Device, method, and storage medium for predicting motion organ disease, learning device, method, and storage medium, and learning completed neural network

Abstract

The present application relates to a moving organ disease prediction device, a moving organ disease prediction method, a moving organ disease prediction program, a learning device, a learning method, a learning program, and a learned neural network. The moving organ disease prediction device includes at least one processor configured to derive, from a first radiograph and a second radiograph obtained by radiographing a subject including a bone portion and a soft portion using radiation having different energy distributions, a bone mineral content of a target bone among bones included in the subject, a muscle amount around the target bone, shape information representing a shape of the target bone, and shape information representing a shape of a bone adjacent to the target bone. The processor is configured to derive a probability of occurrence of a moving organ disease related to the target bone based on the bone mineral content of the target bone, the muscle amount around the target bone, the shape information of the target bone, and the shape information of the bone adjacent to the target bone.

Description

Technical Field

[0001] This invention relates to a device, method and procedure for predicting diseases of the musculoskeletal system, a learning device, method and procedure, and a neural network for learning. Background Technology

[0002] Diseases related to the locomotor organs, such as fractures and dislocations, can lead to bedridden status. In particular, dislocations of the hip joint and fractures of the femur and vertebrae are highly likely to result in bedridden patients. It is known that the 5-year survival rate for bedridden patients is lower than that for cancer. Therefore, various methods have been proposed for evaluating the risk of diseases of the locomotor organs, especially fractures.

[0003] For example, Patent Document 1 proposes a method that acquires bone mass and bone structure from radiographic images and uses a neural network to calculate future fracture risk. Patent Document 2 proposes a method that uses a neural network to estimate bone density from radiographic images and uses the estimated results and an arithmetic expression representing the probability of fracture to predict fracture. Patent Document 3 proposes a method that calculates bone mineral content and muscle mass for each pixel of a radiographic image, calculates statistical values ​​related to the subject based on bone mineral content and muscle mass, and evaluates fracture risk based on these statistical values.

[0004] Patent Document 1: Japanese Patent Publication No. 09-508813

[0005] Patent Document 2: Japanese Republication No. 2020-054738

[0006] Patent Document 3: International Publication No. 2020 / 166561

[0007] However, the goal is to predict diseases of the motor organs with even higher accuracy. Summary of the Invention

[0008] The present invention was made in view of the above circumstances, and its purpose is to be able to predict diseases of the musculoskeletal system with high accuracy.

[0009] The musculoskeletal disease prediction device according to the present invention includes at least one processor.

[0010] The processor derives, based on the first and second radiographic images obtained by photographing a subject containing bones and soft tissues using radiation with different energy distributions, the amount of bone minerals in the bone of the subject, the amount of muscle surrounding the bone, shape information representing the shape of the bone, and shape information representing the shape of the bones adjacent to the bone.

[0011] The probability of developing musculoskeletal diseases related to the target bone is derived based on the amount of bone salts in the target bone, the amount of muscle around the target bone, the shape information of the target bone, and the shape information of the bones adjacent to the target bone.

[0012] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the processor can function as a learning-complete neural network that performs machine learning using teacher data, which includes the amount of bone salts in the target bone, the amount of muscle around the target bone, shape information representing the shape of the target bone, shape information representing the shape of the bones adjacent to the target bone, and positive solution data representing the probability of developing musculoskeletal diseases related to the target bone.

[0013] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the processor can display the derived probability of occurrence of musculoskeletal diseases on a display screen.

[0014] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the processor can display a graph showing the relationship between at least one of bone mineral content and muscle mass and the probability of occurrence of musculoskeletal diseases.

[0015] The chart also displays a plot representing the probability of occurrence of musculoskeletal diseases and a plot representing the value obtained by changing at least one of the probability of occurrence or bone mineral content and muscle mass.

[0016] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the changed value can be a target value of at least one of bone mineral content and muscle mass or a target value of the probability of occurrence of musculoskeletal disease.

[0017] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the processor may also display options for medical interventions to achieve target values ​​for at least one of bone mineral content and muscle mass, or options for medical interventions to achieve target values ​​for musculoskeletal diseases.

[0018] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the medical intervention can be an exercise method for exercising muscles related to the bone of the object.

[0019] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the target bone can be the femur.

[0020] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the target bone can be a vertebra.

[0021] Furthermore, in the musculoskeletal disease prediction device according to the present invention, the musculoskeletal disease can be at least one of fracture and dislocation.

[0022] The learning device according to the present invention includes at least one processor.

[0023] The processor uses data from the bones in the human body, including the amount of bone minerals in the target bone, the amount of muscle surrounding the target bone, shape information representing the shape of the target bone, shape information representing the shape of the bones adjacent to the target bone, and positive solution data representing the probability of developing musculoskeletal diseases related to the target bone, as teacher data to perform machine learning on the neural network. This results in the following learning-complete neural network, which, given the input of the amount of bone minerals in the target bone, the amount of muscle surrounding the target bone, the shape information of the target bone, and the shape information of the bones adjacent to the target bone, outputs the probability of developing musculoskeletal diseases.

[0024] In the method for predicting diseases of the musculoskeletal system according to the present invention, based on a first radiographic image and a second radiographic image obtained by photographing a subject containing bones and soft tissues using radiation with different energy distributions, the amount of bone salts in the target bone, the amount of muscle surrounding the target bone, shape information representing the shape of the target bone, and shape information representing the shape of the bones adjacent to the target bone are derived.

[0025] The probability of developing musculoskeletal diseases related to the target bone is derived based on the amount of bone salts in the target bone, the amount of muscle around the target bone, the shape information of the target bone, and the shape information of the bones adjacent to the target bone.

[0026] In the learning method according to the present invention, a learning completion neural network is constructed by using the bone mineral content of the target bone, the muscle mass surrounding the target bone, shape information representing the shape of the target bone, shape information representing the shape of the bones adjacent to the target bone, and positive solution data representing the probability of developing musculoskeletal diseases related to the target bone as teacher data to perform machine learning on the neural network. This learning completion neural network outputs the probability of developing musculoskeletal diseases if it is input with the bone mineral content of the target bone, the muscle mass surrounding the target bone, the shape information of the target bone, and the shape information of the bones adjacent to the target bone.

[0027] Alternatively, the method and learning method for predicting musculoskeletal diseases according to the present invention can be provided as a program for computer execution.

[0028] In the learning-complete neural network according to the present invention, if the input is the amount of bone salts in the target bone, the amount of muscle around the target bone, the shape information representing the shape of the target bone, and the shape information representing the shape of the bones adjacent to the target bone, then the output is the probability of the occurrence of musculoskeletal diseases related to the target bone.

[0029] Invention Effects

[0030] According to the present invention, diseases of the musculoskeletal system can be predicted with high accuracy. Attached Figure Description

[0031] Figure 1 This is a schematic block diagram illustrating the structure of a radiographic imaging system for a musculoskeletal disease prediction device and learning device according to an embodiment of the present invention.

[0032] Figure 2 This is a diagram showing the schematic structure of the musculoskeletal disease prediction device and learning device according to an embodiment of the present invention.

[0033] Figure 3 This is a diagram illustrating the functional structure of the musculoskeletal disease prediction device and learning device according to an embodiment of the present invention.

[0034] Figure 4 This is a diagram representing the functional structure of the information output section.

[0035] Figure 5 It is a diagram representing a bone structure.

[0036] Figure 6 It is a diagram representing a soft-screen image.

[0037] Figure 7 It is a graph representing the result of segmentation.

[0038] Figure 8 This is a diagram used to illustrate the layout of the joint portion of the femur and the joint portion of the pelvis.

[0039] Figure 9 It is a diagram that represents the shape information of the object's bone and the shape information of the bones adjacent to the object's bone.

[0040] Figure 10 It is a graph showing the relationship between the contrast of the bony and soft parts and the thickness of the subject.

[0041] Figure 11 This is a diagram representing an example of a lookup table.

[0042] Figure 12 This is a diagram showing an example of the energy spectrum of radiation transmitted through muscle tissue and radiation transmitted through adipose tissue.

[0043] Figure 13 It is a diagram used to illustrate the setting of the area around the femur in a muscle image.

[0044] Figure 14 This is a diagram showing a schematic structure of the neural network used in this embodiment.

[0045] Figure 15 This is a graph representing teacher data.

[0046] Figure 16 This is a diagram used to illustrate the learning process of a neural network.

[0047] Figure 17 It is a diagram representing the displayed screen.

[0048] Figure 18 It represents the image displayed on the screen.

[0049] Figure 19 This is a flowchart of the learning process performed in this embodiment.

[0050] Figure 20 This is a flowchart of the musculoskeletal disease prediction process performed in this embodiment.

[0051] Figure 21 This is a schematic block diagram illustrating the structure of a radiographic imaging system for a musculoskeletal disease prediction device and learning device according to another embodiment of the present invention. Detailed Implementation

[0052] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. Figure 1 This is a schematic block diagram illustrating the structure of a radiographic imaging system for predicting and learning musculoskeletal diseases according to an embodiment of the present invention. Figure 1 As shown, the radiographic imaging system according to this embodiment includes an imaging device 1 and a musculoskeletal disease prediction device and learning device (hereinafter, sometimes referred to as a musculoskeletal disease prediction device) 10 according to this embodiment.

[0053] The photographic apparatus 1 is a photographic device for energy subtraction using a so-called one-shot method, which modifies the energy of radiation such as X-rays emitted from radiation source 3 and transmitted through the subject H, and then irradiates the first radiation detector 5 and the second radiation detector 6. During the photographing process, such as... Figure 1 As shown, a first radiation detector 5, a radiation energy conversion filter 7 made of copper plates or the like, and a second radiation detector 6 are arranged sequentially from the side closest to the radiation source 3 to drive the radiation source 3. Furthermore, the first radiation detector 5 and the second radiation detector 6 are in close contact with the radiation energy conversion filter 7.

[0054] Therefore, in the first radiation detector 5, a first radiation image G1 of the subject H, formed by low-energy radiation including so-called soft filaments, is acquired. And in the second radiation detector 6, a second radiation image G2 of the subject H, formed by high-energy radiation from which the soft filaments have been removed, is acquired. The first radiation image G1 and the second radiation image G2 are input into the musculoskeletal disease prediction device 10. Both the first radiation image G1 and the second radiation image G2 are frontal images including the periphery of the hip area of ​​the subject H.

[0055] The first radiation detector 5 and the second radiation detector 6 are radiation detectors capable of repeatedly recording and reading out radiation images. They can be either direct-type radiation detectors that generate charge by directly receiving radiation, or indirect-type radiation detectors that temporarily convert radiation into visible light and then convert that visible light into a charge signal. Furthermore, as the method for reading out the radiation image signal, a TFT readout method, which reads out the radiation image signal by turning a TFT (thin film transistor) switch on or off, or an optical readout method, which reads out the radiation image signal by irradiating a readout light, is preferred. However, they are not limited to these methods, and other methods may also be used.

[0056] In addition, the musculoskeletal disease prediction device 10 is connected to the image storage system 9 via a network (not shown).

[0057] Image storage system 9 is a system for storing image data of radiographic images captured by photographic device 1. Image storage system 9 retrieves the image corresponding to a request from musculoskeletal disease prediction device 10 from the stored radiographic images and sends it to the requesting device. A specific example of image storage system 9 is PACS (Picture Archiving and Communication Systems).

[0058] Next, the musculoskeletal disease prediction device according to this embodiment will be described. First, refer to... Figure 2 The hardware structure of the musculoskeletal disease prediction device according to this embodiment will be described. For example... Figure 2As shown, the musculoskeletal disease prediction device 10 is a computer such as a workstation, server computer, or personal computer, and includes a CPU (Central Processing Unit) 11, non-volatile memory 13, and internal memory 16 as a temporary storage area. Furthermore, the musculoskeletal disease prediction device 10 includes a display 14 such as a liquid crystal display, input devices 15 such as a keyboard and mouse, and a network I / F (Interface) 17 connected to a network not shown. The CPU 11, memory 13, display 14, input devices 15, internal memory 16, and network I / F 17 are connected to a bus 18. The CPU 11 is an example of a processor in this invention.

[0059] The memory 13 is implemented using HDD (Hard Disk Drive), SSD (Solid State Drive), or flash memory, etc. The memory 13, which serves as the storage medium, stores the musculoskeletal disease prediction program 12A and the learning program 12B installed in the musculoskeletal disease prediction device 10. The CPU 11 reads the musculoskeletal disease prediction program 12A and the learning program 12B from the memory 13, expands them into the internal memory 16, and executes the expanded musculoskeletal disease prediction program 12A and the learning program 12B.

[0060] Furthermore, the musculoskeletal disease prediction program 12A and the learning program 12B are stored in a network-connected server computer's storage device or network storage in a state accessible from the outside, and are downloaded and installed on the computer constituting the musculoskeletal disease prediction device 10 upon request. Alternatively, they can be recorded on recording media such as DVD (Digital Versatile Disc) or CD-ROM (Compact Disc Read Only Memory) and distributed, and installed on the computer constituting the musculoskeletal disease prediction device 10 from such recording media.

[0061] Next, the functional structure of the musculoskeletal disease prediction device and learning device according to this embodiment will be described. Figure 3 This is a diagram illustrating the functional structure of the musculoskeletal disease prediction device and learning device according to this embodiment. (See diagram for example.) Figure 3As shown, the musculoskeletal disease prediction device 10 includes an image acquisition unit 21, an information acquisition unit 22, an information derivation unit 23, a probability derivation unit 24, a learning unit 25, and a display control unit 26. Furthermore, the CPU 11 functions as the image acquisition unit 21, information acquisition unit 22, information derivation unit 23, probability derivation unit 24, and display control unit 26 by executing the musculoskeletal disease prediction program 12A, and further functions as the learning-completed neural network 24A, which will be described later. Moreover, the CPU 11 functions as the learning unit 25 by executing the learning program 12B.

[0062] The image acquisition unit 21 acquires frontal images of the subject H near its crotch, namely, the first radiation image G1 and the second radiation image G2, from the first radiation detector 5 and the second radiation detector 6 by having the imaging device 1 capture images of the subject H. When acquiring the first radiation image G1 and the second radiation image G2, imaging conditions are set, including the imaging dose, tube voltage, the distance between the radiation source 3 and the surfaces of the first radiation detector 5 and the second radiation detector 6 (SID, Source Image Receptor Distance), the distance between the radiation source 3 and the surface of the subject H (SOD, Source Object Distance), and the presence or absence of a scattered radiation removal grid.

[0063] SOD and SID, as described later, are used for calculating body thickness distribution. For SOD, it is preferable to obtain it using a T-of-Flight (TOF) camera, for example. For SID, it is preferable to obtain it using a potentiometer, ultrasonic rangefinder, or laser rangefinder, for example.

[0064] The imaging conditions can be set by the operator through the input device 15. The set imaging conditions are stored in the memory 13. Alternatively, in this embodiment, the first radiation image G1 and the second radiation image G2 can be acquired and stored in the memory 13 through a separate program with the musculoskeletal disease prediction program 12A. In this case, the image acquisition unit 21 reads the first radiation image G1 and the second radiation image G2 stored in the memory 13 to acquire these radiation images for processing.

[0065] The information acquisition unit 22 acquires teacher data from the image storage system 9 via the network I / F17 for learning the neural network described later.

[0066] The information export unit 23 exports the amount of bone salts in the target bone contained in the subject H, the amount of muscle surrounding the target bone, shape information representing the shape of the target bone, and shape information representing the shape of the bones adjacent to the target bone. In this embodiment, the target bone is defined as the femur.

[0067] Figure 4 This is a schematic block diagram showing the structure of the information export unit 23. For example... Figure 4 As shown, the information export unit 23 includes a scattering removal unit 31, an image export unit 32, a segmentation unit 33, a bone mineral content export unit 34, and a muscle mass export unit 35. The CPU 11 functions as the scattering removal unit 31, the image export unit 32, the segmentation unit 33, the bone mineral content export unit 34, and the muscle mass export unit 35 by executing the musculoskeletal disease prediction program 12A.

[0068] Here, each of the first radiation image G1 and the second radiation image G2 contains, in addition to the primary ray component of radiation transmitted through the subject H, a scattered ray component based on radiation scattered within the subject H. Therefore, the scattered ray removal unit 31 removes the scattered ray component from the first radiation image G1 and the second radiation image G2. For example, the scattered ray removal unit 31 can use the method described in Japanese Patent Application Publication No. 2015-043959 to remove the scattered ray component from the first radiation image G1 and the second radiation image G2. When using the method described in Japanese Patent Application Publication No. 2015-043959, the volume thickness distribution of the subject H and the scattered ray component for scattering component removal are simultaneously derived.

[0069] The following describes the removal of scattered ray components from the first radiation image G1, but the removal of scattered ray components from the second radiation image G2 can also be performed in the same way. First, the scattered ray removal unit 31 acquires a virtual model K of the subject H having an initial body thickness distribution T0(x, y). The virtual model K is data that virtually represents the subject H in which the body thickness according to the initial body thickness distribution T0(x, y) is correlated with the coordinate positions of each pixel in the first radiation image G1. In addition, the virtual model K of the subject H having the initial body thickness distribution T0(x, y) can be pre-stored in the memory 13. Furthermore, the body thickness distribution T(x, y) of the subject H can also be calculated based on the SID and SOD included in the photographic conditions. In this case, the body thickness distribution can be obtained by subtracting SOD from SID.

[0070] Next, the scattering removal unit 31 generates an image based on the virtual model K as a calculated image of the first radiation image G1 obtained by photographing the subject H. This image is a composite of the calculated primary ray image obtained by photographing the primary ray image obtained by photographing the virtual model K and the calculated scattering image obtained by photographing the scattering image obtained by photographing the virtual model K.

[0071] Next, the scattering removal unit 31 corrects the initial thickness distribution T0(x, y) of the virtual model K to reduce the difference between the calculated image and the first radiation image G1. The scattering removal unit 31 repeats the generation of the calculated image and the correction of the thickness distribution until the difference between the calculated image and the first radiation image G1 meets a predetermined termination condition. The scattering removal unit 31 derives the thickness distribution that meets the termination condition as the thickness distribution T(x, y) of the subject H. Furthermore, the scattering removal unit 31 removes the scattering components contained in the first radiation image G1 by subtracting the scattering components that meet the termination condition from the first radiation image G1.

[0072] The image exporting unit 32 performs energy subtraction processing to export the bone image Gb (extracted from the bone portion) and the soft image Gs (extracted from the soft portion) of the subject H based on the first radiation image G1 and the second radiation image G2. Furthermore, the first radiation image G1 and the second radiation image G2 used in subsequent processing are radiation images with scattered radiation components removed. When exporting the bone image Gb, the image exporting unit 32 performs a weighted subtraction operation between the corresponding pixels of the first radiation image G1 and the second radiation image G2 as shown in equation (1) below. Therefore, as... Figure 5 As shown, bone images Gb are generated from the bone regions of the subject H contained in each of the radiographic images G1 and G2. In equation (1), β1 is a weighting coefficient. In addition, the pixel values ​​of each pixel in the bone region of the bone image Gb are called bone pixel values.

[0073] Gb(x,y)=G1(x,y)-β31×G2(x,y) (1)

[0074] On the other hand, when exporting the soft image Gs, the image exporting unit 32 performs corresponding pixel operations, such as weighted subtraction operations, on the first radiation image G1 and the second radiation image G2 as shown in the following formula (2). Thus, as... Figure 6 As shown, a soft image Gs (energy subtraction) is generated that extracts only the soft parts of the subject H contained in each of the radiation images G1 and G2. In equation (2), β2 is the weighting coefficient.

[0075] Gs(x,y)=G1(x,y)-β2×G2(x,y) (2)

[0076] Additionally, the soft tissue image Gs represents the soft tissue region formed by the soft tissue of the subject H. In this embodiment, the "soft tissue" of the subject H refers to tissue other than bone tissue, specifically including muscle tissue, adipose tissue, blood, and water.

[0077] The segmentation unit 33 segments the bone image Gb into the femur region, pelvic region, and vertebral region, which are the target bones. Segmentation can be performed using a machine learning extraction model that extracts the femur, pelvis, and vertebrae separately from the bone image Gb. Alternatively, templates representing the femur, pelvis, and vertebrae can be pre-stored in the memory 13, and segmentation can be performed by matching these templates with the templates in the bone image Gb.

[0078] Figure 7 This is a diagram showing the result of the segmentation by segmentation part 33. For example... Figure 7 As shown, the bone region in the bone image Gb is segmented into the femur region A1, the pelvis region A2, and the vertebral region A3. Additionally, in Figure 7 In the diagram, different shaded lines are assigned to the femoral region A1, the pelvic region A2, and the vertebral region A3 to illustrate the segmentation results.

[0079] On the other hand, regarding the vertebrae, the bone image Gb only includes the sacrum and lumbar vertebrae. The lumbar vertebrae are anatomically classified as L5, L4, L3, L2, and L1 from the pelvic side towards the neck. Therefore, the segmentation portion 33 preferably divides the sacrum and the five lumbar vertebrae into different regions.

[0080] And, as Figure 8 As shown, the segmentation unit 33 defines a region R1 containing the joint portion of the femur and a region R2 containing the joint portion of the pelvis in the bone image Gb. Additionally, in Figure 8 In this model, regions R1 and R2 are defined only on the left side of the subject H, but are also defined on the right side. Furthermore, by binarizing the femoral region A1 within region R1 and the regions outside it, as shown... Figure 9 As shown, shape information S1 representing the shape of the femur as the object bone is derived. Furthermore, by binarizing the pelvic region A2 within region R2 and the region outside it, as shown... Figure 9 As shown, shape information S2, representing the shape of the pelvis as a bone portion adjacent to the object bone, is derived.

[0081] The bone mineral content extraction unit 34 extracts the bone mineral content for each pixel of the bone image Gb. In this embodiment, the bone mineral content extraction unit 34 extracts the bone mineral content B by converting each pixel value of the bone image Gb into the pixel values ​​of the bone image acquired using reference imaging conditions. More specifically, the bone mineral content extraction unit 34 extracts the bone mineral content by correcting each pixel value of the bone image Gb using a correction coefficient obtained from a lookup table described later.

[0082] Here, the higher the tube voltage in radiation source 3 and the higher the energy of the radiation emitted from radiation source 3, the smaller the contrast between soft and bony areas in the radiation image. Furthermore, during the transmission of radiation through the subject H, beam hardening occurs, where the low-energy components of the radiation are absorbed by the subject H, resulting in high-energy radiation. The greater the thickness of the subject H, the greater the high-energy radiation caused by beam hardening.

[0083] Figure 10 This is a graph showing the relationship between the contrast of the bony and soft parts and the thickness of the subject H. Additionally, in Figure 10 The diagram shows the relationship between the contrast of the skeletal and soft parts relative to the body thickness H of the subject at three tube voltages: 80kV, 90kV, and 100kV. Figure 10 As shown, the higher the tube voltage, the lower the contrast. Furthermore, if the thickness of the subject H exceeds a certain value, the greater the thickness, the lower the contrast. Additionally, the larger the pixel value of the bone region in the bone image Gb, the greater the contrast between the bone and soft tissue. Therefore, the larger the pixel value of the bone region in the bone image Gb, the lower the contrast between the bone and soft tissue. Figure 10 The relationship shown shifts towards the higher contrast side.

[0084] In this embodiment, the difference in contrast in the bone image Gb corresponding to the tube voltage at the time of imaging, and a lookup table for obtaining correction coefficients to correct the decrease in contrast caused by beam hardening, are stored in memory 13. The correction coefficients are coefficients used to correct the pixel values ​​of the bone image Gb.

[0085] Figure 11 This is a diagram illustrating an example of a lookup table stored in memory 13. Figure 11 The example shows a lookup table LUT1 for setting the reference imaging conditions to a tube voltage of 90 kΩ. For example... Figure 11 As shown, in lookup table LUT1, the larger the tube voltage and the greater the thickness of the subject H, the larger the correction factor is set. Figure 11 In the example shown, the reference imaging condition is a tube voltage of 90 kY. Therefore, when the tube voltage is 90 kY and the volume thickness is 0, the correction factor becomes 1. Additionally, in Figure 11 In the diagram, the lookup table LUT1 is shown in two dimensions, but the correction coefficients vary depending on the pixel values ​​of the bone region. Therefore, the lookup table LUT1 is actually a three-dimensional table with axes added to represent the pixel values ​​of the bone region.

[0086] The bone mineral content extraction unit 34 extracts the body thickness distribution T(x, y) of the subject H from the lookup table LUT1 and the correction coefficient CO(x, y) of each pixel corresponding to the imaging conditions, including the tube voltage setting, stored in the memory 13. Furthermore, as shown in equation (3) below, the bone mineral content extraction unit 34 extracts the bone mineral content B(x, y) (g / cm³) of each pixel in the bone region of the bone image Gb by multiplying each pixel (x, y) in the bone region of the bone image Gb by the correction coefficient CO(x, y). 2 The bone mineral content B(x, y) thus derived represents the amount of bone mineral content in the pixel values ​​of the bone region contained in the radiographic image obtained by photographing the subject H using a tube voltage of 90 kV as the reference imaging condition, and in which the effects of beam hardening have been removed.

[0087] B(x,y)=CO(x,y)×Gb(x,y) (3)

[0088] Furthermore, in this embodiment, the target bone is the femur. Therefore, the bone salt quantity extraction unit 34 can extract the bone salt quantity only from the femur region A1 in the bone image Gb.

[0089] The muscle mass extraction unit 35 extracts muscle mass based on the pixel value of each pixel in the soft tissue region of the soft tissue image Gs. As described above, soft tissue includes muscle tissue, adipose tissue, blood, and water. In the muscle mass extraction unit 35 of this embodiment, tissue other than adipose tissue in the soft tissue is considered as muscle tissue. That is, in the muscle mass extraction unit 35 of this embodiment, muscle tissue also includes non-adipose tissue, including blood and water, and is treated as muscle tissue.

[0090] The muscle volume extraction section 35 separates muscle and fat from the soft tissue image Gs by utilizing the differences in energy properties between muscle and adipose tissue. For example... Figure 12 As shown, the amount of radiation transmitted through subject H is reduced compared to the radiation incident before it reaches the human body. Furthermore, muscle tissue and adipose tissue absorb different amounts of energy and have different attenuation coefficients; therefore, the energy spectra of radiation transmitted through muscle tissue and radiation transmitted through adipose tissue differ. For example... Figure 12 As shown, the energy spectrum of the radiation emitted by the subject H and irradiated by the first radiation detector 5 and the second radiation detector 6 depends on the body composition of the subject H, specifically, on the ratio of muscle tissue to adipose tissue. Adipose tissue transmits radiation more readily than muscle tissue; therefore, when the proportion of muscle tissue is higher than that of adipose tissue, the dose of radiation transmitted through the body is lower.

[0091] Therefore, the muscle mass extraction unit 35 separates muscle and fat from the soft tissue image Gs by utilizing the differences in energy characteristics between the muscle and adipose tissues. That is, the muscle mass extraction unit 35 generates muscle and fat images from the soft tissue image Gs. Furthermore, the muscle mass extraction unit 35 extracts the muscle mass of each pixel based on the pixel values ​​of the muscle image.

[0092] Furthermore, the specific method by which the muscle mass extraction unit 35 separates muscle and fat from the soft image Gs is not limited. As an example, the muscle mass extraction unit 35 of this embodiment generates a muscle image from the soft image Gs using the following equations (4) and (5). Specifically, firstly, the muscle mass extraction unit 35 extracts the muscle rate rm(x,y) at each pixel position (x,y) within the soft image Gs using equation (4). In addition, μm in equation (4) is a weighting coefficient corresponding to the attenuation coefficient of muscle tissue, and μf is a weighting coefficient corresponding to the attenuation coefficient of adipose tissue. Furthermore, Δ(x,y) represents the concentration difference distribution. The concentration difference distribution refers to the distribution of concentration changes observed on the image from the concentration obtained by reaching the first radiation detector 5 and the second radiation detector 6 without radiation penetrating the subject H. The distribution of concentration changes on the image is calculated by subtracting the concentration of each pixel in the region of the subject H from the concentration in the directly exposed area obtained by directly irradiating the first radiation detector 5 and the second radiation detector 6 from the concentration in the directly exposed area of ​​the soft image Gs.

[0093] rm(x, y)={μf-Δ(x, y) / T(x, y)} / (μf-μm) (4)

[0094] Furthermore, the muscle volume extraction unit 35 generates a muscle image Gm from the soft tissue image Gs using the following formula (5). In addition, x and y in formula (5) are the coordinates of each pixel in the muscle image Gm.

[0095] Gm(x,y)=rm(x,y)×Gs(x,y) (5)

[0096] Furthermore, as shown in equation (6) below, the muscle mass derivation unit 35 derives the muscle mass M(x,y) (g / cm2) of each pixel of the muscle image Gm by multiplying each pixel (x,y) of the muscle image Gm by a predefined coefficient C1(x,y) representing the relationship between the pixel value and the muscle mass.

[0097] M(x,y)=C1(x,y)×Gm(x,y) (6)

[0098] Furthermore, in this embodiment, the target bone is the femur, therefore, as Figure 13 As shown, muscle mass M can be derived only in the region R3 surrounding the femur in the muscle image Gm.

[0099] Furthermore, the method for deriving muscle mass is not limited to the above-mentioned method. For example, as described in Patent Document 3, muscle mass can also be determined based on the body thickness distribution and the pixel values ​​of the soft tissue image Gs.

[0100] The probability derivation unit 24 derives the probability of developing a musculoskeletal disease related to the target bone based on the bone mineral content, muscle mass surrounding the target bone, shape information of the target bone, and shape information of bones adjacent to the target bone. Therefore, if the bone mineral content, muscle mass surrounding the target bone, shape information of the target bone, and shape information of bones adjacent to the target bone are input, the probability derivation unit 24 uses a learning completion neural network 24A that outputs the probability of developing a musculoskeletal disease related to the target bone to derive the probability of developing such a disease.

[0101] Learning Department 25 constructs a learning neural network 24A by using the bone salt content of the target bone, the muscle mass around the target bone, the shape information representing the shape of the target bone, the shape information representing the shape of the bones adjacent to the target bone, and the positive solution data representing the probability of the occurrence of musculoskeletal diseases related to the target bone as teacher data to perform machine learning on the neural network.

[0102] Examples of neural networks include simple perceptrons, multilayer perceptrons, deep neural networks, convolutional neural networks, deep belief networks, recurrent neural networks, and probabilistic neural networks. In this embodiment, a convolutional neural network is used as the neural network.

[0103] Figure 14 This is a diagram representing the neural network used in this embodiment. For example... Figure 14 As shown, the neural network 60 includes an input layer 61, an intermediate layer 62, and an output layer 63. The intermediate layer 62 includes, for example, multiple convolutional layers 65, multiple pooling layers 66, and a fully connected layer 67. In the neural network 60, the fully connected layer 67 is located before the output layer 63. Furthermore, in the neural network 60, the convolutional layers 65 and pooling layers 66 are alternately arranged between the input layer 61 and the fully connected layer 67.

[0104] Furthermore, the structure of neural network 60 is not limited to Figure 14 For example, neural network 60 may also have a convolutional layer 65 and a pooling layer 66 between the input layer 61 and the fully connected layer 67.

[0105] Figure 15 This is a diagram illustrating an example of teacher data used in the learning process of a neural network. For example... Figure 15As shown, teacher data 40 consists of learning data 41 and correct answer data 42. Learning data 41 consists of the amount of bone minerals in the target bone (bone mineral content) 43, the amount of muscle around the target bone (muscle mass) 44, the shape information of the target bone 45, and the shape information of the bones adjacent to the target bone 46. Correct answer data 42 consists of the probability of fractures (as a disease of the locomotor organ) 47 and the probability of dislocation (hip joint dislocation) 48.

[0106] For multiple patients, the teacher data is derived by statistically recording fracture and dislocation occurrences based on the patient's bone mineral content, muscle mass, target bone shape information, and the shape information of adjacent bones, and is stored in the image storage system 9. The positive solution data 42 in the teacher data 40, i.e., the probability of fracture and dislocation, can be calculated by determining the number of cases of fracture and dislocation after a predetermined number of years (e.g., 1 year, 2 years, or 5 years) for multiple patients with similar bone mineral content, muscle mass, target bone shape information, and the shape information of adjacent bones, and then dividing the calculated number of cases by the number of patients.

[0107] The learning department 25 used multiple teacher data 40 to learn the neural network. Figure 16 This is a diagram used to illustrate the learning process of the neural network 60. During the learning process of the neural network 60, the learning unit 25 inputs learning data 41 into the input layer 61 of the neural network 60. Then, the learning unit 25 outputs the probability of occurrence of musculoskeletal diseases, i.e., fractures and dislocations, from the output layer 63 of the neural network 60 as output data 70. Then, the learning unit 25 derives the difference between the probability of occurrence contained in the output data 70 and the correct solution data 42 as the loss L0.

[0108] The learning unit 25 learns from the neural network 60 based on the loss L0. Specifically, the learning unit 25 adjusts the kernel coefficients in the convolutional layer 65, the connection weights between layers, and the connection weights in the fully connected layer 67 (hereinafter referred to as parameters 71) in a way that reduces the loss L0. For example, backpropagation can be used to adjust parameters 71. The learning unit 25 repeats the adjustment of parameters 71 until the loss L0 falls below a predetermined threshold. Thus, when given inputs of bone mineral content, muscle mass, shape information of the target bone, and shape information of bones adjacent to the target bone, parameters 71 are adjusted to output a more accurate probability of fracture and dislocation, thereby constructing a learned neural network 24A. The constructed learned neural network 24A is stored in the memory 13.

[0109] If the patient's bone mineral content, muscle mass, the shape information of the target bone, and the shape information of the bones adjacent to the target bone are input into the learning neural network 24A constructed in this way, then the learning neural network 24A will output the probability of femoral fracture and the probability of hip dislocation for that patient.

[0110] Additionally, the bone mineral content and muscle mass contained in the learning data 41 are input into the neural network 60 during learning. For example, regarding bone mineral content, the input... Figure 8 The representative value of bone mineral content within the femoral region of region R1, which includes the joint portion of the femur, is shown. The representative value can be set to an average, maximum, minimum, or median value. Furthermore, all values ​​of bone mineral content within the femoral region of region R1 can be input, or the bone mineral content at multiple predefined points within the femoral region of region R1 can be input. Additionally, when the probability derivation unit 24 derives the probability of musculoskeletal diseases, the same bone mineral content as in the case where the neural network 60 has been trained is input into the trained neural network 24A.

[0111] Furthermore, regarding muscle mass, during study... Figure 13 Representative values ​​of the muscle mass around the femur in region R3, which includes the joint portion of the femur, are input into neural network 60. These representative values ​​can be set to average, maximum, minimum, or median values. Furthermore, all values ​​of muscle mass within region R3 around the femur can be input, or muscle mass at multiple predefined points within region R3 can be input. In this case, when the probability derivation unit 24 derives the probability of musculoskeletal disease, the same muscle mass as in the case where neural network 60 has been learned is input into the learned neural network 24A.

[0112] The shape information of the object bone and the shape information of the bones adjacent to the object bone are used to input a binary image representing the shape information into the learning completed neural network 24A.

[0113] The display control unit 26 displays the probability of occurrence of musculoskeletal diseases derived by the probability derivation unit 24 on the display screen 14.

[0114] Figure 17 This is a display screen showing the probability of the occurrence of diseases of the musculoskeletal system. For example... Figure 17As shown, the display screen 50 shows a first graph 51 showing the relationship between average bone mineral content and the probability of fracture occurrence, and a second graph 52 showing the relationship between average muscle mass and the probability of dislocation occurrence. In the first graph 51, the horizontal axis represents average bone mineral content, and the vertical axis represents the probability of fracture occurrence; the smaller the average bone mineral content, the higher the probability of fracture occurrence. Furthermore, the first graph 51 includes a plot 51A with a white circle representing the current probability of fracture occurrence derived by the probability derivation unit 24, and a plot 51B with a star-shaped symbol representing a target value for the probability. The target value is the probability of halving the probability of musculoskeletal diseases derived by the probability derivation unit 24.

[0115] Furthermore, option 53 for medical interventions to achieve target bone mineral density is shown on the right side of Figure 51. Figure 17 The options displayed are "Administer Drug A" and "Perform Exercise B". Drug A is the name of the drug used to achieve the target bone mineral content. Exercise B is an exercise used to strengthen the muscles related to the target bone. Specifically, it could be squats or similar exercises to strengthen the muscles around the hip joint.

[0116] Furthermore, in Figure 52, the horizontal axis represents average muscle mass, and the vertical axis represents the probability of dislocation; the smaller the average muscle mass, the higher the probability of dislocation. Figure 52 also includes a white circle plot 52A representing the current probability of dislocation and a star-shaped plot 52B representing the target probability of a change in muscle mass. The target value is the probability of halving the probability of musculoskeletal diseases derived by the probability derivation unit 24.

[0117] Furthermore, option 54 for medical interventions to achieve target muscle mass is shown on the right side of Figure 52. Figure 17 The options displayed are "Supplement Drug C" and "Perform Exercise D". Drug C is the name of a drug used to achieve the target muscle mass. Exercise D is an exercise used to train the muscles related to the target's bones. Specifically, it includes exercises such as squats that train the muscles around the hip joint.

[0118] Furthermore, in this embodiment, a table defining the relationship between average bone mineral content and / or average muscle mass and the probability of occurrence of musculoskeletal diseases, based on patient information such as age, gender, height, weight, and fracture history of the patient as subject H, is stored in memory 13. The display control unit 26 displays the first graph 51 and the second graph 52 with reference to this table.

[0119] Furthermore, in addition to or in place of Chart 51 and Chart 52, charts showing the probability of fracture relative to average muscle mass and the probability of dislocation relative to average bone mineral mass may be displayed.

[0120] Furthermore, the display control unit 26 can be displayed on the display screen 50 in a manner that allows the user to select "Perform Movement B" and "Perform Movement D" from options 53 and 54 by clicking. In this case, if the operator selects "Perform Movement B" or "Perform Movement D", then... Figure 18 As shown, the display control unit 26 can use another window 56 to display dynamic images of movements used to exercise muscles related to the object's bones.

[0121] Next, the processing performed in this embodiment will be explained. Figure 19 This is a flowchart illustrating the learning process performed in this embodiment. First, the information acquisition unit 22 acquires teacher data from the image storage system 9 (step ST1). The learning unit 25 inputs the learning data 41 contained in the teacher data 40 into the neural network 60 and outputs the probability of occurrence of motor organ diseases. It then learns the neural network 60 using a loss L0 based on the difference from the correct answer data 42 (step ST2) and returns to step ST1. Next, the learning unit 25 repeats steps ST1 and ST2 until the loss L0 reaches a predetermined threshold, and then ends the learning process. Alternatively, the learning unit 25 can end the learning process by repeating it a predetermined number of times. Thus, the learning unit 25 constructs a learning-completed neural network 24A.

[0122] Next, the predictive processing for diseases of the musculoskeletal system in this embodiment will be explained. Figure 20 This is a flowchart illustrating the musculoskeletal disease prediction process in this embodiment. Furthermore, the first radiographic image G1 and the second radiographic image G2 are acquired by imaging and stored in the memory 13. If an instruction to start processing is input from the input device 15, the image acquisition unit 21 acquires the first radiographic image G1 and the second radiographic image G2 from the memory 13 (acquiring radiographic images; step ST11). Next, the scattering component is removed from the first radiographic image G1 and the second radiographic image G2 by the scattering component of the information export unit 23 (step ST12). Furthermore, the image export unit 32 exports the bone image Gb (extracted from the bone portion) and the soft tissue image Gs (extracted from the soft tissue portion) of the subject H from the first radiographic image G1 and the second radiographic image G2 (step ST13). Additionally, the segmentation unit 33 segments the bone image Gb into the femoral region, pelvic region, and vertebral region, which are the target bones (step ST14).

[0123] Next, the bone mineral content exporting unit 34 exports the bone mineral content per pixel of the bone image Gb (step ST15), and the muscle content exporting unit 35 exports the muscle image Gm from the soft tissue image Gs and exports the muscle content per pixel of the muscle image Gm (step ST16).

[0124] Additionally, the probability derivation unit 24 uses a learning-complete neural network 24A to derive the probability of a musculoskeletal disease related to the target bone based on the amount of bone salts in the target bone, the amount of muscle surrounding the target bone, the shape information of the target bone, and the shape information of bones adjacent to the target bone (step ST17). Then, the display control unit 26 displays the probability of the occurrence of the musculoskeletal disease derived by the probability derivation unit 24 on the display 14 (step ST18) and ends the processing.

[0125] Thus, in this embodiment, the probability of musculoskeletal diseases related to the target bone is derived based on the bone mineral content, the muscle mass surrounding the target bone, the shape information of the target bone, and the shape information of the bones adjacent to the target bone. Here, femoral fractures and hip dislocations are prone to occur due to a decrease in bone mineral content and muscle mass, but also due to deformation of the hip joint from its normal state. In this embodiment, the shape information of the femur (the target bone) and the shape information of the pelvis adjacent to the target bone are also used to derive the probability of musculoskeletal diseases, thus enabling more accurate prediction of their occurrence.

[0126] Furthermore, by displaying the probability of musculoskeletal diseases, it is easy to identify the probability of musculoskeletal diseases occurring in the current situation. In particular, by also displaying the probability of musculoskeletal diseases occurring when bone mineral content and muscle mass reach target values, it is easy to identify how much bone mineral content and muscle mass should be increased.

[0127] Furthermore, by displaying options for medical interventions to achieve target values ​​for bone mineral density and muscle mass, it is easy to determine which medication should be administered to the patient or which exercise should be recommended.

[0128] Furthermore, in the above embodiments, the probability of occurrence of fractures and dislocations is derived from the probability of occurrence of diseases of the musculoskeletal system, but the probability of occurrence of either fracture or dislocation can also be derived.

[0129] Furthermore, in the above embodiment, the femur was used as the target bone, but it is not limited to this. Vertebrae, especially lumbar vertebrae, can also be used as target bones. Vertebrae adjacent to the vertebrae can be vertebrae adjacent to the upper side of the vertebrae used as target bones, vertebrae adjacent to the lower side, and vertebrae adjacent to both the upper and lower sides.

[0130] Vertebrae, in particular, experience a decrease in bone mineral content due to osteoporosis. If osteoporosis worsens, the vertebrae are compressed and deformed in the vertical direction of the body, leading to compression fractures. Furthermore, dislocation is rarely considered when the vertebrae are the target bone. Therefore, by using the shape information of the target vertebra and adjacent vertebrae when the target bone is the vertebra, the probability of fracture can be predicted with greater accuracy. Moreover, when the target bone is the vertebra, exercises to strengthen the back muscles are preferably displayed as the medical intervention to be shown.

[0131] Furthermore, in this embodiment, in addition to the femur and vertebrae, any bone around the knee joint, such as the femur and tibia, can be used as the target bone.

[0132] Furthermore, in the above embodiment, the amount of bone minerals and muscle mass were derived using the first radiographic image G1 and the second radiographic image G2 themselves, but this is not a limitation. Alternatively, a moving average of each pixel in the first radiographic image G1 and the second radiographic image G2 with its surrounding pixels can be calculated, and the amount of bone minerals and muscle mass can be derived using the first radiographic image G1 and the second radiographic image G2 with the moving average as the pixel value for each pixel. Here, cortical bone becomes important information when determining the amount of bone minerals; therefore, to maintain a resolution that allows cortical bone identification, for example, a resolution of 2 mm or less based on the actual size of the subject, a moving average of each pixel with its surrounding pixels can be calculated. In this case, the pixels used in the moving average can be appropriately determined based on information about the distances between the radiation source 3, the subject H, and the radiation detectors 5 and 6, as well as information about the pixel sizes of the radiation detectors 5 and 6.

[0133] Furthermore, in the above embodiment, during energy subtraction processing, the first radiation image G1 and the second radiation image G2 were obtained using a single irradiation method, but this is not a limitation. Figure 21As shown, the first radiation image G1 and the second radiation image G2 can also be obtained by a so-called two-irradiation method, which involves taking two images using only one radiation detector. In the case of the two-irradiation method, the position of the subject H included in the first radiation image G1 and the second radiation image G2 may shift due to movement of the subject H. Therefore, in the first radiation image G1 and the second radiation image G2, it is preferable to perform the processing of this embodiment after aligning the subject. For example, the method described in Japanese Patent Application Publication No. 2011-255060 can be used for the alignment processing. The method described in Japanese Patent Application Publication No. 2011-255060 generates multiple first-band images and multiple second-band images representing structures with different frequency bands, respectively, for a first radiation image G1 and a second radiation image G2. The method also obtains the position offsets of corresponding positions in the first-band images and second-band images of the corresponding frequency bands, and aligns the first radiation image G1 and the second radiation image G2 according to the position offsets.

[0134] Furthermore, in the above embodiment, the radiation images acquired in the system that uses the first radiation image G1 and the second radiation image G2 of the subject H to be captured by the first radiation detector 5 and the second radiation detector 6 were used for predictive processing of musculoskeletal diseases. However, the technology of the present invention can also be applied to the case where a fluorophore is used instead of a radiation detector to acquire the first radiation image G1 and the second radiation image G2. In this case, two fluorophores are overlapped and the subject H is irradiated with radiation, and the radiation image information of the subject H is accumulated and recorded in each fluorophore. The first radiation image G1 and the second radiation image G2 are acquired by photoelectrically reading the radiation image information from each fluorophore. In addition, when using a fluorophore to acquire the first radiation image G1 and the second radiation image G2, a two-stage irradiation method can also be used.

[0135] Furthermore, in the above embodiment, the target values ​​for bone mineral content and muscle mass are plotted on the first chart 51 and the second chart 52 on the display screen showing the probability of musculoskeletal diseases, but this is not a limitation. The probability of musculoskeletal diseases occurring when bone mineral content and muscle mass decrease without any further treatment, for example, when bone mineral content and muscle mass decrease to 1 / 4, can also be plotted on the first chart 51 and the second chart 52. This allows patients to maintain motivation for treatment and exercise.

[0136] Furthermore, the radiation used in the above embodiments is not particularly limited; in addition to X-rays, gamma rays or γ rays can also be used.

[0137] Furthermore, in the above embodiments, the hardware structure of processing units that perform various processes, such as image acquisition unit 21, information acquisition unit 22, information derivation unit 23, probability derivation unit 24, learning unit 25, and display control unit 26, can use various processors as shown below. These various processors include, in addition to general-purpose processors (CPUs) that execute software (programs) to function as various processing units as described above, processors such as FPGAs (Field Programmable Gate Arrays) whose circuit structure can be changed after manufacturing (i.e., Programmable Logic Devices (PLDs)) and ASICs (Application Specific Integrated Circuits) which have circuit structures specifically designed for performing specific processes (i.e., dedicated circuits).

[0138] A processing unit can consist of one of these various processors, or it can consist of a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs or a combination of a CPU and an FPGA). Furthermore, a single processor can also constitute multiple processing units.

[0139] As examples of a single processor comprising multiple processing units, there are two main approaches: First, as exemplified by computers such as client machines and servers, a single processor is constructed using a combination of one or more CPUs and software, and this processor functions as multiple processing units. Second, as exemplified by System-on-Chip (SoC), a processor that implements the overall system functionality including multiple processing units is implemented using a single integrated circuit (IC) chip. In this way, various processing units are constructed as hardware using one or more of the aforementioned processors.

[0140] Furthermore, the hardware architecture of these various processors, more specifically, can utilize circuits composed of circuit elements such as semiconductor components.

[0141] Symbol Explanation

[0142] 1-Photographic device, 3-Radiation source, 5, 6-Radiation detector, 7-Radiation energy conversion filter, 9-Image storage system, 10-Motor organ disease prediction device, 11-CPU, 12A-Motor organ disease prediction processing program, 12B-Learning program, 13-Memory, 14-Display, 15-Input device, 16-Internal memory, 17-Network I / F, 18-Bus, 21-Image acquisition unit, 22-Information acquisition unit, 23-Information export unit, 24-Probability export unit, 24A-Learning completed neural network, 25-Learning unit, 26-Display control unit, 31-Scattered radiation removal unit, 32-Image export unit, 33-Segmentation unit, 40-Teacher data, 41-Learning data, 42-Forward solution data, 43-Bone mineral content of the target bone, 44- 45 - Muscle mass around the target bone; 46 - Shape information of the target bone; 47 - Shape information of bones adjacent to the target bone; 48 - Probability of fracture; 59 - Probability of dislocation; 50 - Display screen; 51 - First chart; 51A, 51B, 52A, 52B - Plotting diagrams; 51 - Second chart; 52, 53, 54 - Medical intervention options; 56 - Window; 60 - Neural network; 61 - Input layer; 62 - Intermediate layer; 63 - Output layer; 65 - Convolutional layer; 66 - Pooling layer; 67 - Fully connected layer; 70 - Output data; 71 - Parameters; A1 - Femoral region; A2 - Pelvic region; A3 - Vertebral region; G1 - First radiographic image; G2 - Second radiographic image; Gb - Bone image; Gs - Soft tissue image; LUT1 - Lookup table; S1, S2 - Shape information.

Claims

1. A device for predicting diseases of the musculoskeletal system, comprising at least one processor, The processor derives, based on a first radiographic image and a second radiographic image acquired by photographing a subject containing bones and soft tissues using radiation with different energy distributions, the amount of bone salts in the bone of the subject, the amount of muscle surrounding the bone, shape information representing the shape of the bone, and shape information representing the shape of bones adjacent to the bone. The probability of developing musculoskeletal diseases related to the target bone is derived based on the amount of bone salts in the target bone, the amount of muscle surrounding the target bone, the shape information of the target bone, and the shape information of the bones adjacent to the target bone.

2. The musculoskeletal disease prediction device according to claim 1, wherein, The processor functions as a learning-complete neural network, which uses positive solution data representing the amount of bone salts in the target bone, the amount of muscle around the target bone, shape information representing the shape of the target bone, shape information representing the shape of the bones adjacent to the target bone, and the probability of developing musculoskeletal diseases related to the target bone as teacher data for machine learning.

3. The musculoskeletal disease prediction device according to claim 1 or 2, wherein, The processor displays the derived probability of occurrence of the musculoskeletal disease on a monitor.

4. The musculoskeletal disease prediction device according to claim 3, wherein, The processor displays a graph showing the relationship between at least one of bone mineral content and muscle mass and the probability of developing diseases of the musculoskeletal system. The graph also displays a plot representing the probability of occurrence of the derived musculoskeletal disease and a plot representing the value obtained by changing at least one of the probability of occurrence or bone mineral content and muscle mass.

5. The musculoskeletal disease prediction device according to claim 4, wherein, The value obtained by the change is a target value for at least one of the bone mineral content and the muscle mass, or a target value for the probability of occurrence of the musculoskeletal disease.

6. The musculoskeletal disease prediction device according to claim 5, wherein, The processor also displays options for medical interventions to bring at least one of the bone mineral content and the muscle mass to the target value, or options for medical interventions to bring the musculoskeletal disease to the target value.

7. The musculoskeletal disease prediction device according to claim 6, wherein, The medical intervention is an exercise method used to strengthen the muscles associated with the bones of the subject.

8. The musculoskeletal disease prediction device according to claim 1 or 2, wherein, The bone in question is the femur.

9. The musculoskeletal disease prediction device according to claim 1 or 2, wherein, The bone in question is a vertebra.

10. The musculoskeletal disease prediction device according to claim 1 or 2, wherein, The musculoskeletal disease is at least one of fracture and dislocation.

11. A learning device comprising at least one processor, The processor constructs a learning-complete neural network by using the positive solution data derived from the musculoskeletal disease prediction device of claim 1—which includes the amount of bone salts in a target bone, the amount of muscle surrounding the target bone, shape information representing the shape of the target bone, shape information representing the shape of bones adjacent to the target bone, and the probability of a musculoskeletal disease associated with the target bone—as teacher data to perform machine learning on the neural network. This learning-complete neural network outputs the probability of a musculoskeletal disease occurring if it is input with the amount of bone salts in the target bone, the amount of muscle surrounding the target bone, the shape information of the target bone, and the shape information of bones adjacent to the target bone.

12. A computer-executed method for predicting diseases of the musculoskeletal system, wherein, Based on the first and second radiographic images obtained by photographing a subject containing bones and soft tissues using radiation with different energy distributions, the following information is derived: the amount of bone salts in the bone of the subject, the amount of muscle surrounding the bone, shape information representing the shape of the bone, and shape information representing the shape of the bones adjacent to the bone. The probability of developing musculoskeletal diseases related to the target bone is derived based on the amount of bone salts in the target bone, the amount of muscle surrounding the target bone, the shape information of the target bone, and the shape information of the bones adjacent to the target bone.

13. A learning method, wherein, By using the positive solution data derived from the musculoskeletal disease prediction device of claim 1—including the amount of bone salts in a target bone, the amount of muscle surrounding the target bone, shape information representing the shape of the target bone, shape information representing the shape of bones adjacent to the target bone, and the probability of a musculoskeletal disease associated with the target bone—as teacher data to perform machine learning on a neural network, a learning-complete neural network is constructed that, given the input of the amount of bone salts in the target bone, the amount of muscle surrounding the target bone, the shape information of the target bone, and the shape information of bones adjacent to the target bone, outputs the probability of a musculoskeletal disease occurring.

14. A computer-readable storage medium storing a musculoskeletal disease prediction program, the musculoskeletal disease prediction program being configured to cause a computer to perform the following processing: Based on the first and second radiographic images obtained by photographing a subject containing bones and soft tissues using radiation with different energy distributions, the following are derived: the amount of bone minerals in the bone of the subject; the amount of muscle surrounding the bone; shape information representing the shape of the bone; and shape information representing the shape of bones adjacent to the bone. The probability of developing musculoskeletal diseases related to the target bone is derived based on the amount of bone salts in the target bone, the amount of muscle surrounding the target bone, the shape information of the target bone, and the shape information of the bones adjacent to the target bone.

15. A computer-readable storage medium storing a learning program, the learning program being configured to cause a computer to perform the following processes: By using the positive solution data derived from the musculoskeletal disease prediction device of claim 1—including the amount of bone salts in a target bone, the amount of muscle surrounding the target bone, shape information representing the shape of the target bone, shape information representing the shape of bones adjacent to the target bone, and the probability of a musculoskeletal disease associated with the target bone—as teacher data to perform machine learning on a neural network, a learning-complete neural network is constructed that, given the input of the amount of bone salts in the target bone, the amount of muscle surrounding the target bone, the shape information of the target bone, and the shape information of bones adjacent to the target bone, outputs the probability of a musculoskeletal disease occurring.

Images