TWI885189B - Radiographic image processing method, learning completion model, radiographic image processing module, radiographic image processing program, radiographic image processing system, and machine learning method - Google Patents

Abstract

The present invention provides a radiographic image processing method, a learning model, a radiographic image processing module, a radiographic image processing program, and a radiographic image processing system that can effectively remove noise from radiographic images. The control device 20 includes: an input unit 201 that receives input of condition information indicating either an operation condition of an X-ray source when irradiating X-rays and photographing an object F, or an imaging condition when photographing the object F; a calculation unit 202 that calculates the average energy of the X-rays penetrating the object F based on the condition information; and a screening unit 203 that screens candidates for the learning model 206 from a plurality of learning models 206 that are constructed by machine learning using image data in advance based on the average energy.

Description

Radiographic image processing method, learning completion model, radiographic image processing module, radiographic image processing program, radiographic image processing system, and machine learning method

One embodiment of the present invention relates to a radiation image processing method, a learning completion model, a radiation image processing module, a radiation image processing program, a radiation image processing system, and a machine learning method.

It is previously known that a method of removing noise using a learned model constructed by machine learning such as deep learning is used for image data (for example, refer to the following patent document 1). According to this method, since noise from image data is automatically removed, the object can be observed with high accuracy.

[Prior Art Literature]

[Patent Literature]

[Patent Document 1] Japanese Patent Publication No. 2019-91393

In the above-mentioned previous methods, when the object is a radiographic image generated by allowing radiation such as X-rays to penetrate an object, noise removal is sometimes insufficient. For example, there is a tendency that the relationship between image brightness and noise is easily changed depending on the conditions of the radiation generating source such as the X-ray source, the type of filter used, etc., and noise cannot be effectively removed.

Therefore, one embodiment of the present invention is completed in view of the above-mentioned subject, which is to provide a radiation image processing method, a learning completion model, a radiation image processing module, a radiation image processing program, a radiation image processing system and a machine learning method that can effectively remove the noise of radiation images.

A radiation image processing method of one embodiment includes the following steps: inputting condition information indicating the conditions of the radiation source or the imaging conditions when irradiating radiation and photographing an object; calculating the average energy associated with the radiation penetrating the object based on the condition information; and selecting a candidate for a learning model from a plurality of learning models that are constructed by machine learning using image data in advance based on the average energy.

Alternatively, another embodiment of the learning completion model is a learning completion model used in the above-mentioned radiation image processing method, which is constructed by machine learning using image data, so that the processor performs image processing to remove noise from the radiation image of the object.

Alternatively, a radiation image processing module of another embodiment includes: an input unit that receives input of condition information indicating the conditions of the radiation source or the imaging conditions when the object is irradiated with radiation; a calculation unit that calculates the average energy related to the radiation penetrating the object based on the condition information; and a screening unit that screens candidates for the learning model from a plurality of learning models that are constructed in advance by machine learning using image data based on the average energy.

Alternatively, a radiation image processing program of another embodiment enables a processor to function as an input unit, a calculation unit, and a screening unit. The input unit receives input of condition information indicating the conditions of the radiation source or the imaging conditions when the object is irradiated with radiation and photographed; the calculation unit calculates the average energy related to the radiation penetrating the object based on the condition information; and the screening unit screens candidates for the learning model from a plurality of learning models that are constructed by machine learning using image data in advance based on the average energy.

Alternatively, a radiation image processing system of another embodiment includes: the above-mentioned radiation image processing module; a source that irradiates radiation to an object; and a camera that captures radiation that penetrates the object and obtains a radiation image.

Alternatively, a machine learning method of another embodiment includes a step of constructing a learning model by machine learning, wherein the learning model uses a training image of a radiographic image of an object corresponding to the average energy of radiation penetrating the object, which is calculated based on condition information representing either a condition of a radiation source or an imaging condition when the object is irradiated with radiation and photographed, as training data, and outputs image data after noise removal based on the training image.

According to one or the other of the above aspects, the average energy of radiation penetrating the object is calculated based on the conditions of the radiation source or the imaging conditions when the radiation image of the object is obtained. Then, based on the average energy, the candidate of the learned model used for noise removal is screened from the pre-constructed learned models. In this way, since the learned model corresponding to the average energy of the radiation of the imaged object is used for noise removal, noise removal corresponding to the relationship between the brightness and noise of the radiation image can be realized. As a result, the noise of the radiation image can be effectively removed.

Depending on the implementation form, the noise of the radiation image of the object can be effectively removed.

1: Image acquisition device

10: X-ray detection camera (photographic device)

11a,11b: Flasher

12a,12b: Line scanning camera

13: Sensor control unit

14a,14b:Amplifier

15a,15b:AD converter

16a,16b: Correction circuit

17a,17b: Output interface

18: Amplifier control unit

19: Light filter

20,20A: Control device (radiation image processing module)

30: Display device

40: Input device

50: X-ray irradiator (radiation source)

51: Light filter

60: Belt conveyor (transport mechanism)

101:CPU

102: RAM

103:ROM

104: Communication module

106: Input and output module

201:Input Department

202,202A: Calculation Department

203,203A: Screening Department

204: Selection Department

205: Processing Department

206: Learning completion model

207: Measurement Department

F: Object

G ₁ ,G ₂ ,G _3, G _T :Characteristic curve diagram

P1: Components

P2: Foreign objects

R1, R2: Image area

S1~S11,S101~S108: Steps

TD: Transport direction

Figure 1 is a schematic diagram of the image acquisition device 1 in an implementation form.

FIG2 is a block diagram showing an example of the hardware structure of the control device 20 of FIG1.

FIG3 is a block diagram showing the functional structure of the control device 20 of FIG1.

FIG. 4 is a diagram showing an example of teaching data, i.e., image data, used to construct the learning completion model 206 of FIG. 3 .

FIG. 5 is a flow chart showing the steps for producing teaching data, i.e., image data, for constructing the learning completion model 206 of FIG. 3 .

FIG6 is a diagram showing an example of an X-ray transmission image of the analysis target of the selection section 204 of FIG3 .

FIG. 7 is a diagram showing an example of a thickness-brightness characteristic curve obtained by the selection unit 204 of FIG. 3 .

FIG8 is a diagram showing an example of a characteristic curve diagram of brightness-SNR obtained by the selection unit 204 of FIG3.

Figures 9(a) and (b) are diagrams for explaining the selection function of the learning completion model based on the image characteristics obtained by the selection unit 204 in Figure 3.

FIG. 10 is a diagram showing an example of an X-ray transmission image used to evaluate the resolution obtained by the selection unit 204 of FIG. 3 .

FIG. 11 is a three-dimensional diagram showing an example of the structure of a jig for evaluating the brightness-noise ratio obtained by the selection unit 204 of FIG. 3 .

Figure 12 shows an X-ray penetration image obtained after noise removal processing using the fixture in Figure 11 as the object.

FIG. 13 is a flow chart showing the steps of observation processing using the image acquisition device 1.

FIG. 14 is a diagram showing an example of an X-ray transmission image obtained by the image acquisition device 1 before and after noise removal processing.

FIG. 15 is a diagram showing an example of an X-ray transmission image obtained by the image acquisition device 1 before and after noise removal processing.

FIG. 16 is a block diagram showing the functional structure of the control device 20A of the variation of the present disclosure.

FIG. 17 is a flow chart showing the steps of observation processing of the image acquisition device 1 using a variation.

The following is a detailed description of the implementation of the present invention with reference to the attached drawings. In addition, the same symbols are used for the same elements or elements with the same functions in the description, and repeated descriptions are omitted.

FIG1 is a configuration diagram of the image acquisition device 1, which is a radiation image processing system of the present embodiment. As shown in FIG1, the image acquisition device 1 is a device that irradiates an object F conveyed along a conveying direction TD with X-rays (radiation) and acquires an X-ray penetration image (radiation image) obtained by photographing the object F based on the X-rays that penetrate the object F. The image acquisition device 1 uses the X-ray penetration image to perform foreign body inspection, weight inspection, and cargo inspection on the object F. Examples of the uses include food inspection, carry-on baggage inspection, substrate inspection, battery inspection, and material inspection. The image acquisition device 1 is composed of: a belt conveyor (transport mechanism) 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 various inputs. Furthermore, the radiation image of the embodiment of the present invention is not limited to X-ray images, but also includes images generated by electromagnetic radiation other than X-rays such as gamma rays.

The belt conveyor 60 has a belt portion for carrying the object F, and by moving the belt portion along the conveying direction TD, the object F is conveyed along the conveying direction TD at a specific conveying speed. The conveying speed of the object F is, for example, 48 m/min. The belt conveyor 60 can change the conveying speed to, for example, 24 m/min or 96 m/min as needed. In addition, the belt conveyor 60 can appropriately change the height position of the belt portion to change the distance between the X-ray irradiator 50 and the object F. Furthermore, as the object F conveyed by the belt conveyor 60, for example, there can be cited various items such as edible meat, fish and shellfish, agricultural products, snacks, rubber products such as tires, resource materials such as resin products, metal products, minerals, waste, and electronic parts or electronic substrates. The X-ray irradiator 50 is a device that irradiates (outputs) X-rays toward the object F as an X-ray source. The X-ray irradiator 50 is a point light source that diffusely irradiates X-rays in a specific angle range along a certain irradiation direction. The X-ray irradiator 50 irradiates the belt conveyor 60 in the irradiation direction of X-rays, and the diffused X-rays are arranged above the belt conveyor 60 at a specific distance from the belt conveyor 60 in the entire width direction (the direction intersecting the conveying direction TD) of the object F. In addition, the X-ray irradiator 50 sets a specific segmented range in the longitudinal direction of the object F (the direction parallel to the conveying direction TD) as the irradiation range, and the object F is conveyed in the conveying direction TD by the belt conveyor 60, thereby irradiating the entire longitudinal direction of the object F with X-rays. The control device 20 sets the tube voltage and tube current of the X-ray irradiator 50, and irradiates the belt conveyor 60 with X-rays of specific energy and radiation amount corresponding to the set tube voltage and tube current. In addition, a filter 51 is installed near the belt conveyor 60 side of the X-ray irradiator 50 to allow a specific wavelength band of X-rays to pass through.

The X-ray detection camera 10 detects X-rays that penetrate the object F among the X-rays irradiated to 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 configured with two sets of dual-line X-ray cameras for detecting X-rays. In the image acquisition device 1 of this embodiment, X-ray penetration images are generated based on the X-rays detected by the respective lines (the first line and the second line) of the dual-line X-ray cameras. Moreover, by performing averaging processing or addition processing on the two generated X-ray penetration images, a clear (brighter) image can be obtained with a smaller amount of X-rays compared to the case where an X-ray penetration image is generated based on X-rays detected by one line.

The X-ray detection camera 10 includes a filter 19, scintillators 11a, 11b, line scan cameras 12a, 12b, a sensor control unit 13, amplifiers 14a, 14b, AD converters 15a, 15b, correction circuits 16a, 16b, output interfaces 17a, 17b, and an amplifier control unit 18. The scintillator 11a, line scan camera 12a, amplifier 14a, AD converter 15a, correction circuit 16a, and output interface 17a are electrically connected to form a first circuit. In addition, the scintillator 11b, line scan camera 12b, amplifier 14b, AD converter 15b, correction circuit 16b, and output interface 17b are electrically connected to form a second circuit. The line scanning camera 12a of the first line and the line scanning camera 12b of the second line are arranged in a row along the conveying direction TD. In addition, the following describes the common structure of the first line and the second line, taking the structure of the first line as a representative.

The scintillator 11a is fixed to the line scan camera 12a, and converts the X-rays that penetrate the object F into scintillation light. The scintillator 11a outputs the scintillation light to the line scan camera 12a. The filter 19 allows a specific wavelength band of the X-rays to pass through the scintillator 11a.

The line scan camera 12a detects the flash light from the scintillator 11a, converts it into electric charge, and outputs it to the amplifier 14a as a detection signal (electrical signal). The line scan camera 12a has a plurality of linear sensors arranged in parallel along a direction intersecting the conveying direction TD. The linear sensor is, for example, a CCD (Charge Coupled Device) image sensor or a CMOS (Complementary Metal-Oxide Semiconductor) image sensor, which includes a plurality of photodiodes.

The sensor control unit 13 controls the line scan cameras 12a and 12b to repeatedly shoot at a specific detection cycle so that the line scan cameras 12a and 12b can shoot X-rays that penetrate the same area of the object F. The specific detection cycle can be set as a common cycle 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: source object distance)), and the distance between the X-ray irradiator 50 and the line scan cameras 12a and 12b (FDD (Focus Detector Distance: source detector distance)). In addition, the specific cycle can also be set individually based on the pixel width of the photodiode in the direction orthogonal to the pixel arrangement direction of the linear sensor of each line scan camera 12a and 12b. In this case, the detection cycle offset (delay time) between the specific line scan cameras 12a and 12b can be set according to the distance between the line scan cameras 12a and 12b, the speed of the belt conveyor 60, the distance (FOD) between the X-ray irradiator 50 and the object F on the belt conveyor 60, and the distance (FDD) between the X-ray irradiator 50 and the line scan cameras 12a and 12b, and the individual cycles can be set respectively. The amplifier 14a amplifies the detection signal with a specific set amplification factor and generates 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 specific imaging conditions.

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 specific corrections such as signal amplification on the digital signal and outputs the corrected digital signal to the output interface 17a. The output interface 17a outputs the digital signal to the outside of the X-ray detection camera 10. In FIG1 , the AD converter, the correction circuit, or the output interface exist separately, but they can also be integrated into one.

The control device 20 is, for example, a computer such as a PC (Personal Computer). The control device 20 generates an X-ray transmission image based on the digital signal (amplified signal) output from the X-ray detection camera 10 (more specifically, the output interfaces 17a, 17b). The control device 20 generates an X-ray transmission image by averaging or adding the two digital signals output from the output interfaces 17a, 17b. The generated X-ray transmission image is output to the display device 30 after applying the noise removal process described later, and is displayed by the display device 30. In addition, the control device 20 controls the X-ray irradiator 50, the amplifier control unit 18, and the sensor control unit 13. Furthermore, the control device 20 of this embodiment is an independent device disposed outside the X-ray detection camera 10, but it can also be integrated inside the X-ray detection camera 10.

FIG2 shows the hardware structure of the control device 20. As shown in FIG2, the control device 20 physically includes a CPU (Central Processing Unit) 101 as a processor, a RAM (Random Access Memory) 102 or a ROM (Read Only Memory) 103 as a recording medium, a communication module 104, and an input/output module 106, etc., each of which is electrically connected. Furthermore, the control device 20 may also include a display, a keyboard, a mouse, a touch panel display, etc. as an input device 40 and a display device 30, and may also include a data recording device such as a hard disk drive and a semiconductor memory. In addition, the control device 20 may also include a plurality of computers.

FIG3 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, a screening unit 203, a selection unit 204, and a processing unit 205. Each functional unit of the control device 20 shown in FIG3 is realized by reading a program (radiation image processing program of this embodiment) on hardware such as the CPU 101 and the RAM 102, and operating the communication module 104 and the input/output module 106 under the control of the CPU 101, and reading and writing data from the RAM 102. The CPU 101 of the control device 20 executes the computer program to make the control device 20 function as each functional unit of FIG. 3, and sequentially executes the processing corresponding to the radiation image processing method described later. Furthermore, the CPU can be a single hardware, or it can be installed in a programmable logic device such as FPGA like a software processor. Regarding RAM or ROM, it can be a single hardware, or it can be built into a programmable logic device such as FPGA. All kinds of data required for the execution of the computer program and all kinds of data generated by the execution of the computer program are stored in built-in memories such as ROM 103 and RAM 102, or storage media such as hard disk drives.

Furthermore, in the control device 20, a plurality of learning models 206 for performing noise removal processing on X-ray transmission images are read by the CPU 101 and stored in advance in the CPU 101. The plurality of learning models 206 are learning models generated by machine learning that are pre-constructed using image data as teaching data. Machine learning includes teaching learning, deep learning (deep learning), or reinforcement learning, neural network learning, and the like. In this embodiment, as an example of a deep learning algorithm, a two-dimensional convolutional neural network described in the paper "Beyond Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising" by Kai Zhang et al. is used. A plurality of learning completed models 206 can be generated by an external computer and downloaded to the control device 20, or can be generated within the control device 20.

FIG. 4 shows an example of teaching data, i.e., image data, used for learning to construct the completed model 206. As teaching data, X-ray penetration images with various thicknesses, various materials, and various resolutions as the photographic object can be used. The example shown in FIG. 4 is an example of an X-ray penetration image generated with chicken as the object. The image data can use X-ray penetration images generated with multiple objects as the object using the image acquisition device 1, or image data generated by simulation calculation. Regarding the X-ray penetration image, it can also be obtained using a device different from the image acquisition device 1. In addition, the X-ray penetration image and the image data generated by simulation calculation can be used in combination. The plurality of learning completed models 206 are image data obtained by using penetrating X-rays of different average energies as the object, and the noise distribution is pre-constructed using known image data. The average energy of the X-rays in the image data is pre-set to different values by setting the operating conditions of the X-ray irradiator (radiation source) 50 of the image acquisition device 1, or the imaging conditions of the image acquisition device 1, or by setting the operating conditions or imaging conditions of the X-ray irradiator 50 during simulation calculation (the setting method of the average energy due to the operating conditions or imaging conditions will be described later). That is, the plurality of learned models 206 are constructed by machine learning using, as training data, an X-ray image corresponding to the average energy of the X-rays penetrating the object F calculated based on the condition information such as the operation conditions of the X-ray irradiator (radiation source) 50 when shooting the X-ray penetration image of the object F, or the shooting conditions of the X-ray detection camera 10 (construction step). For example, in this embodiment, the plurality of learned models 206 are constructed using a plurality of image data with the average energy of a plurality of frames (e.g., 20,000 frames) set to 10keV, 20keV, 30keV, ..., and 10keV scale values.

FIG. 5 shows a flow chart of the steps for producing teaching data, i.e., image data, for learning to construct the completed model 206.

The teaching data, i.e., image data (also referred to as teaching image data), is produced by a computer according to the following steps. First, an image of a structure having a specific structure (structure image) is produced (step S101). For example, an image of a structure having a specific structure can be produced by simulation calculation. In addition, an X-ray image of a structure such as a chart having a specific structure can be obtained to produce a structure image. Next, for a pixel selected from a plurality of pixels constituting the structure image, the standard deviation of the pixel value, i.e., the Σ value, is calculated (step S102). Then, based on the Σ value obtained in step S102, a normal distribution (Poisson distribution) representing the noise distribution is set (step S103). In this way, by setting the normal distribution based on the Σ value, teaching data of various noise conditions can be generated. Then, according to the normal distribution set based on the Σ value in step S103, the noise value set randomly is calculated (step S104). Furthermore, by adding the noise value obtained in step S104 to the pixel value of one pixel, the pixel value constituting the teaching data, i.e., the image data, is generated (step S105). For each of the plurality of pixels constituting the structure image, the processing from step S102 to step S105 is performed (step S106), and the teaching image data as the teaching data is generated (step S107). Furthermore, when teaching image data is further needed, it is determined that the processing of steps S101 to S107 is performed on other structural body images (step S108) to generate other teaching image data as teaching data. Furthermore, the other structural body images can be images of structures with the same structure or images of structures with different structures.

Furthermore, a large number of teaching data, i.e., image data, used for learning to construct the completed model 206 need to be prepared. Also, images with less noise are better for structural body images, and ideally, images without noise are better. Therefore, if structural body images are generated by simulation calculation, a larger number of noise-free images can be generated, so it is effective to generate structural body images by simulation calculation.

Next, return to Figure 3 to explain the details of the functions of each functional unit of the control device 20.

The input unit 201 receives input of condition information indicating the operating conditions of the X-ray irradiator (radiation source) 50 or the imaging conditions of the X-ray detection camera 10 when capturing the X-ray penetration image of the object F from the user of the image acquisition device 1. The operating conditions may include all or part of the tube voltage, target angle, target material, etc. As condition information indicating the imaging conditions, the material and thickness of the filters 51, 19 (filters provided by the camera for photographing the object or filters provided by the source) disposed 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 window of the X-ray detection camera 10, etc. can be cited. The type of material, the material and thickness of the scintillator 11a, 11b of the X-ray detection camera 10, the X-ray detection camera information (for example, the gain setting value, the circuit noise value, the saturation charge, the conversion coefficient value (number of electrons/count), the line frequency (Hz) or the line speed (m/min) of the camera), the information of the object, etc. All or part of the information. The input unit 201 can accept the input of the condition information as a direct input of information such as a numerical value, or as a selective input of information such as a numerical value pre-set in the internal memory. The input unit 201 accepts the input of the above-mentioned condition information from the user, but can also obtain part of the condition information (tube voltage, etc.) based on the detection result of the control state executed by the control device 20.

The calculation unit 202 calculates the average energy value of the X-rays (radiation) that are transmitted through the object F using the image acquisition device 1 and 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 using, for example, a well-known approximate formula such as Tucker, based on information such as tube voltage, target angle, target material, material and thickness of the filter and the presence or absence of the window material, and the material and thickness of the scintillators 11a and 11b of the X-ray detection camera 10 included in the condition information. Then, the calculation unit 202 further calculates the spectrum intensity integral value and the photon number integral value based on the X-ray spectrum, and divides the spectrum intensity integral value by the photon number integral value to calculate the average energy value of the X-ray.

The calculation method using the well-known Tucker approximation is described. For example, when the target is specified as tungsten and the target angle is specified as 25°, the calculation unit 202 can determine Em: the kinetic energy of the electron-target collision, T: the kinetic energy of the electron in the target, A: the proportional 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: the function of Z and T that changes slowly, C: the Thomson-Whiddington constant, θ: the target angle, and c: the speed of light in a vacuum. Furthermore, the calculation unit 202 can calculate the irradiated X-ray spectrum by calculating the following formula (1) based on these calculations.

Furthermore, Em can be determined based on information about tube voltage, A, ρ, and μ(E) can be determined based on information about target material, and θ can be determined based on information about target angle.

Next, the calculation unit 202 can use the X-ray attenuation formula of the following formula (2) to calculate the energy spectrum of the X-rays that pass through the filter and the object F and are absorbed by the scintillator.

[Number 2] I = I _O e ^{- μx} (2)

Here, μ is the attenuation coefficient of the target object, filter, scintillator, etc., and x is the thickness of the target object, filter, scintillator, etc. μ can be determined based on information about the materials of the target object, filter, and scintillator, and x can be determined based on information about the thickness of the target object, filter, and scintillator. The X-ray photon number spectrum is obtained by dividing the X-ray energy spectrum by the energy of each X-ray. The calculation unit 202 calculates the average energy of the X-ray by dividing the integral value of the energy intensity by the integral value of the photon number using the following formula (3).

Average energy E = spectral intensity integral value / photon number integral value…(3)

Through the above calculation process, the calculation unit 202 calculates the average energy of the X-ray. Furthermore, for the calculation of the X-ray spectrum, the well-known approximate formulas such as Kramers or Birch can also be used.

The screening unit 203 screens candidates for the learning model from the plurality of pre-constructed learning models 206 based on the average energy value calculated by the calculation unit 202. That is, the screening unit 203 compares the calculated average energy value with the average energy value of X-rays in the image data used to construct the plurality of learning models 206, and screens the plurality of learning models 206 constructed by the image data having similar average energy values as candidates. More specifically, when the average energy value calculated by the calculation unit 202 is 53keV, the screening unit 203 sets the learning completion model 206 constructed by the image data of the average energy values of 40keV, 50keV, and 60keV whose difference from the value does not reach a specific critical value (e.g., 15keV) as a candidate for the learning completion model.

The selection unit 204 selects the final learning model 206 for noise removal processing of the X-ray transmission image of the object F from the candidates filtered by the screening unit 203. Specifically, the selection unit 204 obtains the X-ray transmission image taken by irradiating X-rays with the jig as the object in the image acquisition device 1, and selects the final learning model 206 based on the image characteristics of the X-ray transmission image. At this time, the selection unit 204 analyzes energy characteristics, noise characteristics, or resolution characteristics as image characteristics of the X-ray transmission image, and selects the learning model 206 based on the analysis results.

More specifically, the selection unit 204 obtains an X-ray transmission image for a flat plate-shaped member of a known thickness and material as a jig and a known relationship between the average energy of X-rays and the X-ray transmission rate, compares the brightness of the X-ray image that penetrates the jig with the brightness of the X-ray image that penetrates the air, and calculates the transmission rate of X-rays at one point (or the average of multiple points) of the jig. For example, when the brightness of the X-ray image that penetrates the jig is 5,550 and the brightness of the X-ray image that penetrates the air is 15,000, the transmission rate is calculated to be 37%. Then, the selection unit 204 specifies the average energy of the penetrating X-rays (e.g., 50 keV) estimated based on the transmission rate of 37% as the energy characteristic of the X-ray transmission image of the jig. The selection unit 204 selects a learning completion model 206 constructed by image data with average energy that is closest to a specific average energy value.

In addition, the selection unit 204 can also analyze the characteristics of multiple points of the jig with thickness or material changes as the energy characteristics of the X-ray transmission image of the jig. Figure 6 is a diagram showing an example of an X-ray transmission image of the analysis object of the selection unit 204. Figure 6 is an X-ray transmission image of a jig with a shape with a step-like change in thickness. The selection unit 204 selects multiple measurement regions (ROI: Region Of Interest) with different thicknesses from such an X-ray transmission image, analyzes the average brightness of each of the multiple measurement regions, and obtains a thickness-brightness characteristic curve as an energy characteristic. Figure 7 shows an example of a thickness-brightness characteristic curve obtained by the selection unit 204.

Furthermore, the selection unit 204 similarly obtains the thickness-brightness characteristic curve graph with the image data used for constructing the learned model 206 screened by the screening unit 203, and selects the learned model 206 constructed by the image data having the characteristics closest to the characteristic curve graph obtained with the jig as the object as the final learned model 206. However, the image characteristics of the image data used for constructing the learned model 206 may also refer to those calculated in advance outside the control device 20. In this way, by setting a plurality of measurement areas, the best learned model in noise removal of the X-ray transmission image of the object F can be selected. In particular, differences in X-ray spectra or filter effects when measuring X-ray penetration images can be estimated with high accuracy.

Furthermore, the selection unit 204 can analyze the brightness value and noise of each of the plurality of measurement areas as the noise characteristics of the X-ray transmission image of the fixture, and obtain the characteristic curve diagram of the brightness-noise ratio as the noise characteristics. That is, the selection unit 204 selects a plurality of measurement areas ROIs of different thicknesses or materials from the X-ray transmission image, analyzes the standard deviation of the brightness values of the plurality of measurement areas ROIs and the average value of the brightness values, and obtains the characteristic curve diagram of the brightness-SNR (SNR ratio) as the noise characteristics. At this time, the selection unit 204 calculates the SNR of each measurement area ROI by SNR=(average value of brightness value) ÷ (standard deviation of brightness value). FIG. 8 shows an example of the characteristic curve diagram of brightness-SNR obtained by the selection unit 204. Then, the selection unit 204 selects the learning completion model 206 constructed by the image data having the noise characteristics closest to the obtained characteristic curve as the final learning completion model 206.

Here, the selection unit 204 can also obtain a characteristic curve diagram of noise calculated based on the standard deviation of the brightness value instead of the above-mentioned brightness-SNR characteristic curve diagram as the noise characteristic. By using such a brightness-noise characteristic curve diagram, the dominant noise factor (shot noise, readout noise, etc.) can be specified for each signal amount detected by the X-ray detection camera 10 according to the slope of the curve diagram of the area of each signal amount, and the learning completion model 206 is selected based on the specified result.

FIG9 is a diagram for explaining the selection function of the learning model based on the image characteristics performed by the selection unit 204. In FIG9, part (a) shows the characteristic curve graphs _G1 , _G2 , and _G3 of brightness-SNR of each image data used for the construction of a plurality of learning models 206, and part (b) shows the characteristic curve graph _GT of brightness-SNR of the X-ray transmission image of the captured jig in addition to the characteristic curve graphs _G1 , _G2 , and _G3 . When such characteristic curve graphs _G1 , _G2 , _G3 , and _GT are used as the object, the selection unit 204 functions in a manner of selecting the learning model 206 constructed by the image data of the characteristic curve graph _{G2 having the characteristics closest to the characteristic curve graph GT .} During the selection, the selection unit 204 calculates the error of the SNR of each brightness value at a certain interval between each characteristic graph G ₁ , G ₂ , G ₃ and the characteristic graph _GT , and calculates the root mean square error (RMSE) of the errors, and selects the learned model 206 corresponding to the characteristic graph G ₁ , G ₂ , G ₃ with the smallest root mean square error. In addition, the selection unit 204 can also select the learned model 206 in the same manner when selecting using energy characteristics.

The selection unit 204 can also select the learned model 206 based on the characteristics of the image after applying multiple learned models to perform noise removal processing on the X-ray transmission image of the fixture.

For example, the selection unit 204 uses an X-ray transmission image of a jig with graphs of various resolutions, applies a plurality of learning models 206 to the image, and evaluates the resulting image after noise removal. Then, the selection unit 204 selects the learning model 206 for the image with the smallest change in resolution before and after noise removal. FIG. 10 shows an example of an X-ray transmission image used for resolution evaluation. In the X-ray transmission image, a graph with a step-wise change in resolution along one direction is set as an imaging object. The resolution of the X-ray transmission image can be measured using MTF (Modulation Transfer Function) or CTF (Contrast Transfer Function).

In addition to the evaluation of the change in resolution described above, the selection unit 204 can also evaluate the brightness-noise ratio characteristics of the image after noise removal, and select the learning model 206 for generating the image with the highest characteristic. FIG11 shows an example of the structure of a jig used for the evaluation of the brightness-noise ratio. For example, as a jig, a member P1 with a stepwise change in thickness along one direction can be used, in which foreign objects P2 of various materials and sizes are scattered. FIG12 shows an X-ray transmission image obtained after noise removal processing using the jig of FIG11 as the object. The selection unit 204 selects the image region R1 including the image of the foreign body P2 and the image region R2 near the region R1 not including the image of the foreign body P2 in the X-ray transmission image, and calculates the minimum value L _MIN of the brightness of the image region R1, the average value L _AVE of the brightness of the image region R2, and the standard deviation L _SD of the brightness of the image region R2. Then, the selection unit 204 uses the following formula: CNR=(L _AVE -L _MIN )/L _SD to calculate the brightness-noise ratio CNR. Furthermore, the selection unit 204 calculates the brightness-noise ratio CNR for each of the X-ray transmission images after applying a plurality of learning completion models 206, and selects the learning completion model 206 used for generating the X-ray transmission image with the highest brightness-noise ratio CNR.

Alternatively, the selection unit 204 may also perform calculation using the following formula based on the average value L _{AVE —} R1 of the brightness of the image region R1 , the average value L _{AVE — R2} of the brightness of the image region R2 , and the standard deviation L _SD of the brightness of the image region R2 .

CNR=(L _{AVE_R1} -L _{MIN_R2} )/L _SD

The processing unit 205 generates an output image by applying the learned model 206 selected by the selection unit 204 to the X-ray transmission image obtained with the object F as the object and performing image processing for removing noise. Then, the processing unit 205 outputs the generated output image to the display device 30, etc.

Next, the sequence of observation and processing of the X-ray transmission image of the object F using the image acquisition device 1 of the present embodiment, that is, the flow of the radiation image processing method of the present embodiment, is described. FIG. 13 is a flow chart showing the sequence of observation and processing performed by the image acquisition device 1.

First, the control device 20 receives input of condition information indicating the operating conditions of the X-ray irradiator 50 or the imaging conditions performed by the X-ray detection camera 10 from the operator (user) of the image acquisition device 1 (step S1). Then, the control device 20 calculates the average energy value of the X-rays detected by the X-ray detection camera 10 based on the condition information (step S2).

Furthermore, the control device 20 specifies the value of the average energy of the X-rays for the image data constructed by the learning model 206 stored in the control device 20 (step S3). Thereafter, the specification of the value of the average energy of the X-rays is repeated for all the learning models 206 stored in the control device 20 (step S4).

Next, the control device 20 compares the calculated average energy values of the X-rays to screen out a plurality of candidates for the learning completion model 206 (step S5). Furthermore, the X-ray penetration image of the jig is obtained by setting the jig in the image acquisition device 1 and photographing the jig (step S6).

Afterwards, the image characteristics of the X-ray penetration image of the fixture (the average energy value of the X-ray, the thickness-brightness characteristics, the brightness-noise ratio characteristics, the brightness-noise characteristics, the resolution change characteristics, etc.) are obtained by the control device 20 (step S7). Then, the control device 20 selects the final learning completion model 206 based on the obtained image characteristics (step S8).

Furthermore, by setting the object F in the image acquisition device 1 and photographing the object F, an X-ray transmission image of the object F is acquired (step S9). Then, by using the control device 20 to apply the finally selected learning model 206 to the X-ray transmission image of the object F, noise removal processing is performed on the X-ray transmission image (step S10). Finally, the output image of the X-ray transmission image after the noise removal processing is output to the display device 30 by the control device 20 (step S11).

According to the image acquisition device 1 described above, the average energy of the X-rays penetrating the object F is calculated based on the motion conditions of the X-ray generation source when acquiring the X-ray penetrating image of the object F or the imaging conditions of the X-ray penetrating image. Then, based on the average energy, the candidate of the learning model 206 for noise removal is screened from the pre-constructed learning model 206. In this way, since the learning model 206 corresponding to the average energy of the X-rays of the photographed object is used for noise removal, noise removal corresponding to the relationship between the brightness and noise of the X-ray penetrating image can be realized. As a result, the noise of the X-ray penetrating image can be effectively removed, for example, the foreign body detection performance can be improved. In particular, the noise pattern of X-ray transmission images varies depending on the tube voltage, filter, scintillator, X-ray detection camera conditions (gain setting value, circuit noise value, saturation charge, conversion coefficient value (e-/count), camera line frequency), and object. Therefore, in the case of using machine learning to achieve noise removal, it is necessary to prepare multiple learning models learned under various conditions in advance. Previously, it has not been possible to select a learning model that matches the noise pattern from multiple learning models in accordance with the conditions when measuring X-ray transmission images. According to this embodiment, by selecting the learning model 206 corresponding to the average energy of the X-ray of the imaging object, the selection of a learning model that always conforms to the state of noise is achieved.

Generally speaking, X-ray penetration images contain noise caused by X-ray generation. Consideration In order to improve the SN ratio of X-ray penetration images, the X-ray dose is increased, but in this case, if the X-ray dose is increased, the exposure dose of the sensor increases, which has the problem of shortening the life of the sensor and the life of the X-ray generation source, and it is difficult to take into account both the improvement of the SN ratio and the longevity. In this embodiment, there is no need to increase the X-ray dose, so the improvement of the SN ratio and the longevity can be taken into account.

Furthermore, the control device 20 of the present embodiment has a function of performing image processing, which uses the selected learning model 206 to remove noise from the X-ray transmission image of the object F. With such a function, noise removal corresponding to the relationship between the brightness and noise of the X-ray transmission image can be achieved, and the noise of the X-ray transmission image can be effectively removed.

In addition, the control device 20 of the present embodiment has the following function: by comparing the average energy value of the X-ray calculated based on the selection information with the average energy value specified in the image data used to construct the learning model 206, the candidate for the learning model is screened. With such a function, noise removal corresponding to the relationship between the brightness and noise of the X-ray penetration image can be achieved reliably.

Furthermore, the control device 20 of this embodiment has a function of selecting a learned model 206 from candidates based on the image characteristics of the X-ray transmission image of the fixture. With such a function, the best learned model 206 can be selected for noise removal of the X-ray transmission image of the object F. As a result, noise removal corresponding to the relationship between the brightness and noise of the X-ray transmission image can be more accurately achieved.

Figures 14 and 15 show examples of X-ray penetration images obtained by the image acquisition device 1 before and after noise removal processing. Figures 14 and 15 respectively show images of cheese with foreign matter such as metal and glass as the object and images of chicken with bones of various sizes remaining as the object, with the image before noise processing being shown on the left and the image after noise processing being shown on the right. As can be seen, according to this embodiment, noise removal is effectively performed on various objects.

The various embodiments of the present invention have been described above, but the present invention is not limited to the above embodiments and can be modified within the scope of the gist of each patent application, or can be applied to other embodiments.

For example, the X-ray detection camera 10 is described as a dual-line X-ray camera, but is not limited to this, and may also be a single-line X-ray camera, a dual-energy X-ray camera, a TDI (Time Delay Integration) scanning X-ray camera, a multi-line X-ray camera having more than two lines, a two-dimensional X-ray camera, an X-ray flat panel sensor, X-ray I.I, a direct conversion X-ray camera that does not use a scintillator (a-Se, Si, CdTe, CdZnTe, TlBr, PbI2, etc.), or a camera that uses an optical lens achieved by lens coupling to observe a scintillator. Furthermore, the X-ray detection camera 10 may also be a radioactive tube or a point sensor that is sensitive to radioactive rays.

Furthermore, the control device 20 of the above-mentioned embodiment selects the candidate of the learning completion model 206 based on the value of the average energy of the X-ray calculated according to the condition information, but it can also have the function of responding to the performance degradation of the X-ray detection camera 10, the output change or performance degradation of the X-ray irradiator 50, as in the control device 20A of the variation shown below.

Furthermore, the image acquisition device 1 is not limited to the above-mentioned embodiment, and may be a radiation image processing system such as a CT (Computed Tomography) device that takes images while the object F is stationary. Furthermore, it may be a radiation image processing system that takes images while the object F is rotated.

FIG16 is a block diagram showing the functional configuration of the control device 20A of the variation. The control device 20A is different from the control device 20 of the above-mentioned embodiment in that it has a measuring unit 207 and the functions of the calculation unit 202A and the screening unit 203A.

In the control device 20, the performance degradation of the X-ray detection camera 10 and the output variation or performance degradation of the X-ray irradiator 50 are set to none, and the relationship between the brightness and noise of the X-ray penetration image can be estimated based on the average energy of the X-rays, and the learning completion model 206 is screened. In contrast, in the control device 20A of this variation, the following functions are provided: considering the performance degradation of the X-ray detection camera 10, the output variation of the X-ray irradiator 50, or its performance degradation, the X-ray conversion coefficient is calculated, and the learning completion model 206 is screened based on the X-ray conversion coefficient. The X-ray conversion coefficient is a parameter that represents the efficiency of the X-ray from being converted into visible light by the scintillator to being converted into electrons (electrical signals) by the camera sensor.

Generally speaking, when the average energy of X-rays is set to E [keV], the luminescence of the scintillator is set to EM [photon/keV], the coupling efficiency of the sensor is set to C, and the quantum efficiency of the sensor is set to QE, the following formula can be used: _FT = E × EM × C × QE

Calculate the X-ray conversion coefficient _FT . In addition, by using the X-ray conversion coefficient _FT , the number of X-ray photons _NP and the camera readout noise Nr _, the following formula is used: SNR = _FTNP /{( _FTNP ⁺ _Nr2 ) ^1/2 }

The signal-to-noise ratio (SNR) of the X-ray transmission image is obtained, and the relationship between the brightness and noise of the X-ray transmission image after the performance degradation of the camera is considered can be estimated based on the X-ray conversion coefficient _FT .

The measuring unit 207 of the control device 20A has a function of measuring a decrease in the amount of light emitted EM as a performance degradation of the scintillators 11a and 11b, a decrease in the quantum efficiency QE of the sensor as a performance degradation of the line scan cameras 12a and 12b, and a change in the average energy E as an output variation and performance degradation of the X-ray irradiator 50. For example, the measuring unit 207 measures a decrease in the amount of light emitted between the state of the scintillators 11a and 11b without performance degradation (new state) and the current state of the scintillators 11a and 11b, and estimates the current amount of light emitted EM based on the decrease. Furthermore, the measuring unit 207 measures the brightness reduction between the state of the line scan cameras 12a and 12b without performance degradation (new state) and the current line scan cameras 12a and 12b, and estimates the current quantum efficiency QE based on the reduction. Furthermore, the measuring unit 207 estimates the current average energy E based on the change in average energy between the state of the X-ray irradiator 50 without performance degradation (new state) and the current X-ray irradiator 50. The average energy E can be obtained based on the imaging data of a flat plate component whose thickness and material are known and the relationship between the average energy of the X-ray and the X-ray transmittance is known, or can be obtained based on the imaging data of multiple points of a jig whose thickness or material changes, etc.

The calculation unit 202A of the control device 20A calculates the X-ray conversion coefficient _FT using the calculated average energy E of the X-rays and the emission amount EM and quantum efficiency QE estimated by the measurement unit 207. The screening unit 203 of the control device 20A has a function of screening candidates for the learning completion model 206 by comparing the calculated X-ray conversion coefficient _FT with the X-ray conversion coefficient _FT of the image data used for construction of the learning completion model 206.

Furthermore, the control device 20 of the above-mentioned embodiment selects the learned model based on the characteristics of the image obtained by photographing the jig after screening the candidates of the learned model, but it is also possible to perform noise removal processing on the X-ray penetration image of the object without photographing the jig. FIG. 17 is a flow chart showing the sequence of observation processing performed by the image acquisition device 1 of the variation. In this way, the processing of steps S6 to S8 in FIG. 13 can also be omitted, and the learned model based on the average energy screening can be used to perform noise removal processing.

In the above-mentioned embodiment, it is preferable to further include a step of performing image processing for removing noise from a radiation image of an object using a candidate. In the above-mentioned embodiment, it is preferable to further include a processing unit for performing image processing for removing noise from a radiation image of an object using a candidate. In this way, noise removal corresponding to the relationship between brightness and noise of a radiation image can be achieved, and noise of a radiation image can be effectively removed.

Furthermore, in the step of filtering, it is preferred to filter the candidates by comparing the average energy with the average energy specified based on the image data. Furthermore, the filtering unit preferably filters the candidates by comparing the average energy with the average energy specified based on the image data. In this case, the learned model is filtered by comparing with the average energy specified based on the image data used to construct the learned model, so that noise removal corresponding to the relationship between the brightness and noise of the radiation image can be reliably achieved.

Furthermore, the condition information preferably includes at least one of the tube voltage of the generating source, the filter information of the camera used to photograph the object, the filter information of the generating source, the strobe information of the camera, the distance between the generating source and the imaging device, the information related to the X-ray detection camera used to photograph the object, and the information related to the object. In this case, since the average energy of the radiation penetrating the object can be calculated with high accuracy, noise removal corresponding to the relationship between the brightness and noise of the radiation image can be performed.

Furthermore, it is preferable to further include the following steps: irradiating the jig with radiation, photographing the jig, and obtaining a radiation image, and selecting a learned model from candidates based on the image characteristics of the radiation image. Furthermore, it is preferable to further include a selection unit, which irradiates the jig with radiation, photographs the jig, and obtains a radiation image, and selects a learned model from candidates based on the image characteristics of the radiation image. According to the above structure, since the learned model is selected based on the image characteristics of the radiation image obtained by photographing the actual jig, the best learned model can be selected for noise removal of the radiation image of the object. As a result, noise removal corresponding to the relationship between the brightness and noise of the radiation image can be more reliably achieved.

[Industrial availability]

The implementation form is to use a radiation image processing method, a learning completion model, a radiation image processing module, a radiation image processing program, and a radiation image processing system, which can effectively remove noise from radiation images.

20: Control device (radiation image processing module) 201: Input unit 202: Calculation unit 203: Screening unit 204: Selection unit 205: Processing unit 206: Learning completion model

Claims (18)

A radiographic image processing method comprises the following steps: Inputting condition information representing either the condition of the radiation source or the imaging condition when irradiating radiation and photographing an object; Based on the condition information, calculating the average energy associated with the radiation penetrating the object; Based on the average energy, selecting a candidate for a learned model from a plurality of learned models constructed by machine learning using image data in advance for each of the plurality of average energies; and Performing image processing for removing noise from the radiographic image of the object using the candidate. A radiographic image processing method as claimed in claim 1, wherein in the aforementioned screening step, the aforementioned candidates are screened by comparing the aforementioned average energy with an average energy specified based on the aforementioned image data. A radiation image processing method as claimed in claim 1, wherein the aforementioned conditional information includes at least: the tube voltage of the aforementioned generating source, information of a filter possessed by a camera used for photographing the aforementioned object, information of a filter possessed by the aforementioned generating source, information of a strobe possessed by the aforementioned camera, the distance between the aforementioned generating source and an imaging device, information related to an X-ray detection camera used for photographing the aforementioned object, and any one of information related to the aforementioned object. A radiation image processing method as claimed in claim 2, wherein the aforementioned conditional information includes at least: the tube voltage of the aforementioned generating source, information of a filter possessed by a camera used for photographing the aforementioned object, information of a filter possessed by the aforementioned generating source, information of a strobe possessed by the aforementioned camera, the distance between the aforementioned generating source and an imaging device, information related to an X-ray detection camera used for photographing the aforementioned object, and any one of information related to the aforementioned object. The radiographic image processing method of any one of claims 1 to 4 further comprises the following steps: irradiating the fixture with radiation and photographing the fixture to obtain a radiographic image, and selecting a learning completion model from the aforementioned candidates based on the image characteristics of the radiographic image. A radiographic image processing method as claimed in any one of claims 1 to 4, wherein the machine learning is deep learning. A radiographic image processing method as claimed in claim 5, wherein the machine learning is deep learning. A learning completion model is a program that enables a computer to execute a radiation image processing method of any one of claim items 1 to 7, and is constructed by machine learning using image data, so that a processor of the aforementioned computer executes image processing for removing noise from the radiation image of the aforementioned object. A radiation image processing module comprises: an input unit that receives condition information indicating the conditions of the radiation source or the imaging conditions when irradiating radiation and photographing an object; a calculation unit that calculates the average energy associated with the radiation that penetrates the object based on the condition information; a screening unit that screens candidates for a learning model from a plurality of learning models that are constructed by machine learning using image data for each of the plurality of average energies based on the average energy; and a processing unit that performs image processing for removing noise from the radiation image of the object using the candidate. A radiation image processing module as claimed in claim 9, wherein the aforementioned candidate is screened by comparing the aforementioned average energy with an average energy specified based on the aforementioned image data. A radiation image processing module as claimed in claim 9, wherein the aforementioned conditional information includes at least: the tube voltage of the aforementioned generating source, information of a filter possessed by a camera used for photographing the aforementioned object, information of a filter possessed by the aforementioned generating source, information of a strobe possessed by the aforementioned camera, the distance between the aforementioned generating source and an imaging device, information related to an X-ray detection camera used for photographing the aforementioned object, and any one of information related to the aforementioned object. A radiation image processing module as claimed in claim 10, wherein the aforementioned conditional information includes at least: the tube voltage of the aforementioned generating source, information of a filter possessed by a camera used for photographing the aforementioned object, information of a filter possessed by the aforementioned generating source, information of a strobe possessed by the aforementioned camera, the distance between the aforementioned generating source and an imaging device, information related to an X-ray detection camera used for photographing the aforementioned object, and any one of information related to the aforementioned object. A radiation image processing module as claimed in any one of claims 9 to 12, further comprising a selection unit, which irradiates radiation and photographs the fixture to obtain a radiation image, and selects a learning completion model from the aforementioned candidates based on the image characteristics of the radiation image. A radiographic image processing module as claimed in any one of claims 9 to 12, wherein the machine learning is deep learning. A radiographic image processing module as claimed in claim 13, wherein the aforementioned machine learning is deep learning. A radiographic image processing program, which enables a processor to function as an input unit, a calculation unit, a screening unit, and a processing unit. The input unit receives condition information indicating the conditions of the source of the aforementioned radiation or any one of the imaging conditions when irradiating radiation and photographing an object; The calculation unit calculates the average energy associated with the aforementioned radiation penetrating the aforementioned object based on the aforementioned condition information; The screening unit screens candidates for a learning model from among a plurality of learning models constructed for each of a plurality of average energies by machine learning using image data in advance based on the aforementioned average energy; The processing unit performs image processing for removing noise from the radiographic image of the aforementioned object using the aforementioned candidate. A radiation image processing system, comprising: a radiation image processing module of any one of claims 9 to 15; the aforementioned generating source, which irradiates the aforementioned object with radiation; and an imaging device, which captures the radiation that penetrates the aforementioned object and obtains the aforementioned radiation image. A machine learning method includes a step of constructing a learning model by machine learning, wherein the learning model uses a radiation image of the object corresponding to the average energy calculated based on condition information representing the condition of the source of the radiation or the imaging condition when the object is irradiated with radiation, and the average energy related to the radiation penetrating the object, i.e., a training image, as training data, and outputs image data after noise is removed based on the training image; The training image has: a known noise distribution based on the average energy.