JP2022012392A - Image processing method, x-ray phase imaging device and creation method of learned model - Google Patents

Abstract

To provide an image processing method which can accurately remove an artifact component while reducing the learning time of a machine learning model when removing the artifact component with the machine learning model.SOLUTION: An image processing method comprises the steps of: acquiring an X-ray radiographic image 11 before contraction; acquiring an X-ray radiographic image 12 after contraction; acquiring an artifact image 14 after contraction by inputting the X-ray radiographic image 12 after contraction to a machine learning model 13; enlarging the artifact image 14 after contraction; and removing a low frequency component by using an enlarged artifact image 15. The machine learning model 13 is learned with the low frequency component image 19 after contraction indicating the low frequency component as a teacher image with respect to the input image created with contraction by adding the low frequency component to the original image and reducing a data capacity.SELECTED DRAWING: Figure 3

Description

The present invention relates to an image processing method, an X-ray phase imaging apparatus, and a method for creating a trained model.

Conventionally, an image processing method for an X-ray photographed image is known (see, for example, Non-Patent Document 1).

The non-patent document 1 discloses an image processing method for an X-ray photographed image. In this image processing method, an X-ray image containing an artifact component generated by X-ray photography is input, and a machine learning model that outputs an X-ray image with the artifact component removed is used from the X-ray image. The artifact component is removed.

J. Chen et al., "Automatic image-domain Moire artifact reduction method in grating-based x-ray interferometry imaging", Physics in Medicine and Biology, 2019, Vol. 64, No. 19.

The shape of the artifact generated in actual X-ray photography changes according to the device configuration. In order to deal with artifacts of all shapes, it is necessary to prepare a large amount of training data, and there is an inconvenience that the amount of data increases and learning takes time.

The present invention has been made to solve the above-mentioned problems, and one object of the present invention is to shorten the learning time of the trained model when the artifact component is removed by the trained model. At the same time, it is an object of the present invention to provide an image processing method capable of accurately removing an artifact component, an X-ray phase imaging device, and a method for creating a trained model.

In order to achieve the above object, the image processing method according to the first aspect of the present invention includes a step of acquiring an X-ray image by performing an X-ray image by an X-ray phase measurement method using a plurality of lattices. , The step of reducing the radiographed image in the form of reducing the data capacity, the step of preparing the trained model, the step of inputting the radiographed image into the trained model and acquiring the artifact image, and the data capacity The trained model adds low-frequency components to the original image, with steps to magnify the artifact image in an increasing manner and to remove low-frequency components from the radiographed image using the magnified artifact image. In addition, a low-frequency component image showing a low-frequency component is learned as a teacher image for an input image created by reducing the data capacity. Note that "reducing an image" is a broad concept that includes not only a process of reducing an image size but also a process of reducing an image size in image processing by acquiring a typical pixel value.

The X-ray phase imaging apparatus according to the second aspect of the present invention includes an X-ray source, a detector for detecting X-rays emitted from the X-ray source, and a plurality of X-ray sources arranged between the X-ray source and the detector. The image processing unit is provided with a grid and an image processing unit, and the image processing unit acquires an X-ray photographed image by performing X-ray photography by an X-ray phase measurement method using a plurality of lattices, and reduces the data capacity. Reduce the X-ray image in the form, input the reduced X-ray image to the trained model, acquire the artifact image, enlarge the artifact image in the form of increasing the data capacity, and use the enlarged artifact image. , The trained model is configured to remove low frequency components from the radiographed image, and the trained model is an input image created by adding low frequency components to the original image and reducing the data capacity. On the other hand, a low-frequency component image showing a low-frequency component is learned as a teacher image.

The method for creating a trained model according to the third aspect of the present invention shows a step of creating an input image by adding a low frequency component to the original image and reducing the data capacity, and a low frequency component. Obtained by performing X-ray photography by the X-ray phase measurement method using multiple grids by performing machine learning based on the step of creating a low-frequency component image as a teacher image and the input image and the teacher image. Create a trained model that inputs a reduced image of the X-ray photographed image and outputs an artifact image.

The image processing method according to the fourth aspect of the present invention has already been learned: a step of acquiring an X-ray photographed image by performing X-ray photography, and a step of reducing the X-ray photographed image in a form of reducing the data capacity. Using the step of preparing the model, the step of inputting the reduced X-ray image to the trained model and acquiring the artifact image, the step of enlarging the artifact image by increasing the data capacity, and the step of enlarging the artifact image. The trained model is provided with a step of removing the low frequency component from the X-ray photographed image, and the trained model is added to the input image created by adding the low frequency component to the original image and reducing the data capacity. On the other hand, a low frequency component image showing a low frequency component is learned as a teacher image.

The image processing method according to the fifth aspect of the present invention includes a step of acquiring an X-ray photographed image, a step of preparing a trained model, and a step of inputting an X-ray photographed image into the trained model and acquiring an artifact image. With the steps of removing the low frequency component from the X-ray image using the artifact image, the trained model has a low frequency with respect to the input image created by adding the low frequency component to the original image. A low-frequency component image showing the components is learned as a teacher image.

The image processing method according to the first aspect of the present invention, the X-ray phase imaging apparatus according to the second aspect, the method for creating a trained model according to the third aspect, and the image processing method according to the fourth aspect are as described above. In addition, the trained model is trained using a low-frequency component image showing a low-frequency component as a teacher image for an input image created by adding a low-frequency component to the original image and reducing the data capacity. ing. As a result, the image size of the trained image can be reduced as compared with the case where the trained model is trained using the unreduced image. As a result, the learning time of the trained model can be shortened. In addition, an artifact image is acquired, the artifact image is enlarged in a form that increases the data capacity, and the enlarged artifact image is used to remove low-frequency components from the radiographed image. Here, the high-frequency component, which is a fine component, is easily erased when the image is reduced or enlarged, while the low-frequency component, which is a gentle component, is difficult to be erased when the image is reduced or enlarged. Therefore, by enlarging the artifact image as described above, it is possible to obtain an enlarged artifact image that accurately shows the low frequency component as the artifact component. Further, as described above, if the low frequency component is removed from the X-ray photographed image using the enlarged artifact image, the artifact component can be obtained by using the enlarged artifact image that accurately shows the low frequency component as the artifact component. It can be removed accurately. As a result, when the artifact component is removed by the trained model, the artifact component can be removed accurately while shortening the learning time of the trained model.
Further, in the image processing method according to the fifth aspect of the present invention, the trained model uses a low-frequency component image showing a low-frequency component as a teacher image for an input image created by adding a low-frequency component to the original image. Is being learned as. As a result, it is possible to create a trained model with a simple image as compared with the case of creating a trained model using an image containing no artifact component (for example, the original image of the input image) as a teacher image. As a result, a trained model can be easily created. Also, the artifact image is used to remove low frequency components from the radiographed image. This makes it possible to remove artifact components while easily creating a machine learning model. Further, when the image is reduced or enlarged and used, the artifact component can be accurately removed while shortening the learning time of the trained model as in the invention according to the first to fourth aspects described above. can.

It is a figure which showed the structure of the X-ray phase imaging apparatus by one Embodiment. It is a flowchart for demonstrating low frequency component removal processing by one Embodiment. It is a figure for demonstrating the low frequency component removal process by one Embodiment. It is a flowchart for demonstrating the method of making a trained model by one Embodiment. It is a figure for demonstrating the method of making a trained model by one Embodiment. It is a figure for demonstrating the machine learning model of the X-ray phase imaging apparatus of one Embodiment. It is a figure for demonstrating the low frequency component removal processing of the X-ray phase imaging apparatus by the modification of one Embodiment.

Hereinafter, embodiments embodying the present invention will be described with reference to the drawings.

First, with reference to FIG. 1, the overall configuration of the X-ray phase imaging apparatus 100 according to the embodiment will be described.

(Structure of image processing method)
As shown in FIG. 1, the X-ray phase imaging device 100 is a device that generates an image of the inside of the subject 200 by using the X-rays that have passed through the subject 200. Specifically, the X-ray phase imaging device 100 is a device that generates an image of the inside of the subject 200 by utilizing the Talbot effect. The X-ray phase imaging apparatus 100 can be used for imaging the inside of an object, for example, in non-destructive inspection applications.

The X-ray phase imaging apparatus 100 includes an X-ray source 1, a first grid 2, a second grid 3, a third grid 4, a detector 5, an image processing unit 6, a control unit 7, and a grid movement. It is equipped with a mechanism 8. In the specification of the present application, the direction from the X-ray source 1 toward the first lattice 2 is the Z2 direction, and the direction opposite to the direction is the Z1 direction. Further, the left-right direction in the plane orthogonal to the Z direction is the X direction, the direction toward the back of the paper surface is the X2 direction, and the direction toward the front side of the paper surface is the X1 direction. Further, the vertical direction in the plane orthogonal to the Z direction is the Y direction, the upward direction is the Y1 direction, and the downward direction is the Y2 direction.

The X-ray source 1 is configured to generate X-rays by applying a high voltage and to irradiate the generated X-rays in the Z2 direction.

The first grid 2 has a plurality of slits 2a arranged in the Y direction with a predetermined period (pitch) d1 and an X-ray phase changing portion 2b. Each of the slits 2a and the X-ray phase changing portion 2b is formed so as to extend linearly. Further, each of the slits 2a and the X-ray phase changing portion 2b is formed so as to extend in parallel. The first grid 2 is a so-called phase grid.

The first grid 2 is arranged between the X-ray source 1 and the second grid 3, and X-rays are emitted from the X-ray source 1. The first grid 2 is provided to form a self-image (not shown) of the first grid 2 by the Talbot effect. When X-rays having coherence pass through the lattice in which the slit is formed, an image of the lattice (self-image) is formed at a position separated from the lattice by a predetermined distance (Talbot distance). This is called the Talbot effect.

The second lattice 3 has a plurality of X-ray transmitting portions 3a and X-ray absorbing portions 3b arranged in the Y direction with a predetermined period (pitch) d2. The X-ray absorbing portion 3b extends along the direction in which the X-ray phase changing portion 2b extends. Each of the X-ray transmitting portion 3a and the X-ray absorbing portion 3b is formed so as to extend linearly. Further, each X-ray transmitting portion 3a and X-ray absorbing portion 3b are formed so as to extend in parallel. The second grid 3 is a so-called absorption grid. The first grid 2 and the second grid 3 are grids having different roles, but the slit 2a and the X-ray transmitting portion 3a each transmit X-rays. Further, the X-ray absorbing unit 3b plays a role of shielding X-rays, and the X-ray phase changing unit 2b changes the phase of X-rays depending on the difference in refractive index from the slit 2a.

The second grid 3 is arranged between the first grid 2 and the detector 5, and is irradiated with X-rays that have passed through the first grid 2. Further, the second grid 3 is arranged at a position separated from the first grid 2 by a Talbot distance. The second grid 3 interferes with the self-image of the first grid 2 to form moire fringes (not shown) on the detection surface of the detector 5.

The third lattice 4 has a plurality of X-ray transmitting portions 4a and X-ray absorbing portions 4b arranged in the Y direction with a predetermined period (pitch) d3. Each of the X-ray transmitting portion 4a and the X-ray absorbing portion 4b is formed so as to extend linearly. Further, each X-ray transmitting portion 4a and X-ray absorbing portion 4b are formed so as to extend in parallel. The third grid 4 is a so-called multi-slit.

The third grid 4 is arranged between the X-ray source 1 and the first grid 2. The third lattice 4 is configured to convert X-rays from the X-ray source 1 into a multi-point light source by using the X-rays that have passed through each X-ray transmitting portion 4a as a line light source. The X-rays emitted from the X-ray source 1 when the pitch of the three grids (first grid 2, the second grid 3, and the third grid 4) and the distance between the grids satisfy certain conditions. It is possible to increase the interferability. As a result, the interference strength can be maintained even if the focal size of the tube of the X-ray source 1 is large.

The detector 5 is configured to detect X-rays, convert the detected X-rays into an electric signal, and read the converted electric signal as an image signal. The detector 5 is, for example, an FPD (Flat Panel Detector). The detector 5 is composed of a plurality of conversion elements (not shown) and pixel electrodes (not shown) arranged on the plurality of conversion elements. The plurality of conversion elements and pixel electrodes are arranged in an array in the X direction and the Y direction at a predetermined period (pixel pitch). Further, the detector 5 is configured to output the acquired image signal to the image processing unit 6.

The image processing unit 6 is configured to generate an X-ray photographed image (X-ray photographed image before reduction) 11 (see FIG. 3) of the subject 200 based on the image signal output from the detector 5. In the example shown in FIG. 3, the subject 200 as a vial is reflected in the X-ray photographed image (X-ray photographed image before reduction) 11. The image processing unit 6 includes, for example, a GPU (Graphics Processing Unit) and a processor such as an FPGA (Field-Programmable Gate Array) configured for image processing.

The control unit 7 is configured to step-move the first grid 2 in the grid plane in a direction orthogonal to the grid direction by the grid movement mechanism 8. In the X-ray phase imaging apparatus 100, a method (striping scanning method) of acquiring an image from a plurality of moire fringes (images) obtained by scanning the first lattice 2 at regular periodic intervals is used. Further, the control unit 7 includes, for example, a processor such as a CPU (Central Processing Unit).

The grid movement mechanism 8 is configured to step-move the first grid 2 in the grid plane (in the XY plane) in a direction orthogonal to the grid direction (Y direction in FIG. 1) based on the signal from the control unit 7. Has been done. Specifically, the grid moving mechanism 8 divides the period d1 of the first grid 2 into n, and moves the first grid 2 in steps by d1 / n. The grid movement mechanism 8 is configured to move the first grid 2 in steps for at least one cycle d1 of the first grid 2. Note that n is a positive integer, for example, 9. Further, the grid moving mechanism 8 includes, for example, a stepping motor, a piezo actuator, and the like.

(Low frequency component removal processing)
Next, the low frequency component removal process by the X-ray phase imaging apparatus 100 of the embodiment will be described with reference to the flowchart of FIG. 2 and FIG. The low frequency component removal process is performed by the image processing unit 6.

First, as shown in FIGS. 2 and 3, first, in step 101, an X-ray photographed image 11 before reduction is acquired by performing an X-ray photograph by an X-ray phase measurement method using a plurality of lattices 2 to 4. Will be done. Here, when the X-ray imaging data is acquired by the fringe scanning method, if there is an error in the step position of the first grid 2 which is the grid to be stepped, the pre-reduction X-ray imaging image 11 generated by the imaging process. In addition, artifacts similar to moire fringes occur. This is called a moire artifact. Since this moire artifact occurs even with a minute step position error of several nm to several tens of nm, it is difficult to suppress the occurrence. The X-ray photographed image 11 before reduction is an example of the "X-ray photographed image" in the claims.

Then, in step 102, the pre-reduction X-ray image 11 is reduced in a manner that reduces the data capacity. That is, in step 102, the reduced X-ray captured image 12 obtained by reducing the reduced pre-reduced X-ray captured image 11 is acquired. Specifically, in step 102, an image reduction process for reducing the number of pixels (resolution) of the pre-reduction X-ray image 11 is performed at a predetermined reduction rate with respect to the pre-reduction X-ray image 11. The reduced and reduced X-ray photographed image 12 is acquired. In the X-ray photographed image 12 after reduction, the number of pixels is reduced at a predetermined reduction rate with respect to the X-ray photographed image 11 before reduction, and the data capacity is reduced. The image reduction process is not particularly limited, but for example, an average pixel method, a bilinear interpolation method, a bicubic interpolation method, or the like can be adopted. The reduced X-ray image 12 is an example of the "X-ray image" in the claims.

Then, in step 103, the reduced X-ray photographed image 12 is input to the pre-learned and prepared machine learning model 13, and the reduced reduced artifact image 14 is acquired as an output image. The details of the machine learning model 13 will be described later. The reduced artifact image 14 is an image showing a low frequency component of the image as an artifact component. In this embodiment, the low frequency component is a shading component of the image. Further, the low frequency component is a component showing a gentle change in the shade of the image. Specifically, the low-frequency component is a moire artifact component that is a striped shade generated due to the misalignment of the first lattice 2. The reduced-reduced artifact image 14 does not contain components other than the moire artifact component such as the subject 200, but contains only the periodic striped moire artifact component. Further, the reduced-reduced artifact image 14 has the same image size as the reduced-reduced X-ray photographed image 12. The machine learning model 13 is an example of a "learned model" in the claims. Further, the reduced artifact image 14 is an example of an "artifact image" within the scope of the claims.

Then, in step 104, the reduced artifact image 14 is enlarged in a manner that increases the data capacity. That is, in step 104, the enlarged artifact image 15 obtained by enlarging the reduced artifact image 14 as an output image to the same image size as the pre-reduced X-ray photographed image 11 is acquired. Specifically, in step 104, an image enlargement process (image interpolation process) for increasing the number of pixels (resolution) of the reduced artifact image 14 is performed to obtain a predetermined enlargement ratio for the reduced artifact image 14. By enlarging with, an enlarged artifact image 15 enlarged to the same image size as the pre-reduction X-ray image 11 is acquired. In the enlarged artifact image 15, the number of pixels is increased by a predetermined enlargement ratio with respect to the reduced artifact image 14. Further, the predetermined enlargement ratio is the reciprocal of the predetermined reduction ratio in step 102. The image enlargement processing is not particularly limited, but for example, a bilinear interpolation method, a bicubic interpolation method, or the like can be adopted. The enlarged artifact image 15 is an example of the "artifact image" in the claims.

Then, in step 105, the magnified artifact image 15 is used to remove the low frequency component from the pre-reduction X-ray photographed image 11. That is, in step 105, the low frequency component is removed from the pre-reduction X-ray image 11 by differentiating the magnified artifact image 15 from the pre-reduction X-ray image 11. Specifically, in step 105, the X-ray photographed image after removing the low frequency component from which the low frequency component as the moire artifact component is removed by performing the process of subtracting the magnified artifact image 15 from the X-ray photographed image 11 before reduction. 16 is acquired.

If the cycle of the moire artifact component in the pre-reduction X-ray image 11 is excessively small, the moire artifact component may be erased during the image reduction process in step 102. On the other hand, the period of the moire artifact component correlates with the period of the moire fringes formed on the detection surface of the detector 5. Therefore, it is preferable to adjust the period of the moire fringes formed on the detection surface of the detector 5 in advance so that the period of the moire artifact component becomes sufficiently large. The period of the moire fringes formed on the detection surface of the detector 5 can be adjusted by adjusting the positions of the grids 2 to 4.

(How to create a trained model)
Next, a method of creating a machine learning model 13 (method of creating a trained model) included in the X-ray phase imaging apparatus 100 of one embodiment will be described with reference to the flowchart of FIG. 4 and FIG. The machine learning model 13 is learned by supervised learning.

As shown in FIGS. 4 and 5, first, in step 111, a pre-reduction low frequency component image 17 showing the low frequency component of the image showing the artifact component is created. Specifically, in step 111, the pre-reduction low frequency component image 17 is created by simulation calculation. Further, the pre-reduction low-frequency component image 17 may be created by extracting from an actual X-ray photographed image (pre-reduction X-ray photographed image 11) including an artifact component. In this embodiment, the low frequency component is a shading component of the image. Further, the low frequency component is a shading component in which the change in the pixel value of the image is gentle. Specifically, the low-frequency component is a moire artifact component that is a striped shade generated due to the misalignment of the first lattice 2. The pre-reduced low-frequency component image 17 and the post-reduced low-frequency component image 19 do not contain any components other than the moire artifact component, and contain only the periodic striped moire artifact component. Further, in step 111, a plurality of unreduced low frequency component images 17 having different at least one of the amplitude and period of the moire artifact component are created.

Then, in step 112, a pre-reduction low-frequency component composite image 18 is created in which the low-frequency component of the image showing the artifact component is added to the original image. Specifically, in step 112, the pre-reduction low-frequency component composite image 18 is created by adding (combining) the pre-reduction low-frequency component image 17 created in step 111 to the original image. Further, in step 112, a plurality of pre-reduction low-frequency component composite images 18 having different at least one of the amplitude and period of the moire artifact component are created. The original image to which the low frequency component is added may be an image that appropriately contains the high frequency component and the low frequency component, and does not have to be an X-ray photographed image. As the original image, for example, an X-ray photographed image, a landscape photograph image (natural image), or the like can be adopted. Note that, for convenience, FIG. 5 shows an image similar to the pre-reduction X-ray photographed image 11 (see FIG. 3) as a pre-reduction low-frequency component composite image 18. Further, the pre-reduction low-frequency component image 17 and the pre-reduction low-frequency component composite image 18 have the same image size.

Then, in step 113, the reduced low-frequency component image 19 after reducing the reduced low-frequency component image 17 in a form of reducing the data capacity is created as a teacher output image (teacher image), and the pre-reduced low-frequency component synthesis is performed. A reduced low-frequency component composite image 20 obtained by reducing the image 18 in a form of reducing the data capacity is created as a teacher input image (input image). Specifically, in step 113, an image reduction process for reducing the number of pixels (resolution) of the low-frequency component image 17 before reduction is performed, so that the low-frequency component image 17 before reduction is reduced at a predetermined reduction ratio. The reduced-reduced low-frequency component image 19 is acquired. In the reduced low-frequency component image 19, the number of pixels is reduced at a predetermined reduction rate with respect to the pre-reduced low-frequency component image 17, and the data capacity is reduced. Similarly, in step 113, an image reduction process for reducing the number of pixels (resolution) of the pre-reduction low-frequency component composite image 18 is performed, so that the pre-reduction low-frequency component composite image 18 is reduced at a predetermined reduction ratio. The reduced-reduced low-frequency component composite image 20 is acquired. In the reduced low-frequency component composite image 20, the number of pixels is reduced at a predetermined reduction rate with respect to the pre-reduced low-frequency component composite image 18, and the data capacity is reduced. The image reduction process is not particularly limited, but for example, an average pixel method, a bilinear interpolation method, a bicubic interpolation method, or the like can be adopted. Further, the reduced low frequency component image 19 and the reduced low frequency component composite image 20 have the same image size. The reduced low-frequency component image 19 is an example of the "low-frequency component image" in the claims.

Further, the reduction ratio at the time of image reduction processing may be any reduction ratio as long as the low frequency component as the moire artifact component is not erased (not crushed). Whether or not the low frequency component is erased (crushed) by the reduction can be determined by using an index. Specifically, by comparing the image before reduction with the image restored by reducing and enlarging it once, the similarity (matching degree) between the image before reduction and the restored image is acquired. .. Then, when the acquired similarity is equal to or less than the threshold value, it can be determined that the low frequency component is erased (crushed) by the reduction. Further, when the acquired similarity exceeds the threshold value, it can be determined that the low frequency component is not erased (not crushed) by the reduction.

Then, in step 114, a plurality of lattices 2 to 4 are formed by performing machine learning based on the teacher input image (reduced low frequency component composite image 20) and the teacher output image (reduced low frequency component image 19). The pre-reduction X-ray image 11 (see FIG. 3) acquired by performing X-ray photography by the X-ray phase measurement method used is reduced, and the reduced post-reduction X-ray image 12 (see FIG. 3) is input and reduced. A trained machine learning model 13 (see FIG. 3) is created that outputs the post-artifact image 14 (see FIG. 3). Specifically, in step 114, the trained machine learning model 13 is constructed by constructing a convolutional neural network in which the reduced X-ray image 12 (see FIG. 3) is input and the reduced artifact image 14 is output. (See Figure 3) is created.

(Machine learning model)
Further, in the present embodiment, as shown in FIG. 6, the pre-reduction X-ray image 11 (see FIG. 3) includes an absorption image 11a, a phase differential image 11b, and a dark field image 11c. The image processing unit 6 (see FIG. 1) is configured to be capable of generating an absorption image 11a, a phase differential image 11b, and a dark field image 11c based on the image signal output from the detector 5 (see FIG. 1). ing. The absorption image 11a is an image of the contrast generated by the difference in absorption of X-rays by the subject 200. Further, the phase differential image 11b is an image of the contrast generated by the change in the phase of the X-ray by the subject 200. The dark field image 11c is an image of the contrast generated by the refraction (scattering) of X-rays due to the fine structure inside the subject 200. In other words, the dark field image 11c is an image of the decrease in visibility due to the subject 200, and the decrease in visibility depends on the degree of scattering of the subject 200. That is, the dark field image 11c is an image of the X-ray scattering of the subject 200.

Here, in the absorption image 11a, the phase differential image 11b, and the dark field image 11c, different moire artifacts are generated. Specifically, in the absorption image 11a, a moire artifact having the same period as the moire fringes formed on the detection surface of the detector 5 is generated. Further, in the phase differential image 11b, a moire artifact having a period of half of the moire fringes formed on the detection surface of the detector 5 is generated. Therefore, in the phase differential image 11b, a moire artifact having a smaller period than that of the absorption image 11a is generated. Further, in the dark field image 11c, the moire fringes having the same period as the moire fringes formed on the detection surface of the detector 5 and the moire fringes having half the period of the moire fringes formed on the detection surface of the detector 5 Moire artifacts that overlap with each other occur. Therefore, the dark field image 11c has the same period as the phase differential image 11b, but unlike the absorption image 11a and the phase differential image 11b, the moire artifacts have different degrees of shading (brightness) between the bright parts and the dark parts. Occur.

Therefore, in the present embodiment, the machine learning model 13 (see FIG. 3) is a machine learning model 13a for the absorption image 11a, a machine learning model 13b for the phase differential image 11b, and a machine learning model for the dark field image 11c. Contains 13c. The machine learning model 13a for the absorption image 11a, the machine learning model 13b for the phase differential image 11b, and the machine learning model 13c for the dark field image 11c have reduced low frequencies containing moire artifact components peculiar to each image. The component image 19 (see FIG. 5) is learned as a teacher output image. Specifically, the machine learning model 13a for the absorption image 11a is trained using the reduced low-frequency component image 19 including the moire artifact component having a larger period than the phase differential image 11b and the dark field image 11c as a teacher output image. ing. Further, the machine learning model 13b for the phase differential image 11b is trained using the reduced low-frequency component image 19 including the moire artifact component whose period is smaller than that of the absorption image 11a as a teacher output image. Further, unlike the absorption image 11a and the phase differential image 11b, the machine learning model 13c for the dark field image 11c contains a moire artifact component in which the degree of shading (brightness) differs between the bright parts and the dark parts, and the reduced low frequency. The component image 19 is learned as a teacher output image. The machine learning model 13a for the absorption image 11a, the machine learning model 13b for the phase differential image 11b, and the machine learning model 13c for the dark field image 11c are each "learning for the absorption image" within the scope of the patent claim. It is an example of "trained model", "trained model for phase differential image", and "trained model for dark field image".

(Effect of this embodiment)
In this embodiment, the following effects can be obtained.

In the present embodiment, as described above, the machine learning model 13 is an input image created by adding a low frequency component to the original image and reducing the data capacity (reduced low frequency component composite image 20). ), A low-frequency component image showing a low-frequency component (reduced low-frequency component image 19) is learned as a teacher image. As a result, the image size of the image to be learned can be reduced as compared with the case where the machine learning model 13 is trained using the unreduced image. As a result, the learning time of the machine learning model 13 can be shortened. Further, the reduced-reduced artifact image 14 is enlarged in a manner that increases the data capacity, and the enlarged reduced-reduced artifact image 14 (enlarged artifact image 15) is used to remove low-frequency components from the pre-reduction X-ray photographed image 11. Here, the high-frequency component, which is a fine component, is easily erased when the image is reduced or enlarged, while the low-frequency component, which is a gentle component, is difficult to be erased when the image is reduced or enlarged. Therefore, if the reduced-reduced artifact image 14 is enlarged as described above, the enlarged artifact image 15 that accurately shows the low-frequency component as the artifact component can be obtained. Further, as described above, if the low frequency component is removed from the pre-reduction X-ray image 11 using the magnified artifact image 15, the magnified artifact image 15 that accurately shows the low frequency component as the artifact component is used. Artifact components can be removed with high accuracy. As a result, when the artifact component is removed by the machine learning model 13, the artifact component can be removed accurately while shortening the learning time of the machine learning model 13.

Further, in the above embodiment, the machine learning model 13 is a low-frequency component image showing a low-frequency component with respect to an input image (reduced low-frequency component composite image 20) created by adding a low-frequency component to the original image. (Low frequency component image 19 after reduction) is learned as a teacher image. As a result, the machine learning model 13 can be created with a simple image as compared with the case where the machine learning model 13 is created using an image containing no artifact component (for example, the original image of the input image) as a teacher image. As a result, the machine learning model 13 can be easily created. Also, the magnified artifact image 15 is used to remove low frequency components. This makes it possible to remove the artifact component while easily creating the machine learning model 13.

Further, in the present embodiment, as described above, the low frequency component is a shading component of the image. As a result, since the low frequency component is a simple shading component, it is possible to prevent the low frequency component as an artifact component from being erased more when the image is reduced or enlarged. As a result, the magnified artifact image 15 that accurately shows the low frequency component as the artifact component can be effectively obtained.

Further, in the present embodiment, as described above, the low frequency component is a moire artifact component which is a striped shade generated due to the misalignment of the grids 2 to 4. As a result, when the moire artifact component is removed by the machine learning model 13, the moire artifact component can be accurately removed while shortening the learning time of the machine learning model 13.

Further, in the present embodiment, as described above, the pre-reduction X-ray image 11 includes the absorption image 11a, the phase differential image 11b, and the dark field image 11c. Further, the machine learning model 13 includes a machine learning model 13a for the absorption image 11a, a machine learning model 13b for the phase differential image 11b, and a machine learning model 13c for the dark field image 11c. As a result, the artifact component can be accurately removed from the absorption image 11a, the phase differential image 11b, or the dark field image 11c.

[Modification example]
It should be noted that the embodiments disclosed this time are exemplary in all respects and are not considered to be restrictive. The scope of the present invention is shown by the scope of claims, not the description of the above-described embodiment, and further includes all modifications (modifications) within the meaning and scope equivalent to the scope of claims.

For example, in the above embodiment, the present invention has been shown as an example in which the present invention is applied to an X-ray phase imaging apparatus, but the present invention is not limited to this. The present invention may be applied to an X-ray imaging device other than the X-ray phase imaging device (for example, a normal X-ray imaging device capable of capturing only an absorption image). Also in this case, the artifact component contained in the X-ray photographed image generated by the ordinary X-ray imaging apparatus can be accurately removed.

Further, in the above embodiment, the present invention has been shown as an example in which the present invention is applied to the removal of a low frequency component as a moire artifact component, but the present invention is not limited to this. The present invention can be applied to an artifact component other than the moire artifact component as long as the low frequency component is the main artifact component. For example, as in the modification shown in FIG. 7, the low-frequency component as an artifact component may be a beam hardening artifact component generated due to radiation hardening of X-rays. Here, when the X-ray absorption rate of the subject 200 (see FIG. 1) is high, the X-ray photographed image 11d before reduction has a gradation-like shade due to the curing of the X-ray quality inside the subject 200. Beam hardening artifacts occur. In the example shown in FIG. 7, the subject 200 as a rectangular aluminum plate is reflected in the pre-reduction X-ray photographed image 11d. Originally, since the inside of the aluminum plate has a uniform structure, the portion of the subject 200 in the image should have a uniform pixel value. However, in the example shown in FIG. 7, a beam hardening artifact having a mountainous pixel value is generated due to the influence of X-ray quality hardening inside the subject 200. This beam hardening artifact can also be removed from the pre-reduction X-ray image 11d by the same method as in the above embodiment. As a result, when the beam hardening artifact component is removed by the machine learning model, the beam hardening artifact component can be removed accurately while shortening the learning time of the machine learning model.

Further, in the above embodiment, an example is shown in which the machine learning model (trained model) includes a machine learning model for an absorption image, a machine learning model for a phase differential image, and a machine learning model for a dark field image. However, the present invention is not limited to this. In the present invention, the trained model includes only one or two of the trained model for the absorption image, the trained model for the phase differential image, and the trained model for the dark field image. You may. Further, the trained model may include one trained model in which any of the absorption image, the phase differential image, and the dark field image can be input.

Further, in the above embodiment, an example in which the third lattice is provided is shown, but the present invention is not limited to this. In the present invention, the third grid may not be provided.

Further, in the above embodiment, an example in which the first lattice is a phase lattice is shown, but the present invention is not limited to this. In the present invention, the first lattice may be an absorption lattice.

Further, in the above embodiment, an example in which the first lattice is step-moved in the lattice plane is shown, but the present invention is not limited to this. In the present invention, any of the plurality of grids may be stepped.

Further, in the above embodiment, an example is shown in which a machine learning model is trained by using a reduced low-frequency component composite image as a teacher input image and a reduced low-frequency component image as a teacher output image, but the present invention is limited to this. do not have. In the present invention, the machine learning model may be trained using the pre-reduction low-frequency component composite image of the above embodiment as a teacher input image and the pre-reduction low-frequency component image of the above embodiment as a teacher output image. In this case, the pre-reduction X-ray image is input to the machine learning model, and the pre-reduction low-frequency component image is acquired as the output image. Then, the low frequency component is removed from the pre-reduction X-ray image by differentiating the pre-reduction low-frequency component image from the pre-reduction X-ray image.

[Aspect]
It will be understood by those skilled in the art that the above-mentioned exemplary embodiments are specific examples of the following embodiments.

(Item 1)
A step of acquiring an X-ray photographed image by performing an X-ray photograph by an X-ray phase measurement method using a plurality of grids, and a step of acquiring an X-ray photographed image.
The step of reducing the X-ray image in a way that reduces the data capacity,
Steps to prepare the trained model and
A step of inputting the reduced X-ray image into the trained model and acquiring an artifact image, and
Steps to magnify the artifact image and
It comprises a step of removing low frequency components from the radiographed image using the magnified artifact image.
In the trained model, a low-frequency component image showing the low-frequency component is trained as a teacher image with respect to an input image created by adding a low-frequency component to the original image and reducing the data capacity. Image processing method.

(Item 2)
The image processing method according to item 1, wherein the low frequency component is a shading component of an image.

(Item 3)
The image processing method according to item 2, wherein the low frequency component is a moire artifact component which is a striped shade generated due to the misalignment of the lattice.

(Item 4)
The image processing method according to item 2 or 3, wherein the low frequency component is a beam hardening artifact component generated due to radiation hardening of X-rays.

(Item 5)
The radiographed image includes at least one of an absorption image, a phase differential image, and a dark field image.
Item 1 to 4, wherein the trained model includes at least one of a trained model for the absorption image, a trained model for the phase differential image, and a trained model for the dark field image. The image processing method according to any one item.

(Item 6)
X-ray source and
A detector that detects X-rays emitted from the X-ray source, and
A plurality of grids arranged between the X-ray source and the detector,
Equipped with an image processing unit
The image processing unit
An X-ray photographed image is acquired by performing X-ray photography by the X-ray phase measurement method using the plurality of lattices.
The X-ray photographed image is reduced in a way that reduces the data capacity,
Input the reduced X-ray image into the trained model, acquire the artifact image, and
Enlarge the artifact image in a way that increases the data capacity,
It is configured to remove low frequency components from the radiographed image using the magnified artifact image.
In the trained model, a low-frequency component image showing the low-frequency component is trained as a teacher image with respect to an input image created by adding a low-frequency component to the original image and reducing the data capacity. X-ray phase imaging device.

(Item 7)
A step to create an input image by adding a low frequency component to the original image and reducing the data capacity.
The step of creating a low-frequency component image showing the low-frequency component as a teacher image, and
By performing machine learning based on the input image and the teacher image, a reduced image obtained by performing X-ray imaging by an X-ray phase measurement method using a plurality of lattices is input. And how to create a trained model that outputs an artifact image.

(Item 8)
The step of acquiring an X-ray photographed image by performing X-ray photography, and
The step of reducing the X-ray image in a way that reduces the data capacity,
Steps to prepare the trained model and
A step of inputting the reduced X-ray image into the trained model and acquiring an artifact image, and
The step of enlarging the artifact image in a way that increases the data capacity,
It comprises a step of removing low frequency components from the radiographed image using the magnified artifact image.
In the trained model, a low-frequency component image showing the low-frequency component is trained as a teacher image with respect to an input image created by adding a low-frequency component to the original image and reducing the data capacity. Image processing method.

(Item 9)
Steps to acquire X-ray images and
Steps to prepare the trained model and
The step of inputting the X-ray photographed image into the trained model and acquiring the artifact image, and
It comprises a step of removing low frequency components from the radiographed image using the artifact image.
The trained model is an image processing method in which a low-frequency component image showing the low-frequency component is trained as a teacher image with respect to an input image created by adding a low-frequency component to the original image.

(Item 10)
Item 9. The image processing according to item 9, wherein the step of acquiring the X-ray photographed image includes a step of acquiring the X-ray photographed image by performing X-ray photography by an X-ray phase measurement method using a plurality of grids. Method.

(Item 11)
The image processing method according to item 9 or 10, wherein the low frequency component is a shading component of an image.

(Item 12)
The image processing method according to item 11, wherein the low-frequency component is a moire artifact component that is a striped shade generated due to a misalignment of the lattice.

(Item 13)
The image processing method according to item 11 or 12, wherein the low frequency component is a beam hardening artifact component generated due to radiation hardening of X-rays.

(Item 14)
The radiographed image includes at least one of an absorption image, a phase differential image, and a dark field image.
Item 9-13, wherein the trained model includes at least one of the trained model for the absorption image, the trained model for the phase differential image, and the trained model for the dark field image. The image processing method according to any one item.

1 X-ray source 2-4 grids (multiple grids)
5 Detector 6 Image processing unit 11, 11d X-ray image before reduction (X-ray image)
11a Absorption image 11b Phase differential image 11c Dark field image 12 X-ray image after reduction (X-ray image)
13 Machine learning model (learned model)
13a Machine learning model for absorption image (trained model for absorption image)
13b Machine learning model for phase differential image (trained model for phase differential image)
13c Machine learning model for darkfield image (trained model for darkfield image)
14 Artifact image after reduction (artifact image)
15 Enlarged artifact image (artifact image)
17 Low-frequency component image before reduction 18 Low-frequency component composite image before reduction 19 Low-frequency component image after reduction (teacher image, low-frequency component image)
20 Reduced low frequency component composite image (input image)
100 X-ray phase imaging device

Claims (14)

A step of acquiring an X-ray photographed image by performing an X-ray photograph by an X-ray phase measurement method using a plurality of grids, and a step of acquiring an X-ray photographed image.
The step of reducing the X-ray image in a way that reduces the data capacity,
Steps to prepare the trained model and
A step of inputting the reduced X-ray image into the trained model and acquiring an artifact image, and
The step of enlarging the artifact image in a way that increases the data capacity,
It comprises a step of removing low frequency components from the radiographed image using the magnified artifact image.
In the trained model, a low-frequency component image showing the low-frequency component is trained as a teacher image with respect to an input image created by adding a low-frequency component to the original image and reducing the data capacity. Image processing method. The image processing method according to claim 1, wherein the low frequency component is a shading component of an image. The image processing method according to claim 2, wherein the low-frequency component is a moire artifact component which is a striped shade generated due to the misalignment of the lattice. The image processing method according to claim 2 or 3, wherein the low frequency component is a beam hardening artifact component generated due to radiation hardening of X-rays. The radiographed image includes at least one of an absorption image, a phase differential image, and a dark field image.
The trained model includes at least one of the trained model for the absorption image, the trained model for the phase differential image, and the trained model for the dark field image, claims 1 to 4. The image processing method according to any one of the above items. X-ray source and
A detector that detects X-rays emitted from the X-ray source, and
A plurality of grids arranged between the X-ray source and the detector,
Equipped with an image processing unit
The image processing unit
An X-ray photographed image is acquired by performing X-ray photography by the X-ray phase measurement method using the plurality of lattices.
The X-ray photographed image is reduced in a way that reduces the data capacity,
Input the reduced X-ray image into the trained model, acquire the artifact image, and
Enlarge the artifact image in a way that increases the data capacity,
It is configured to remove low frequency components from the radiographed image using the magnified artifact image.
In the trained model, a low-frequency component image showing the low-frequency component is trained as a teacher image with respect to an input image created by adding a low-frequency component to the original image and reducing the data capacity. X-ray phase imaging device. A step to create an input image by adding a low frequency component to the original image and reducing the data capacity.
The step of creating a low-frequency component image showing the low-frequency component as a teacher image, and
By performing machine learning based on the input image and the teacher image, a reduced image obtained by performing X-ray imaging by an X-ray phase measurement method using a plurality of lattices is input. And how to create a trained model that outputs an artifact image. The step of acquiring an X-ray photographed image by performing X-ray photography, and
The step of reducing the X-ray image in a way that reduces the data capacity,
Steps to prepare the trained model and
A step of inputting the reduced X-ray image into the trained model and acquiring an artifact image, and
The step of enlarging the artifact image in a way that increases the data capacity,
It comprises a step of removing low frequency components from the radiographed image using the magnified artifact image.
In the trained model, a low-frequency component image showing the low-frequency component is trained as a teacher image with respect to an input image created by adding a low-frequency component to the original image and reducing the data capacity. Image processing method. Steps to acquire X-ray images and
Steps to prepare the trained model and
The step of inputting the X-ray photographed image into the trained model and acquiring the artifact image, and
It comprises a step of removing low frequency components from the radiographed image using the artifact image.
The trained model is an image processing method in which a low-frequency component image showing the low-frequency component is trained as a teacher image with respect to an input image created by adding a low-frequency component to the original image. The image according to claim 9, wherein the step of acquiring the X-ray photographed image includes a step of acquiring the X-ray photographed image by performing X-ray photography by an X-ray phase measurement method using a plurality of lattices. Processing method. The image processing method according to claim 9 or 10, wherein the low frequency component is a shading component of an image. The image processing method according to claim 11, wherein the low-frequency component is a moire artifact component that is a striped shade generated due to a misalignment of the lattice. The image processing method according to claim 11 or 12, wherein the low frequency component is a beam hardening artifact component generated due to radiation hardening of X-rays. The radiographed image includes at least one of an absorption image, a phase differential image, and a dark field image.
The trained model includes at least one of the trained model for the absorption image, the trained model for the phase differential image, and the trained model for the dark field image, claims 9 to 13. The image processing method according to any one of the above items.

Images