Information processing device, information processing method, and program

Abstract

To provide an information processing device, an information processing method, and a program capable of accurately correcting a defective pixel when imaging a physical phantom.SOLUTION: An information processing device executes processing for correcting a defective pixel of a radiation image acquired by radiographing a physical phantom in which a test pattern is formed, and includes a processor. The processor detects a region including the test pattern by inputting an image based on the radiation image to a first machine learned model, obtains an edge direction of the test pattern in the region, and corrects a pixel to be corrected using a normal pixel other than the defective pixel in the region present in the edge direction from the pixel to be corrected with the defective pixel in the region as the pixel to be corrected.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The technology of the present disclosure relates to an information processing apparatus, an information processing method, and a program.

Background Art

[0002] In a radiation imaging system that uses radiation to image a subject, a radiation detector such as an FPD (Flat Panel Detector) is used. In the radiation detector, a plurality of pixels that generate and accumulate signal charges according to the incident radiation dose are arranged in a two-dimensional array.

[0003] There may be defective pixels in the plurality of pixels provided in the radiation detector. Since defective pixels cannot obtain appropriate signal charges, it is necessary to perform correction by interpolation processing or the like using the pixel values of surrounding normal pixels.

[0004] Also, in maintenance management, in order for a medical physicist to evaluate the sharpness of a radiation detector, etc., radiation imaging may be performed with a physical phantom (for example, a resolution chart) as the imaging target. In order to accurately perform sharpness evaluation etc. based on the radiation image obtained by imaging the physical phantom, it is necessary to correct defective pixels. However, since the physical phantom includes a high-contrast and high-frequency test pattern, even if defect correction is performed, correction residuals may be visually recognized as artifacts.

[0005] Patent Document 1 discloses a correction method when there is a line defect on a high-frequency test pattern. Specifically, in Patent Document 1, it is proposed to perform regression analysis using a plurality of normal pixels around a defective pixel and correct the defective pixel based on the regression curve obtained as a result of the regression analysis.

[0006] In Patent Document 2, it is proposed to correct a defective pixel using the average value of the pixel values of a pair of normal pixels that exist point-symmetrically with respect to the defective pixel and have the smallest difference value among them.

Prior Art Documents

[0007] [Patent Document 1] Japanese Patent Publication No. 2020-170959 [Patent Document 2] Japanese Patent Publication No. 2002-197450 [Overview of the project] [Problems that the invention aims to solve]

[0008] The defective pixel correction methods described in Patent Documents 1 and 2 both correct the defective pixel using normal pixels located around it. However, the normal pixels used for correction are not necessarily included in the region of the physical phantom that has a test pattern (hereinafter referred to as the test pattern region). If a defective pixel included in the test pattern region is corrected using normal pixels not included in the test pattern region, the correction accuracy will decrease.

[0009] Furthermore, since the physical phantom is attached to any position on the detection surface by a medical physicist, the position and angle of the test pattern area differ each time the physical phantom is attached. For this reason, conventional techniques cannot correct defective pixels using only normal pixels within the test pattern area, and sufficient correction accuracy cannot be obtained.

[0010] Furthermore, in the correction method described in Patent Document 1, as shown in Figure 18, when a thin line with a width of about 1 pixel intersects a line defect at an angle, interpolation is performed using only the horizontally normal pixels adjacent to the line defect. As a result, the correction accuracy is insufficient and correction residuals occur. In such cases, it is necessary to perform interpolation using diagonally oriented normal pixels.

[0011] The present invention aims to provide an information processing device, an information processing method, and a program that enable accurate correction of defective pixels when photographing a physical phantom. [Means for solving the problem]

[0012] To achieve the above objective, the information processing device of the present disclosure is an information processing device that performs processing to correct defective pixels in a radiographic image obtained by radiographic imaging of a physical phantom on which a test pattern has been formed, and comprises a processor, the processor inputs an image based on the radiographic image to a first machine learning model to detect a region containing the test pattern, determines the edge direction of the test pattern within the region, designates the defective pixels within the region as pixels to be corrected, and corrects the pixels to be corrected using normal pixels other than the defective pixels within the region that are located in the edge direction from the pixels to be corrected.

[0013] The first machine learning model is preferably a deep learning model that detects the above-mentioned regions by performing segmentation.

[0014] The physical phantom is a rectangular resolution chart, and it is preferable that the processor determines the edge direction based on a straight line representing the boundary of the region detected by the Hough transform.

[0015] The processor preferably determines the edge direction by inputting the image containing the above region into a second machine learning model.

[0016] The physical phantom is a radial resolution chart, and it is preferable that the processor determines the edge direction based on the relative positional relationship between the pixel to be corrected and the center coordinates of the region.

[0017] The processor preferably corrects the target pixel using a plurality of normal pixels, which are pre-configured according to the edge direction, from among the normal pixels within a mask set with the target pixel at the center.

[0018] The processor preferably corrects the target pixel by assigning weights to multiple normal pixels according to their distance from the target pixel.

[0019] The processor preferably corrects the pixel to be corrected by assigning a weight according to the difference angle, which is the angle formed by the direction of the normal pixel from the pixel to be corrected and the edge direction, and the distance from the pixel to be corrected to each normal pixel within the mask set centered on the pixel to be corrected.

[0020] The information processing method of the present disclosure is an information processing method for performing a process of correcting defective pixels in a radiation image obtained by radiographing a physical phantom on which a test pattern is formed. The method includes detecting a region including the test pattern by inputting an image based on the radiation image into a first pre-trained machine learning model, obtaining an edge direction of the test pattern within the region, setting defective pixels within the region as pixels to be corrected, and correcting the pixels to be corrected using normal pixels other than the defective pixels in the region that exist in the edge direction from the pixels to be corrected.

[0021] The program of the present disclosure is a program that causes a computer to execute a process of correcting defective pixels in a radiation image obtained by radiographing a physical phantom on which a test pattern is formed. The process includes detecting a region including the test pattern by inputting an image based on the radiation image into a first pre-trained machine learning model, obtaining an edge direction of the test pattern within the region, setting defective pixels within the region as pixels to be corrected, and causing the computer to execute a process of correcting the pixels to be corrected using normal pixels other than the defective pixels in the region that exist in the edge direction from the pixels to be corrected.

[0022] According to the technology of the present disclosure, it is possible to provide an information processing apparatus, an information processing method, and a program that can accurately correct defective pixels when photographing a physical phantom.

Brief Description of the Drawings

[0023] [Figure 1] It is a diagram showing a configuration example of a radiation imaging system. [Figure 2] It is a diagram showing an example of a physical phantom. [Figure 3]It is a diagram showing an example of a physical phantom attached to the detection surface of a cassette holder. [Figure 4] It is a diagram schematically showing an example of processing by a detection unit. [Figure 5] It is a diagram for explaining angle information representing an edge direction. [Figure 6] It is a diagram showing an example of a plurality of masks used for correcting point defects. [Figure 7] It is a diagram showing another example of a plurality of masks used for correcting point defects. [Figure 8] It is a diagram showing an example of a plurality of masks used for correcting line defects. [Figure 9] It is a diagram showing another example of a plurality of masks used for correcting line defects. [Figure 10] It is a diagram for explaining a process of excluding pixels outside a test pattern area from a mask. [Figure 11] It is a flowchart showing an example of a correction process flow by a correction unit. [Figure 12] It is a diagram for explaining a point defect correction process by a correction unit according to a modification example. [Figure 13] It is a diagram for explaining a line defect correction process by a correction unit according to a modification example. [Figure 14] It is a flowchart showing an example of a correction process flow by a correction unit according to a modification example. [Figure 15] It is a diagram showing a physical phantom according to a modification example. ​​​​​​​​​​​​​​​​​

[0025] [Embodiment] Figure 1 shows an example configuration of the radiography system 2. The radiography system 2 includes a radiation generator 3, a radiation tube 4, an FPD 5, and an information processing device 6.

[0026] The radiation generator 3 generates radiation R by applying a high-voltage pulse to the radiation tube 4 in response to the user operating an exposure switch (not shown). For example, radiation R is X-rays. The radiation R generated by the radiation tube 4 is irradiated onto the subject H. A portion of the radiation R passes through the subject H and reaches the FPD 5.

[0027] The FPD5 is detachably housed in the cassette holder 7. Radiation R that has passed through the object H passes through the detection surface 7A of the cassette holder 7 and enters the FPD5. A scatter removal grid 8 can also be detachably attached to the cassette holder 7. The scatter removal grid 8 is inserted into the detection surface 7A side of the FPD5. When the scatter removal grid 8 is attached to the cassette holder 7, radiation R that enters the detection surface 7A enters the FPD5 via the scatter removal grid 8. The scattered radiation is removed as the radiation R passes through the scatter removal grid 8.

[0028] The FPD5 has a pixel array in which multiple pixels are arranged in a two-dimensional array, each generating and accumulating a signal charge corresponding to the incident dose of radiation R. Each pixel contains a photoelectric conversion element. The photoelectric conversion element converts radiation R, which has been converted into visible light by a phosphor, into a signal charge and accumulates it. The FPD5 generates a radiation image corresponding to the signal charge of each pixel and transmits the generated radiation image to the information processing device 6 wirelessly or via a wired connection. Note that the FPD5 is not limited to an indirect type of radiation detector that first converts radiation R into visible light and then converts that visible light into a signal charge, but may also be a direct type of radiation detector that directly converts the irradiation of radiation R into a signal charge.

[0029] The information processing device 6 includes a control unit 10, a display 11, an operation unit 12, a storage unit 13, and an image processing unit 14. The information processing device 6 performs processing based on the radiation image received from the FPD 5.

[0030] The control unit 10 comprises one or more processors (not shown) and implements various functions by executing programs 15 stored in the memory unit 13. The memory unit 13 is composed of, for example, ROM (Read Only Memory), RAM (Random Access Memory), etc. The memory unit 13 stores radiation images received by the control unit 10 from the FPD 5, images after image processing by the image processing unit 14, and various data used by the image processing unit 14 for image processing.

[0031] Furthermore, the memory unit 13 stores defective pixel data 16. The defective pixel data 16 is information representing the location, type, etc., of defective pixels in the pixel array of the FPD 5. The types of defective pixels include point defects and line defects. Point defects are isolated defective pixels. Line defects are defective pixels that are continuous in a linear fashion. Line defects occur in the row or column direction of the pixel array. The defective pixel data 16 is acquired by calibration radiography, which is performed in addition to normal radiography. Calibration is performed at the time of product shipment, installation, and during periodic maintenance.

[0032] The image processing unit 14 performs image processing on the radiographic image received by the control unit 10 from the FPD 5. In this disclosure, in order to evaluate the sharpness of the FPD 5, defective pixel correction is performed on the radiographic image obtained by a medical physicist performing radiography on a physical phantom (e.g., a resolution chart) as the target. The image processing unit 14 may have a function to evaluate the sharpness based on the radiographic image after defective pixel correction.

[0033] The image processing unit 14 has an acquisition unit 20, a detection unit 21, and a correction unit 22 as its functional configuration. As will be described in detail later, the acquisition unit 20 acquires the radiation image received by the control unit 10 from the FPD 5. The detection unit 21 detects the area and edge direction of the physical phantom that appears in the radiation image. The correction unit 22 performs defective pixel correction based on the defective pixel data 16 and the detected area and edge direction.

[0034] These functional configurations may be realized by the processor of the control unit 10 executing processing based on the program 15. Alternatively, these functional configurations may be realized by one or more processors in the image processing unit 14 executing processing based on the program 15 read from the storage unit 13.

[0035] The processors of the control unit 10 and the image processing unit 14 are composed of, for example, a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). Each part of the image processing unit 14 may be composed of an integrated circuit or the like, as long as it performs a similar function.

[0036] The display 11 displays the radiation image received by the control unit 10 from the FPD 5, the image processed by the image processing unit 14, etc. The operation unit 12 allows input of instructions to the control unit 10, FPD 5, etc., and accepts input of instructions to the FPD 5 via a user interface (not shown).

[0037] Figure 2 shows an example of a physical phantom. The physical phantom 30 shown in Figure 2 is a rectangular resolution chart. Specifically, the physical phantom 30 is a rectangular chart in which a test pattern is formed by creating a striped grid 31 on a rectangular plate. The physical phantom 30 has a pair of opposing long sides 32 and a pair of opposing short sides 33. The grid 31 extends along the short sides 33.

[0038] When the physical phantom 30 is radiographically scanned, a test pattern corresponding to the shape of the grid 31 appears in the radiographic image. The edge direction of this test pattern is parallel to the short side 33 of the physical phantom 30. In Figure 2, reference numeral 34 indicates the region containing the test pattern (hereinafter referred to as the test pattern region).

[0039] Figure 3 shows an example of a physical phantom 30 attached to the detection surface 7A of the cassette holder 7. When evaluating the sharpness of the FPD 5 during maintenance, the medical physicist attaches the physical phantom 30 to the detection surface 7A using tape or the like. The medical physicist performs radiography with the physical phantom 30 attached to the detection surface 7A. In this way, the physical phantom 30 is used by the medical physicist by being attached to an arbitrary position on the detection surface 7A, so the position and angle of the test pattern area 34 will differ each time the physical phantom 30 is attached.

[0040] Figure 4 schematically shows an example of processing by the detection unit 21. The detection unit 21 receives the radiation image D1 acquired by the acquisition unit 20 from the control unit 10. The radiation image D1 shows the physical phantom 30. The detection unit 21 detects the test pattern region 34 using the machine learning model LM1 and detects the edge direction by utilizing the fact that the edge direction of the test pattern is parallel to the short side 33 of the physical phantom 30. The machine learning model LM1 corresponds to the "first machine learning model" related to the technology of this disclosure.

[0041] Specifically, first, the detection unit 21 generates a LOG-converted image D2 by performing a LOG conversion on the radiation image D1. LOG conversion is a type of grayscale conversion. By performing a LOG conversion on the radiation image D1, whose pixel values ​​depend linearly on the radiation dose, an image D2 is generated that is less dependent on the radiation dose.

[0042] Next, the detection unit 21 performs a cropping process on image D2 to generate image D3 in which the region containing the physical phantom 30 is extracted. Then, the detection unit 21 performs an average reduction process on image D3 to generate image D4, which is sized to be suitable as input to the machine learning model LM1. This is to improve the processing speed of the machine learning model LM1.

[0043] Next, the detection unit 21 inputs image D4 to the machine-learned model LM1. The machine-learned model LM1 is a deep learning model such as a CNN (Convolutional Neural Network) and performs so-called segmentation. The machine-learned model LM1 detects the test pattern region 34 of the physical phantom 30 from image D4 and outputs image D5, which represents the region containing the test pattern region 34. The machine-learned model LM1 is a deep learning model that has been pre-machine-trained using training data. The region included in the training data as the correct data is the smallest rectangular region containing the test pattern region 34.

[0044] Furthermore, the machine learning model LM1 may be one in which machine learning has been performed to detect the entire physical phantom 30 as the test pattern region 34. In this case, the region included in the training data as the ground truth data can be the entire physical phantom 30.

[0045] Next, the detection unit 21 generates an image D6 representing the edges of the test pattern region 34 by performing edge detection processing on image D5 using a filter such as a Laplacian filter. Next, the detection unit 21 detects a straight line representing the boundary of the test pattern region 34 by performing a Hough transform on image D6. The detected straight line is parallel to the long side 32 or short side 33 of the physical phantom 30. Then, the detection unit 21 determines the edge direction by measuring the angle of the straight line representing the short side 33. The detection unit 21 supplies the angle information representing the edge direction to the correction unit 22.

[0046] Next, the detection unit 21 generates an enlarged image D8 by performing a linear enlargement process on image D5. Then, the detection unit 21 uses the linear information (hereinafter referred to as linear information) detected by the Hough transform to detect the test pattern region 34 within image D8. When performing a linear enlargement process, the boundary may become zigzag rather than linear, but by using the linear information, the test pattern region 34 within image D8 can be detected with high accuracy.

[0047] The detection unit 21 then supplies the information representing the test pattern region 34 detected from the image D8 to the correction unit 22 as region information.

[0048] The correction unit 22 performs defective pixel correction on the radiation image D1 using the defective pixel data 16 and the angle information and region information supplied from the detection unit 21. Based on the defective pixel data 16 and region information, the correction unit 22 selects defective pixels within the test pattern region 34 as pixels to be corrected, sets a mask centered on the pixels to be corrected, and corrects the pixels to be corrected by performing interpolation processing.

[0049] Figure 5 illustrates the angular information representing the edge direction. Assume that the pixels constituting the radiation image D1 are arranged parallel to the X and Y directions. The X and Y directions are orthogonal to each other. Let α be the angle between the X direction and the edge direction. Hereafter, angle α will be referred to as the edge angle α. The correction unit 22 selects a mask corresponding to the edge angle α from a plurality of masks, which will be described later, and corrects the pixels to be corrected.

[0050] Figure 6 shows an example of multiple masks used for correcting point defects. The mask size shown in Figure 6 is 3x3 pixels. Ps indicates the pixel to be corrected. Pixels with hatched lines are normal pixels Pi used for interpolation. The same applies to the following examples.

[0051] Masks are set at 45° intervals within the range of possible edge angles α. The position of normal pixels Pi differs in each mask. Normal pixels Pi are positioned in a direction corresponding to the edge angle α, centered on the pixel to be corrected Ps. For example, in the mask used when the edge angle α is in the range of 0 ± 22.5°, a pair of normal pixels Pi are positioned opposite each other in the X direction, centered on the pixel to be corrected Ps. Similarly, in the mask used when the edge angle α is in the range of 45 ± 22.5°, a pair of normal pixels Pi are positioned opposite each other in a direction that forms a 45° angle with respect to the X direction, centered on the pixel to be corrected Ps. The same applies to other masks.

[0052] As shown in Figure 6, when the mask size is 3x3 pixels, the correction unit 22 performs interpolation using normal pixels Pi that are located in the edge direction from the pixel to be corrected Ps, selected from the 8 pixels in the mask excluding the pixel to be corrected Ps. In this way, the normal pixels Pi used in the interpolation process are located almost in the edge direction from the pixel to be corrected Ps, so the pixel to be corrected Ps can be corrected with high accuracy. In this disclosure, "normal pixels located in the edge direction from the pixel to be corrected" is not limited to normal pixels that are located in a direction that perfectly coincides with the edge direction from the pixel to be corrected, but also includes normal pixels that are located within a certain angular range based on the edge direction.

[0053] Figure 7 shows another example of several masks used for correcting point defects. The mask size shown in Figure 7 is 5x5 pixels. In the example shown in Figure 7, masks are set at 11.25° intervals within the range of possible edge angles α. Similar to the example shown in Figure 6, the normal pixels Pi are set in the direction corresponding to the edge angle α, centered on the pixels Ps to be corrected.

[0054] By setting the mask size to 5x5 pixels, it is possible to set the mask at smaller angle intervals than in the case of 3x3 pixels, thus enabling more accurate correction of the target pixel Ps. In the example shown in Figure 7, interpolation may also be performed by weighting the pixel value of the normal pixel Pi according to its distance from the target pixel Ps. For example, the greater the distance from the target pixel Ps, the smaller the weight applied to the pixel value of the normal pixel Pi.

[0055] The examples shown in Figures 6 and 7 are not the only ones; the mask size may be 7x7 pixels or larger. This allows masks to be set at smaller angle intervals, enabling more accurate correction of the target pixels Ps.

[0056] Figure 8 shows an example of multiple masks used to correct line defects. In particular, Figure 8 shows an example of a mask when the line defect extends in the X direction. In the case of line defects, the mask contains defective pixels other than the target pixel Ps, so the position of the normal pixels Pi is changed so that the defective pixels are not used as normal pixels Pi in the interpolation process. Specifically, when the edge angle α is in the range of 0 ± 22.5°, the pixels used in the interpolation process will be on the line defect, so pixels whose angle with respect to the target pixel Ps is as close to 0° as possible are set as the normal pixels Pi used in the interpolation process.

[0057] The correction unit 22 performs interpolation in the case of line defects as well as point defects, using normal pixels Pi located in the edge direction from the target pixel Ps, selected from the eight pixels in the mask excluding the target pixel Ps. Since the normal pixels Pi used in the interpolation process are located almost in the edge direction from the target pixel Ps, the target pixel Ps can be corrected with high accuracy.

[0058] Figure 9 shows another example of several masks used to correct line defects. The mask size shown in Figure 9 is 5x5 pixels. In the example shown in Figure 9, as in the example shown in Figure 8, the position of normal pixels Pi is changed so that defective pixels within the mask are not used as normal pixels Pi in the interpolation process.

[0059] The examples shown in Figures 8 and 9 are not the only ones; the mask size may be 7x7 pixels or larger. This allows for setting masks at smaller angle intervals, enabling more accurate correction of the target pixel Ps. The same applies when the line defect extends in the Y direction.

[0060] Furthermore, as shown in Figure 10, the correction unit 22 excludes pixels outside the test pattern area 34 from the mask in both the case of point defects and line defects, so as not to be used for correction. In other words, the correction unit 22 corrects the target pixel Ps using only normal pixels Pi included in the test pattern area 34. Since there are no edges outside the test pattern area 34, the target pixel Ps can be corrected with high accuracy by excluding pixels not included in the test pattern area 34 (i.e., pixels that do not contain edge information).

[0061] Furthermore, if weighting is applied to the pixel values ​​of normal pixels Pi according to their distance from the pixels Ps to be corrected, the weights should be normalized using a normalization coefficient corresponding to the number of normal pixels Pi used for correction.

[0062] Figure 11 shows an example of the correction process flow by the correction unit 22. First, the correction unit 22 acquires defective pixel data 16 from the storage unit 13 (step S10). Next, the correction unit 22 acquires angle information and region information from the detection unit 21 (step S11).

[0063] The correction unit 22 selects one defective pixel within the test pattern region 34 as the pixel Ps to be corrected based on the defective pixel data 16 and region information (step S12). Next, the correction unit 22 selects a mask corresponding to the angle α included in the angle information (step S13). Next, the correction unit 22 excludes pixels outside the test pattern region 34 from the mask (step S14). Then, the correction unit 22 corrects the pixel Ps to be corrected using only the normal pixels Pi included in the test pattern region 34 (step S15).

[0064] Next, the correction unit 22 determines whether or not the correction of all defective pixels in the test pattern area 34 has been completed (step S16). If the correction unit 22 determines that the correction of all defective pixels in the test pattern area 34 has not been completed (step S16: NO), the process returns to step S12. In step S12, the correction unit 22 selects the defective pixels that have not been corrected as the pixels to be corrected Ps.

[0065] If the correction unit 22 determines that the correction of all defective pixels in the test pattern area 34 has been completed (step S16: YES), it terminates the process.

[0066] In conventional correction methods, if a defective pixel exists on a high-contrast edge, a correction residual is visible in the corrected radiographic image. This correction residual is particularly pronounced in physical phantoms with high-contrast, high-frequency test patterns. In contrast, the correction method according to this disclosure corrects the target pixel Ps using normal pixels Pi other than the defective pixels within the test pattern region 34 that exists in the edge direction from the target pixel Ps, thus enabling accurate correction of defective pixels. Furthermore, when line defects are the target of correction, accurate correction can be achieved not only for line defects with a width of one pixel, but also for line defects with a width of two or more pixels by performing correction along the edge direction.

[0067] [First variation] Next, a modified example of the correction process by the correction unit 22 will be described. In this modified example, the correction unit 22 sets a mask centered on the pixel Ps to be corrected, and corrects the pixel Ps to be corrected using a weighted sum of the pixel values ​​of the normal pixels included within the mask.

[0068] Figure 12 illustrates the point defect correction process by the correction unit 22 in a modified example. The mask size shown in Figure 12 is 5 × 5 pixels. QL(X,Y) represents the pixel value of one pixel within the mask. θ(X,Y) represents the angle made with respect to the X direction by the line connecting the pixel and the pixel to be corrected Ps. D(X,Y) represents the distance between the pixel and the pixel to be corrected Ps. The coordinates of the pixel to be corrected Ps are set as the origin.

[0069] The correction unit 22 replaces the pixel value of the pixel Ps to be corrected with QL(0,0) calculated based on the following equations (1) to (3). The Σ in equation (1) means that the addition is performed by changing X and Y within the ranges of -2≦X≦2 and -2≦Y≦2 (excluding the origin).

number

[0070] Furthermore, the weight W(X,Y) in equation (1) above is expressed by equation (2) below. In equation (2) below, n and m are parameters, which are set to positive integers, for example.

number

[0071] Furthermore, the difference angle δ(X,Y) in equation (2) above is expressed by equation (3) below. The difference angle δ(X,Y) is the angle between the line connecting the normal pixel used for correction and the pixel Ps to be corrected and the edge direction.

number

[0072] Equations (1) to (3) above represent the correction of the target pixel Ps by assigning a weight W(X,Y) to each normal pixel within the mask according to the difference angle δ(X,Y). The weight W(X,Y) increases as the difference angle δ(X,Y) decreases. In other words, the weight W(X,Y) is largest for normal pixels that are located in the edge direction from the target pixel Ps. As a result, good correction results with little correction residual can be obtained even for defects that span edges.

[0073] Note that the function representing the weight W(X,Y) is not limited to equation (2) above; any function that is maximized at δ(X,Y)=0° and δ(X,Y)=180°, and minimized at δ(X,Y)=90°, is acceptable. For example, equation (2A) below may be used instead of equation (2) above.

number

[0074] In this modified example, pixels outside the test pattern area 34 are excluded from the mask so as not to be used for correction (see Figure 10). That is, the correction unit 22 corrects the target pixels Ps using only the normal pixels included in the test pattern area 34.

[0075] Figure 13 illustrates the line defect correction process by the correction unit 22 in a modified example. The line defect correction process is the same as the point defect correction process, and QL(0,0) is calculated by excluding defective pixels from the pixels included in the mask.

[0076] Figure 14 shows an example of the correction process flow by the correction unit 22 in a modified example. The correction process flow in the modified example differs from the correction process flow in the above embodiment only in steps S13 to S15.

[0077] In this modified example, in step S13, the correction unit 22 excludes pixels outside the test pattern area 34 from the mask. In step S14, the correction unit 22 calculates QL(0,0) based on equations (1) to (3) above. In step S15, the correction unit 22 replaces the pixel value of the correction target pixel Ps with QL(0,0).

[0078] Furthermore, in this disclosure, "using normal pixels that exist in the edge direction from the pixel to be corrected" does not preclude the use of normal pixels with small weights other than the normal pixels in the edge direction from the pixel to be corrected, as in this modified example.

[0079] [Second variation] Next, a modified version of the physical phantom 30 will be described. Figure 15 shows the modified physical phantom 30A. The physical phantom 30A is a radial resolution chart used for evaluating focus performance. A radial grid 31A is formed on the physical phantom 30A. That is, in the radiographic image obtained by performing radiography using the physical phantom 30A, multiple radial edges appear.

[0080] In this modified example, as shown in Figure 16, the detection unit 21 detects a circular test pattern region 34A from the radiation image and calculates the centroid of the detected test pattern region 34A to determine the center coordinate C. The detection unit 21 then determines the edge direction (i.e., the edge angle α) based on the relative positional relationship between the pixel to be corrected Ps and the center coordinate C. For example, the detection unit 21 calculates the angle that the line connecting the pixel to be corrected Ps and the center coordinate C makes with respect to the X direction, and sets the calculated value as the edge angle α.

[0081] In this modified example, it is necessary to detect the edge direction each time a pixel Ps to be corrected is selected, but the method for correcting the pixel Ps to be corrected is the same as in the above embodiment.

[0082] The physical phantom 30 is not limited to a radial resolution chart, but may also be a chart with other rotationally symmetrical patterns. When using a chart with such a rotationally symmetrical pattern, the edge direction can be determined based on the relative positional relationship between the pixels to be corrected Ps and the central coordinate C, similar to the example shown in Figure 16.

[0083] [Third variation] Next, a modified example of the edge direction detection process by the detection unit 21 will be described. In the above embodiment, the detection unit 21 determines the edge direction based on the linear information detected by the Hough transform, but in this modified example, the edge direction is determined by a machine learning model.

[0084] Figure 17 shows the edge direction detection process by the detection unit 21 in a modified example. The detection unit 21 acquires angular information including the edge direction by inputting an image containing the test pattern region 34 detected from the radiation image D1 into the machine learning model LM2. The machine learning model LM2 is a deep learning model that has been pre-machine-trained using training data so that when an image of a resolution chart is input, it outputs angular information including the edge direction. The machine learning model LM2 corresponds to the "second machine learning model" in the technology of this disclosure.

[0085] In the above embodiment, for example, the hardware structure of the Processing Unit that performs various processes such as the acquisition unit 20, the detection unit 21, and the correction unit 22 can be the various processors shown below. As mentioned above, the various processors include a CPU, which is a general-purpose processor that executes software (programs) and functions as various processing units, as well as a PLD (Programmable Logic Device), which is a processor whose circuit configuration can be changed after manufacturing, such as an FPGA (Field Programmable Gate Array), and a dedicated electrical circuit, which is a processor with a circuit configuration specifically designed to perform a particular process, such as an ASIC (Application Specific Integrated Circuit).

[0086] A single processing unit may consist of one of these various processors, or it may consist of a combination of two or more processors of the same or different types (for example, a combination of multiple FPGAs, and / or a combination of a CPU and an FPGA). Alternatively, multiple processing units may be composed of a single processor.

[0087] Examples of configuring multiple processing units with a single processor include, firstly, a configuration where one or more CPUs and software combine to form a single processor, as exemplified by client and server computers, and this processor functions as multiple processing units. Secondly, a configuration using a processor that realizes the functions of the entire system, including multiple processing units, on a single IC (Integrated Circuit) chip, as exemplified by System-on-a-Chip (SoC). Thus, various processing units are configured, in terms of hardware structure, using one or more of the above-mentioned processors.

[0088] Furthermore, the hardware structure of these various processors can more specifically utilize electrical circuits, which are combinations of circuit elements such as semiconductor devices.

[0089] All documents, patent applications, and technical standards described herein are incorporated by reference to the same extent as if each individual document, patent application, and technical standard were specifically and individually noted to be incorporated by reference. [Explanation of symbols]

[0090] 2. Radiography System 3. Radiation Generating Devices 4 Radiation tubes 6. Information Processing Device 7 Cassette holder 7A Detection surface 8. Scatter Removal Grid 10 Control Unit 11 displays 12 Control section 13 Storage section 14 Image Processing Unit 15 Programs 16. Defective pixel data 20 Acquisition Department 21 Detection unit 22 Correction section 30,30A Physical Phantom 31,31A lattice 32 Long side 33 Short side 34,34A Test Pattern Area α edge angle C center coordinates D1 Radiological image Images D2-D8 H Subject LM1, LM2 Machine Learning Pre-programmed Models Pi Normal Pixel Ps Correction Target Pixels R radiation

Claims

1. An information processing device that performs processing to correct defective pixels in a radiographic image obtained by radiographically imaging a physical phantom on which a test pattern has been formed, Equipped with a processor, The aforementioned processor, By inputting the image based on the aforementioned radiation image into the first machine learning model, the region containing the test pattern is detected. Determine the edge direction of the test pattern within the region, The defective pixels within the aforementioned region are designated as correction target pixels, and the correction target pixels are corrected using normal pixels other than the defective pixels within the aforementioned region that are located in the edge direction from the correction target pixels. Information processing device.

2. The first machine-learned model is a deep learning model that detects the region by performing segmentation. The information processing apparatus according to claim 1.

3. The aforementioned physical phantom is a rectangular resolution chart, The processor determines the edge direction based on a straight line representing the boundary of the region detected by the Hough transform. The information processing apparatus according to claim 1.

4. The processor determines the edge direction by inputting the image including the region into a second machine learning model. The information processing apparatus according to claim 1.

5. The aforementioned physical phantom is a radial resolution chart, The processor determines the edge direction based on the relative positional relationship between the pixel to be corrected and the center coordinates of the region. The information processing apparatus according to claim 1.

6. The processor corrects the target pixel using a plurality of normal pixels, which are set in advance according to the edge direction, from among the normal pixels within a mask set with the target pixel as the center. The information processing apparatus according to claim 1.

7. The processor corrects the target pixels by assigning weights to the plurality of normal pixels according to their distance from the target pixels. The information processing apparatus according to claim 6.

8. The processor corrects each of the normal pixels within a mask set with the target pixel as the center, by assigning a weight to each of the normal pixels, which is the difference angle between the direction of the normal pixel from the target pixel and the edge direction, and the distance from the target pixel to the normal pixel. The information processing apparatus according to claim 1.

9. An information processing method for correcting defective pixels in a radiographic image obtained by radiographically imaging a physical phantom on which a test pattern has been formed, By inputting the image based on the aforementioned radiation image into the first machine learning model, the region containing the test pattern is detected. To determine the edge direction of the test pattern within the said region, The defective pixels within the aforementioned region are designated as correction target pixels, and the correction target pixels are corrected using normal pixels other than the defective pixels within the aforementioned region that are located in the edge direction from the correction target pixels. Information processing methods including

10. A program that causes a computer to perform a process to correct defective pixels in a radiographic image obtained by radiographically imaging a physical phantom on which a test pattern has been formed, By inputting the image based on the aforementioned radiation image into the first machine learning model, the region containing the test pattern is detected. To determine the edge direction of the test pattern within the said region, The defective pixels within the aforementioned region are designated as correction target pixels, and the correction target pixels are corrected using normal pixels other than the defective pixels within the aforementioned region that are located in the edge direction from the correction target pixels. A program that causes a computer to perform a process that includes [a specific action].

Images