, noise in the radiation image can be more effectively removed.

[0109] Also, in the image acquisition step, it is preferable that the jig is irradiated with radiation and a radiation image of the jig through which the radiation has passed through the jig is acquired, and in the noise map generation step, related data is derived from the radiation image of the jig. Also, in the image acquisition unit, it is preferable that the jig is irradiated with radiation and a radiation image of the jig through which the radiation has passed through the jig is acquired, and in the noise map generation unit, related data is derived from the radiation image of the jig. According to such a configuration, since the related data is generated based on the radiation image actually obtained by imaging the jig and the noise map is generated, noise removal corresponding to the relationship between the pixel value and the spread of noise can be realized. As a result, noise in the radiation image can be removed more effectively.

[0110] Also, in the image acquisition step, it is preferable to acquire a plurality of radiation images imaged in a state where there is no object, and in the noise map generation step, to derive related data from the plurality of radiation images, and the plurality of radiation images are a plurality of images in which at least one of the conditions of the radiation source and the imaging conditions are different from each other. Also, in the image acquisition unit, it is preferable to acquire a plurality of radiation images imaged in a state where there is no object, and in the noise map generation unit, to derive related data from the plurality of radiation images, and the plurality of radiation images are a plurality of images in which at least one of the conditions of the radiation source and the imaging conditions are different from each other. According to such a configuration, since the related data is generated based on the radiation image actually obtained by imaging and the noise map is generated, noise removal corresponding to the relationship between the pixel value and the spread of noise can be realized. As a result, noise in the radiation image can be removed more effectively.

[0111] Also, it is preferable that the evaluation value is the standard deviation of the noise value. Thereby, since the spread of the noise value in the pixel value of each pixel of the radiation image is evaluated more precisely, noise removal corresponding to the relationship between the pixel value and the noise can be realized. As a result, noise in the radiation image can be removed more effectively.

[0112] The machine learning method according to the above embodiment uses a radiation image as a training image, and based on relationship data representing the relationship between the pixel values and the evaluation value obtained by evaluating the spread of the noise values, a noise map generated from the training image and noise removal image data which is data obtained by removing noise from the training image are used as training data, and a trained model that outputs noise removal image data based on the training image and the noise map is constructed by machine learning. In the above other aspect, a construction unit that further includes a construction step of constructing, by machine learning, a trained model that uses, as training data, a training image that is a radiation image, a noise map generated from the training image, and noise removal image data which is data obtained by removing noise from the training image, and outputs noise removal image data based on the training image and the noise map, is preferably provided. According to such a configuration, the trained model used for noise removal of the radiation image is constructed by machine learning using the training data. As a result, when a radiation image and a noise map generated from the radiation image are input to the trained model, noise removal corresponding to the relationship between the pixel values and the spread of the noise can be realized. As a result, the noise in the radiation image of the object can be removed more effectively.

[0113] Alternatively, the trained model according to the above embodiment is a trained model constructed by the above construction step, and causes a processor to execute image processing for removing noise from a radiation image of an object. In the above other aspect, the spread of the noise value evaluated from the pixel value of each pixel of the radiation image is considered, and the noise is removed from the radiation image by machine learning. As a result, noise removal corresponding to the relationship between the pixel value and the spread of the noise in the radiation image can be realized using the trained model. As a result, the noise in the radiation image can be effectively removed.

[0114] Furthermore, in order to generate a noise map which is training data for the machine learning method according to the above-described embodiment, the preprocessing method of the machine learning method derives an evaluation value from the pixel value of each pixel of the radiation image based on the relational data representing the relationship between the evaluation value obtained by evaluating the spread of the pixel value and the noise value, and generates a noise map which is data associating the evaluation value derived for each pixel of the radiation image. With such a configuration, the noise map which is training data for the above-described machine learning method corresponds to the relationship between the pixel value and the evaluation value obtained by evaluating the spread of the pixel value and the noise value. Thereby, when a radiation image and a noise map generated from the radiation image are input to the learned model constructed by the above-described machine learning method, noise removal corresponding to the relationship between the pixel value and the spread of the noise can be realized. As a result, noise in the radiation image of the object can be removed more effectively.

Explanation of Signs

[0115] 10…X-ray detection camera (imaging device) (camera), 12C…2D sensor (line sensor) (flat panel sensor), 20, 20A, 20B…control device (radiation image processing module), 50…X-ray irradiator (radiation source), 201…input unit, 202, 202A…calculation unit, 203, 203B…image acquisition unit, 204, 204A, 204B…noise map generation unit, 205…processing unit, 206…construction unit, 207…learned model, G5…noise standard deviation map (noise map), G3, G23, G24, G25…relationship graph (relational data) representing the correspondence between the pixel value and the standard deviation of the noise value, G7…training image, G26…X-ray image of jig (radiation image), F…object, P1…member (jig).

Claims

1. An image acquisition step of irradiating an object with radiation and acquiring a radiation image obtained by imaging the radiation transmitted through the object; A noise map generation step of deriving the evaluation value from the pixel value of each pixel of the radiation image based on relationship data representing the relationship between the evaluation value obtained by evaluating the spread of the pixel value and the noise value, and generating a noise map which is data associating the derived evaluation value with each pixel of the radiation image; A processing step of performing image processing for removing noise from the radiation image based on the radiation image and the noise map; Comprising: In the processing step, in the image processing, the radiation image and the noise map are input into a pre-constructed learned model by machine learning, and an output image which is an image with noise removed from the radiation image is obtained from the learned model. A radiation image processing method.

2. In the noise map generation step, the average energy of the radiation transmitted through the object and the evaluation value are derived from the pixel value of each pixel of the radiation image. The radiation image processing method according to Claim 1.

3. An input step of receiving an input of condition information indicating either the conditions of the radiation source or the imaging conditions when imaging the object by irradiating with radiation; A calculation step of calculating the average energy based on the condition information. Further comprising: The condition information includes at least any one of the tube voltage of the radiation source, information about the object, information about the filter provided in the camera used for imaging the object, information about the filter provided in the radiation source, and information about the scintillator provided in the camera. The radiation image processing method according to Claim 2.

4. Further comprising a calculation step of calculating the average energy from the pixel value of each pixel of the radiation image. The radiation image processing method according to Claim 2.

5. In the image acquisition step, a radiation image of a jig irradiated with radiation and imaged with the radiation transmitted through the jig is acquired; In the noise map generation step, the relationship data is derived from the radiation image of the jig. The radiation image processing method according to Claim 1.

6. In the image acquisition step, a plurality of radiation images imaged in a state where the object is not present are acquired; In the noise map generation step, the relationship data is derived from the plurality of radiation images. The plurality of radiographic images are a plurality of images in which at least one of the conditions of the radiation source and the imaging conditions is different from each other. The radiographic image processing method according to claim 1.

7. The evaluation value is the standard deviation of the noise value, and the radiographic image processing method according to any one of claims 1 to 6.

8. A construction step of constructing, by machine learning, a learned model that outputs the noise removal image data based on the training image and the noise map, using, as training data, a noise map that is data in which the evaluation value is derived from the pixel value of each pixel of the training image based on relationship data representing the relationship between the pixel value of the radiographic image used as the training image and the evaluation value obtained by evaluating the spread of the noise value, and the noise removal image data obtained by removing noise from the training image.

9. A preprocessing method of the machine learning method according to claim 8, A noise map generation step of generating the noise map based on the relationship data, which is a preprocessing method of the machine learning method.

10. A learned model that causes a processor to execute image processing for removing noise from a radiographic image of an object, wherein in the image processing, based on relationship data representing the relationship between the radiographic image in which the object is irradiated with radiation and the radiation transmitted through the object is imaged, and the evaluation value obtained by evaluating the spread of the pixel value and the noise value, the evaluation value is derived from the pixel value of each pixel of the radiographic image, and the noise map, which is data in which the evaluation value derived for each pixel of the radiographic image is associated, are input, and an output image in which noise is removed from the radiographic image is output.

11. An image acquisition unit that acquires a radiographic image in which an object is irradiated with radiation and the radiation transmitted through the object is imaged; A noise map generation unit that generates a noise map, which is data in which the evaluation value is derived from the pixel value of each pixel of the radiographic image based on relationship data representing the relationship between the pixel value and the evaluation value obtained by evaluating the spread of the noise value, and the evaluation value derived for each pixel of the radiographic image is associated; A processing unit that executes image processing for removing noise from the radiographic image based on the radiographic image and the noise map; Comprising In the image processing, the processing unit inputs the radiation image and the noise map into a pre-trained model constructed by machine learning in advance, and obtains an output image, which is an image with noise removed from the radiation image, from the pre-trained model. The radiation image processing module.

12. The noise map generation unit derives the evaluation value from the average energy of the radiation transmitted through the object and the pixel value of each pixel of the radiation image. The radiation image processing module according to claim 11.

13. An input unit that receives an input of condition information indicating either the conditions of the radiation source or the imaging conditions when irradiating the object with radiation to capture an image of the object; A calculation unit that calculates the average energy based on the condition information; and further includes The condition information includes at least any one of the tube voltage of the radiation source, information about the object, information about the filter provided in the camera used for imaging the object, information about the filter provided in the radiation source, and information about the scintillator provided in the camera used for imaging the object. The radiation image processing module according to claim 12.

14. It further includes a calculation unit that calculates the average energy from the pixel value of each pixel of the radiation image. The radiation image processing module according to claim 12.

15. The image acquisition unit acquires a radiation image of a jig irradiated with radiation and imaged with the radiation transmitted through the jig. In the noise map generation unit, the relationship data is derived from the radiation image of the jig. The radiation image processing module according to claim 11, further including

16. The image acquisition unit acquires a plurality of radiation images captured in a state where the object is not present. The noise map generation unit derives the relationship data from the plurality of radiation images. The plurality of radiation images are a plurality of images in which at least one of the conditions of the radiation source and the imaging conditions are different from each other. The radiation image processing module according to claim 11.

17. The evaluation value is the standard deviation of the noise value. The radiation image processing module according to any one of claims 11 to 16.

18. A training image that is a radiographic image, a noise map generated from the image based on the relationship data, and noise-removed image data that is data with noise removed from the training image are used as training data, and a learned model that outputs the noise-removed image data based on the training image and the noise map is constructed by machine learning. The radiation image processing module according to any one of claims 11 to 17, further comprising a construction unit.

19. A processor, An image acquisition unit that irradiates an object with radiation and acquires a radiation image obtained by imaging the radiation that has passed through the object, Based on relationship data representing the relationship between the pixel value and the evaluation value obtained by evaluating the spread of the noise value, the evaluation value is derived from the pixel value of each pixel of the radiation image, and the evaluation value derived for each pixel of the radiation image is A noise map generation unit that generates data associated with the noise map, and A processing unit that executes image processing for removing noise from the radiation image based on the radiation image and the noise map. In the image processing, the radiation image and the noise map are input to a learned model constructed in advance by machine learning, and an output image that is an image with noise removed from the radiation image is obtained from the learned model. A radiation image processing program that functions as the processing unit.

20. The radiation image processing module according to any one of claims 11 to 18, A radiation source that irradiates the object with radiation, An imaging device that images the radiation that has passed through the object and acquires the radiation image, A radiation image processing system comprising:

21. The imaging device has a line sensor. The radiation image processing system according to claim 20.

22. The imaging device has a two-dimensional sensor. The radiation image processing system according to claim 20.

Images