Estimation device, method, and program

Abstract

To accurately predict a locomotorium disease in an estimation device, method, and program.SOLUTION: An estimation device includes at least one processor. The processor functions as a learned neural network for deriving an estimation result related to bone density of a bone part from a plain radiation image acquired by executing plain radiography of a subject including the bone part or a DXA scan image acquired by imaging the subject by a DXA method. The learned neural network is learned by using, as teacher data, a composite two-dimensional image showing the subject derived by composing three-dimensional CT images of the subject, and information related to the bone density of the subject.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present disclosure relates to an estimation device, a method, and a program. [Background technology]

[0002] DXA (Dual X-ray Absorptiometry) is one of the representative bone mineral quantification methods used to diagnose bone density in bone diseases such as osteoporosis. DXA uses the attenuation coefficient μ (cm) of radiation that enters and penetrates the human body, which depends on the materials that make up the human body (e.g., bone). 2 / g) and its density ρ(g / cm 3 This method calculates bone mineral density from the pixel values ​​of radiographic images taken with two types of radiation energy, utilizing the attenuation of bone, characterized by the radiation energy (t) and thickness (cm).

[0003] Various methods for evaluating bone density using radiographic images acquired by photographing a subject have also been proposed. For example, Patent Documents 1 and 2 propose a method for estimating information related to bone density from images of bones by using a trained neural network constructed by training a neural network. In the method described in Patent Document 1, the neural network is trained using images of bones acquired by simple radiography and bone density as training data. In the method described in Patent Document 1, the neural network is trained using images of bones acquired by simple radiography, bone density, and information related to bone density (e.g., age, sex, weight, drinking habits, smoking habits, fracture history, body fat percentage, subcutaneous fat percentage, etc.) as training data.

[0004] Note that plain radiography is an imaging method in which a subject is irradiated with radiation once to obtain a single two-dimensional image that is a transmission image of the subject. In the following description, a radiographic image obtained by plain radiography will be referred to as a plain radiographic image. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] U.S. Patent No. 6,064,716 [Patent Document 2] International Publication No. 2020 / 054738 Summary of the Invention [Problem to be solved by the invention]

[0006] However, there is a demand for a more accurate estimation of bone mineral density.

[0007] The present disclosure has been made in consideration of the above circumstances, and aims to enable bone mineral density to be estimated with high accuracy. [Means for solving the problem]

[0008] An estimation device according to the present disclosure includes at least one processor, The processor functions as a trained neural network that derives an estimation result related to bone density of a bone portion from a plain radiographic image obtained by plain radiography of a subject including a bone portion or a DXA scanned image obtained by photographing the subject using a DXA method; The trained neural network is trained using a synthesized two-dimensional image representing the subject, which is derived by synthesizing three-dimensional CT images of the subject, and information related to the subject's bone density as training data.

[0009] In addition, in the estimation device according to the present disclosure, the composite two-dimensional image may be derived by deriving the radiation attenuation coefficient for the composition at each position in three-dimensional space and projecting the CT image in a predetermined direction based on the attenuation coefficient.

[0010] In addition, in the estimation device according to the present disclosure, information related to bone density may be derived by identifying a bone region in a CT image, deriving the attenuation coefficient of radiation in the bone region, and deriving the bone density at each position in the bone region based on the attenuation coefficient of radiation and the mass attenuation coefficient in the bone region.

[0011] In addition, in the estimation device according to the present disclosure, information related to bone density may be derived by projecting the bone density at each position in the bone region in a predetermined direction.

[0012] In addition, in the estimation device according to the present disclosure, the information related to bone density may include at least one of bone density per unit area, bone density per unit volume, an assessment value of the subject's fracture risk, and information representing the healing state of the bone after treatment.

[0013] In addition, in the estimation device according to the present disclosure, the processor functions as a trained neural network that derives an estimation result related to the bone mineral density of the bone portion from the DXA scan image, The trained neural network may be trained using a low-resolution composite two-dimensional image obtained by processing the composite two-dimensional image to reduce its resolution and information related to the bone density of the subject as training data.

[0014] Furthermore, in the estimation device according to the present disclosure, the low-resolution composite two-dimensional image is an image in which the pixel value of a plurality of adjacent pixels in the composite two-dimensional image is the average value of the pixel values ​​of the adjacent pixels, and the size of the adjacent pixels may correspond to the size of one pixel in the DXA scan image.

[0015] In addition, in the estimation device according to the present disclosure, the low-resolution composite two-dimensional image may be an image obtained by performing moving average processing in one direction of the composite two-dimensional image, and the one direction may correspond to the scanning direction of the DXA scan image.

[0016] Furthermore, in the estimation device according to the present disclosure, the low-resolution composite two-dimensional image may be an image generated by generating a first low-resolution image in which the pixel value of a plurality of adjacent pixels in the composite two-dimensional image is the average value of the pixel values ​​of the adjacent plurality of pixels, and performing a moving average process in one direction on the first low-resolution image, wherein the size of the adjacent plurality of pixels corresponds to the size of one pixel of the DXA scan image, and the one direction corresponds to the scanning direction of the DXA scan image.

[0017] The estimation method according to the present disclosure is a method for deriving an estimation result related to bone density from a plain radiographic image or a DXA scanned image by using a trained neural network that derives an estimation result related to bone density from a plain radiographic image obtained by simply photographing a subject including a bone portion or a DXA scanned image obtained by photographing the subject using a DXA method, the method comprising: The trained neural network is trained using a synthesized two-dimensional image representing the subject, derived by synthesizing three-dimensional CT images of the subject, and information related to the bone density of the subject, as training data.

[0018] The estimation method according to the present disclosure may be provided as a program for causing a computer to execute the method. [Effects of the Invention]

[0019] According to the present disclosure, bone density can be estimated with high accuracy. [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 1 is a schematic block diagram showing the configuration of a radiographic imaging system to which an estimation device according to a first embodiment of the present disclosure is applied. [Figure 2] FIG. 1 is a diagram showing a schematic configuration of an estimation device according to a first embodiment. [Figure 3] FIG. 1 is a diagram showing a functional configuration of an estimation device according to a first embodiment. [Figure 4] FIG. 1 is a diagram showing a schematic configuration of a neural network used in this embodiment. [Figure 5] Diagram showing training data [Figure 6] FIG. 1 is a diagram showing a schematic configuration of an information derivation device according to a first embodiment; [Figure 7] FIG. 1 is a diagram showing a functional configuration of an information derivation device according to a first embodiment; [Figure 8] A diagram for explaining the derivation of a synthetic 2D image. [Figure 9] A diagram for explaining the derivation of a synthetic 2D image. [Figure 10] Diagram to explain CT value [Figure 11] Diagram showing the relationship between radiation energy and mass attenuation coefficient [Figure 12] Diagram to explain neural network learning [Figure 13] Conceptual diagram of the processing performed by a trained neural network [Figure 14] Figure showing the display screen of the estimation results [Figure 15] 1 is a flowchart of a learning process performed in the first embodiment. [Figure 16] 1 is a flowchart of an estimation process performed in the first embodiment. [Figure 17] Another example of training data [Figure 18] FIG. 10 is a diagram showing a functional configuration of an information derivation device according to a second embodiment. [Figure 19] Graph showing the relationship between statistical values ​​and the probability of fracture occurrence within 10 years [Figure 20] Another example of training data [Figure 21] FIG. 10 is a diagram showing a functional configuration of an information derivation device according to a third embodiment. [Figure 22] A diagram showing an example of an artificial bone embedded in the bone of a subject. [Figure 23] Graph showing an example of the relationship between the distance from the stem inside the femur and the bone mineral content at each stage after surgery. [Figure 24] A cross-sectional view showing an example of the cross-sectional structure of a human bone [Figure 25] Another example of training data [Figure 26] FIG. 10 is a diagram showing another example of the display screen for the estimation results. DETAILED DESCRIPTION OF THE INVENTION

[0021] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. Fig. 1 is a schematic block diagram showing the configuration of a radiographic image capturing system to which an estimation device according to a first embodiment of the present disclosure is applied. As shown in Fig. 1, the radiographic image capturing system according to the first embodiment includes an imaging device 1, a CT device 7, an image storage system 9, an estimation device 10 according to the first embodiment, and an information derivation device 50. The imaging device 1, the CT (Computed Tomography) device 7, the estimation device 10, and the information derivation device 50 are connected to the image storage system 9 via a network (not shown).

[0022] The imaging device 1 is an imaging device that can acquire a simple radiographic image G0 of the subject H by irradiating a radiation detector 5 with radiation such as X-rays that are emitted from a radiation source 3 and transmitted through the subject H. The acquired simple radiographic image G0 is input to the estimation device 10. The simple radiographic image G0 is, for example, a front image including the groin area of ​​the subject H.

[0023] The radiation detector 5 is capable of repeatedly recording and reading out radiation images, and may be a so-called direct type radiation detector that generates electric charges upon direct exposure to radiation, or a so-called indirect type radiation detector that converts radiation into visible light and then converts the visible light into an electric charge signal. The radiation image signal readout method is preferably a TFT readout method in which the radiation image signal is read out by turning a TFT (thin film transistor) switch on and off, or an optical readout method in which the radiation image signal is read out by irradiating the detector with readout light, but is not limited to these, and other methods may also be used.

[0024] The CT device 7 acquires a plurality of tomographic images representing a plurality of tomographic planes of the subject H as a three-dimensional CT image V0. The CT value of each pixel (voxel) in the CT image is a numerical representation of the radiation absorption rate of the components that make up the human body. The CT value will be described later.

[0025] The image storage system 9 is a system that stores image data of radiographic images acquired by the radiography device 1 and image data of CT images acquired by the CT device 7. The image storage system 9 extracts images from the stored radiographic images and CT images in response to requests from the estimation device 10 and the information derivation device 50, and transmits them to the requesting device. A specific example of the image storage system 9 is a PACS (Picture Archiving and Communication Systems). In this embodiment, the image storage system 9 stores a large amount of training data for training a neural network, which will be described later.

[0026] Next, an estimation device according to a first embodiment will be described. First, with reference to FIG. 2, the hardware configuration of the estimation device according to the first embodiment will be described. As shown in FIG. 2, the estimation device 10 is a computer such as a workstation, a server computer, or a personal computer, and includes a CPU (Central Processing Unit) 11, non-volatile storage 13, and memory 16 as a temporary storage area. The estimation device 10 also includes a display 14 such as a liquid crystal display, an input device 15 such as a keyboard and a mouse, and a network I / F (Interface) 17 connected to a network (not shown). The CPU 11, the storage 13, the display 14, the input device 15, the memory 16, and the network I / F 17 are connected to a bus 18. The CPU 11 is an example of a processor in the present disclosure.

[0027] The storage 13 is realized by a hard disk drive (HDD), a solid state drive (SSD), a flash memory, etc. The storage 13 as a storage medium stores the estimation program 12A and the learning program 12B installed in the estimation device 10. The CPU 11 reads out the estimation program 12A and the learning program 12B from the storage 13, expands them in the memory 16, and executes the expanded estimation program 12A and the learning program 12B.

[0028] The estimation program 12A and the learning program 12B are stored in a state accessible from the outside in a storage device of a server computer connected to a network or in a network storage, and are downloaded and installed in response to a request into a computer constituting the estimation device 10. Alternatively, they are recorded on a recording medium such as a DVD (Digital Versatile Disc) or a CD-ROM (Compact Disc Read Only Memory) and distributed, and are installed from the recording medium into a computer constituting the estimation device 10.

[0029] Next, the functional configuration of the estimation device according to the first embodiment will be described. Fig. 3 is a diagram showing the functional configuration of the estimation device according to the first embodiment. As shown in Fig. 3, the estimation device 10 includes an image acquisition unit 21, an information acquisition unit 22, an estimation unit 23, a learning unit 24, and a display control unit 25. The CPU 11 executes an estimation program 12A to function as the image acquisition unit 21, the information acquisition unit 22, the estimation unit 23, and the display control unit 25. The CPU 11 executes a learning program 12B to function as the learning unit 24.

[0030] The image acquisition unit 21 causes the imaging device 1 to perform simple imaging of the subject H, thereby acquiring from the radiation detector 5 a simple radiographic image G0, which is, for example, a frontal image of the area around the crotch of the subject H. When acquiring the simple radiographic image G0, imaging conditions are set, such as the imaging dose, radiation quality, tube voltage, SID (Source Image Receptor Distance) which is the distance between the radiation source 3 and the surface of the radiation detector 5, SOD (Source Object Distance) which is the distance between the radiation source 3 and the surface of the subject H, and the presence or absence of an anti-scatter grid.

[0031] The imaging conditions may be set by the operator through input device 15. The set imaging conditions are stored in storage 13. The plain radiographic image G0 and the imaging conditions are also transmitted to and stored in image storage system 9.

[0032] In this embodiment, the simple radiographic image G0 may be acquired by a program separate from the estimation program 12A and stored in the storage 13. In this case, the image acquisition unit 21 acquires the simple radiographic image G0 stored in the storage 13 by reading it from the storage 13 for processing.

[0033] The information acquisition unit 22 acquires training data for training a neural network, which will be described later, from the image storage system 9 via the network I / F 17.

[0034] The estimation unit 23 derives an estimation result related to the bone density of the bone portion included in the subject H from the simple radiographic image G0. In this embodiment, the estimation result related to the bone density is an estimation result of the bone density of the target bone in the bone portion region included in the simple radiographic image G0. To this end, the estimation unit 23 derives the estimation result related to the bone density using a trained neural network 23A that outputs bone density when the simple radiographic image G0 is input.

[0035] The learning unit 24 constructs a trained neural network 23A by machine learning a neural network using training data. Examples of neural networks include a simple perceptron, a multilayer perceptron, a deep neural network, a convolutional neural network, a deep belief network, a recurrent neural network, and a probabilistic neural network. In this embodiment, a convolutional neural network is used as the neural network.

[0036] Fig. 4 is a diagram showing a neural network used in this embodiment. As shown in Fig. 4, the neural network 30 includes an input layer 31, an intermediate layer 32, and an output layer 33. The intermediate layer 32 includes, for example, a plurality of convolutional layers 35, a plurality of pooling layers 36, and a fully connected layer 37. In the neural network 30, the fully connected layer 37 is located before the output layer 33. In the neural network 30, the convolutional layers 35 and the pooling layers 36 are alternately arranged between the input layer 31 and the fully connected layer 37.

[0037] The configuration of the neural network 30 is not limited to the example shown in Fig. 4. For example, the neural network 30 may include one convolutional layer 35 and one pooling layer 36 between the input layer 31 and the fully connected layer 37.

[0038] 5 is a diagram showing an example of training data used for training a neural network. As shown in FIG. 5, training data 40 consists of training data 41 and correct answer data 42. In this embodiment, the data input to trained neural network 23A to obtain a bone mineral density estimation result is a plain radiographic image G0, but training data 41 includes a composite two-dimensional image C0 representing subject H derived by combining a CT image V0.

[0039] The correct answer data 42 is the bone density of the target bone (i.e., femur) of the subject from which the learning data 41 was obtained. In this embodiment, the bone density per unit area is estimated from the two-dimensional plain radiographic image G0, so the unit of bone density is (g / cm 2 ) The synthetic two-dimensional image C0, which is the learning data 41, and the bone density, which is the correct answer data 42, are derived by the information derivation device 50. The bone density, which is the correct answer data 42, is an example of information related to the bone density of the bones of the subject. The information derivation device 50 will be described below.

[0040] Fig. 6 is a schematic block diagram showing the configuration of an information derivation device according to a first embodiment. As shown in Fig. 6, information derivation device 50 according to the first embodiment is a computer such as a workstation, a server computer, or a personal computer, and includes a CPU 51, non-volatile storage 53, and memory 56 as a temporary storage area. Information derivation device 50 also includes a display 54 such as a liquid crystal display, an input device 55 including a keyboard, a pointing device such as a mouse, and a network I / F 57 connected to a network (not shown). CPU 51, storage 53, display 54, input device 55, memory 56, and network I / F 57 are connected to a bus 58.

[0041] The storage 53 is realized by an HDD, an SSD, a flash memory, or the like, similar to the storage 13. The storage 53 as a storage medium stores an information derivation program 52. The CPU 51 reads the information derivation program 52 from the storage 53, expands it in the memory 56, and executes the expanded information derivation program 52.

[0042] Next, the functional configuration of the information derivation device according to the first embodiment will be described. Fig. 7 is a diagram showing the functional configuration of the information derivation device according to the first embodiment. As shown in Fig. 7, the information derivation device 50 according to the first embodiment includes an image acquisition unit 61, a synthesis unit 62, and a bone density derivation unit 63. When the CPU 51 executes the information derivation program 52, the CPU 51 functions as the image acquisition unit 61, the synthesis unit 62, and the bone density derivation unit 63.

[0043] The image acquisition unit 61 acquires the CT image V0 for deriving the learning data 41 from the image storage system 9. Note that the image acquisition unit 61 may acquire the CT image V0 by causing the CT device 7 to capture an image of the subject H, similar to the image acquisition unit 21 of the estimation device 10.

[0044] The synthesis unit 62 synthesizes the CT images V0 to derive a synthesized two-dimensional image C0 representing the subject H. FIG. 8 is a diagram for explaining the derivation of the synthesized two-dimensional image C0. For the sake of explanation, FIG. 8 shows the three-dimensional CT image V0 in two dimensions. As shown in FIG. 8, the subject H is included in the three-dimensional space represented by the CT image V0. The subject H is composed of multiple components, including bones, fat, muscles, and internal organs.

[0045] Here, the CT value V0(x, y, z) of each pixel of the CT image V0 can be expressed by the following formula (1) using the attenuation coefficient μi of the composition at that pixel and the attenuation coefficient μw of water. (x, y, z) are coordinates that represent the pixel position of the CT image V0. In the following explanation, attenuation coefficient means the radiation source attenuation coefficient unless otherwise specified. The attenuation coefficient represents the degree (proportion) of attenuation of radiation due to absorption or scattering, etc. The attenuation coefficient differs depending on the specific composition (density, etc.) and thickness (mass) of the structure through which the radiation passes. V0(x,y,z)=(μi-μw) / μw×1000 (1)

[0046] The attenuation coefficient μw of water is known. Therefore, by solving equation (1) for μi, the attenuation coefficient μi of each composition can be calculated as shown in equation (2) below. μi=V0(x,y,z)×μw / 1000+μw (2)

[0047] As shown in FIG. 8, the synthesis unit 62 virtually irradiates the subject H with radiation at an exposure dose I0, and derives a synthetic two-dimensional image C0 by virtually detecting the radiation transmitted through the subject H using a radiation detector (not shown) installed on a virtual plane 64. The exposure dose I0 and radiation energy of the virtual radiation are set according to predetermined imaging conditions. Specifically, the incident dose I0 can be set by referring to a table prepared in advance corresponding to imaging conditions such as tube voltage, mAs value, and SID. The radiation energy can be set by referring to a table prepared in advance corresponding to tube voltage. In this case, the incident dose I1(x, y) for each pixel of the synthetic two-dimensional image C0 is transmitted through one or more components in the subject H. Therefore, the incident dose I1(x, y) can be calculated using the following equation (3) using the attenuation coefficient μi of one or more components transmitted by the radiation at the exposure dose I0. The incident dose I1(x, y) becomes the pixel value of each pixel of the synthetic two-dimensional image C0. I1(x,y)=I0×exp(-∫μi·dt) (3)

[0048] If the irradiating radiation source is assumed to be a surface light source, the attenuation coefficient μi used in equation (3) can be derived by equation (2) from the CT values ​​of each pixel arranged in the vertical direction as shown in Fig. 8. If the irradiating radiation source is assumed to be a point light source, as shown in Fig. 9, pixels on the path of the radiation that reaches each pixel can be identified based on the geometric positional relationship between the point light source and each position on a virtual plane 64, and the attenuation coefficient μi derived by equation (2) from the CT values ​​of the identified pixels can be used.

[0049] The bone density deriving unit 63 derives the bone density of the subject H for each pixel of the composite two-dimensional image C0 using the CT image V0. Here, the CT value will be explained. FIG. 10 is a diagram for explaining the CT value. The CT value is a numerical representation of the X-ray absorption rate in the human body. Specifically, as shown in FIG. 10, the CT value is determined according to the composition of the human body, with water having a CT value of 0 and air having a CT value of -1000 (unit: HU).

[0050] The bone density deriving unit 63 first identifies a bone region in the CT image V0 based on the CT value of the CT image V0. Specifically, a region consisting of pixels with a CT value of 100 to 1000 is identified as a bone region by threshold processing. Note that instead of threshold processing, the bone region may be identified using a trained neural network that has been trained to detect bone regions from the CT image V0. Alternatively, the CT image V0 may be displayed on the display 54, and the bone region may be identified by manually specifying the bone region on the displayed CT image V0.

[0051] Here, the density per unit volume of the composition at each pixel of the CT image is ρ [g / cm 3 ] is the attenuation coefficient μi [1 / cm] of the composition and the mass attenuation coefficient μe [cm] of that composition 2 / g] can be derived using the following formula (4). ρ=μi / μe (4)

[0052] FIG. 11 is a diagram showing the relationship between radiation energy and mass attenuation coefficient for various compositions of the human body. FIG. 11 also shows the relationship between radiation energy and mass attenuation coefficient for bones, muscles, etc., and fat. Note that "muscles, etc." refers to muscle, blood, and water. In this embodiment, the relationship between radiation energy and mass attenuation coefficient shown in FIG. 11 is stored as a table in storage 53. In this embodiment, the mass attenuation coefficient of bones is required, so the mass attenuation coefficient of bones is obtained by referring to the relationship for bones in the table shown in FIG. 11 based on virtual radiation energy. Furthermore, the attenuation coefficient μb for each pixel in the bone region is derived using the above formula (2). Then, the bone density ρ per unit volume for each pixel in the bone region included in the CT image V0 is derived using the above formula (4).

[0053] Since the CT image V0 is a three-dimensional image, the unit of bone density per unit volume calculated by the above formula (4) is [g / cm 3 In this embodiment, the bone density deriving unit 63 derives the bone density per unit area for each pixel of the composite two-dimensional image C0. For this reason, the bone density deriving unit 63 projects the bone density per unit volume ρ derived by the above formula (4) onto a virtual plane 64 in the same manner as when the composite two-dimensional image C0 was derived, thereby deriving the bone density per unit area B [g / cm 2 ] is derived.

[0054] During projection, a representative value of bone density of each pixel in the CT image V0 on a path from the virtual radiation source to each pixel in the composite two-dimensional image C0 may be derived. The representative value may be an integrated value, an average value, a maximum value, a median value, a minimum value, or the like. Furthermore, in this embodiment, the bone density deriving unit 63 may derive a representative value of bone density only for the target bone. For example, if the target bone is the femur, the bone density deriving unit 63 derives a representative value of bone density of the femur region by deriving a representative value of bone density of each pixel in the femur region in the composite two-dimensional image C0. The representative value may be an average value, a median value, a minimum value, a maximum value, or the like. In this embodiment, the representative value of bone density of the femur, which is the target bone, is used as the correct answer data 42.

[0055] The bone density used as the correct answer data 42 is derived at the same time as the learning data 41 is acquired and transmitted to the image storage system 9. In the image storage system 9, the learning data 41 and the correct answer data 42 are associated with each other and stored as training data 40. To improve the robustness of learning, additional training data 40 may be created and stored, including images obtained by performing at least one of the following on the same image: enlargement / reduction, contrast change, translation, in-plane rotation, inversion, and noise addition.

[0056] Returning to the estimation device 10, the learning unit 24 trains the neural network using a large amount of training data 40. FIG. 12 is a diagram for explaining the training of the neural network 30. When training the neural network 30, the learning unit 24 inputs training data 41, i.e., a synthetic two-dimensional image C0, to the input layer 31 of the neural network 30. Then, the learning unit 24 causes the output layer 33 of the neural network 30 to output the bone mineral density of the target bone as output data 47. Then, the learning unit 24 derives the difference between the output data 47 and the ground truth data 42 as the loss L0.

[0057] The learning unit 24 trains the neural network 30 based on the loss L0. Specifically, the learning unit 24 adjusts the kernel coefficients in the convolutional layer 35, the connection weights between layers, the connection weights in the fully connected layer 37, and the like (hereinafter referred to as parameters 48) so as to reduce the loss L0. The parameters 48 can be adjusted, for example, by backpropagation. The learning unit 24 repeatedly adjusts the parameters 48 until the loss L0 becomes equal to or less than a predetermined threshold. In this way, when a plain radiographic image G0 is input, the parameters 48 are adjusted so as to output the bone mineral density of the target bone, and a trained neural network 23A is constructed. The constructed trained neural network 23A is stored in the storage 13.

[0058] Fig. 13 is a conceptual diagram of the processing performed by the trained neural network 23A. As shown in Fig. 13, when a plain radiographic image G0 of a patient is input to the trained neural network 23A constructed as described above, the trained neural network 23A outputs the bone density of the target bone (i.e., the femur) included in the input plain radiographic image G0.

[0059] The display control unit 25 displays the bone density estimation results estimated by the estimation unit 23 on the display 14. FIG. 14 is a diagram showing a display screen for the estimation results. As shown in FIG. 14, a display screen 70 has an image display area 71 and a bone density display area 72. A plain radiographic image G0 of the subject H is displayed in the image display area 71. Furthermore, a representative value of the bone density around the femoral joint, based on the bone density estimated by the estimation unit 23, is displayed in the bone density display area 72.

[0060] Next, the processing performed in the first embodiment will be described. FIG. 15 is a flowchart showing the learning processing performed in the first embodiment. First, the information acquisition unit 22 acquires training data 40 from the image storage system 9 (step ST1). The learning unit 24 inputs training data 41 included in the training data 40 into the neural network 30 to output bone mineral density, and trains the neural network 30 using a loss L0 based on the difference from the ground truth data 42 (step ST2), and returns to step ST1. The learning unit 24 then repeats the processing in steps ST1 and ST2 until the loss L0 reaches a predetermined threshold value, and ends the learning processing. Note that the learning unit 24 may end the learning processing by repeating the learning a predetermined number of times. In this way, the learning unit 24 constructs a trained neural network 23A.

[0061] Next, the estimation process in the first embodiment will be described. Fig. 16 is a flowchart showing the estimation process in the first embodiment. It is assumed that the simple radiographic image G0 is acquired by radiography and stored in the storage 13. When an instruction to start the process is input from the input device 15, the image acquisition unit 21 acquires the simple radiographic image G0 from the storage 13 (step ST11). Next, the estimation unit 23 derives an estimation result related to bone density from the simple radiographic image G0 (step ST12). Then, the display control unit 25 displays the estimation result related to bone density derived by the estimation unit 23 on the display 14 together with the simple radiographic image G0 (step ST13), and the process ends.

[0062] As described above, in this embodiment, an estimation result related to the bone density of the subject H included in the plain radiographic image G0 is derived using a trained neural network 23A constructed by training using the composite two-dimensional image C0 derived from the CT image V0 and the bone density derived from the CT image V0 as training data. Here, in this embodiment, the composite two-dimensional image C0 derived from the CT image V0 and the bone density derived from the CT image V0 are used for training the neural network. Therefore, compared to using a single radiographic image and information related to bone density derived from the radiographic image as training data, the trained neural network 23A can derive an estimation result related to bone density from the plain radiographic image G0 with greater accuracy. Therefore, this embodiment can derive an estimation result related to bone density with greater accuracy.

[0063] In the first embodiment, bone density per unit area is derived as the correct answer data 42, but this is not limiting. In the first embodiment, bone density per unit volume obtained in the process of deriving bone density per unit area may be derived as the correct answer data. The bone density per unit volume can be determined by using a representative value of bone density for pixels within the region of the target bone in the CT image V0. The representative value can be the mean, median, minimum, maximum, or the like. The training data in this case is shown in FIG. 17. As shown in FIG. 17, training data 40A consists of training data 41 including a composite two-dimensional image C0 and correct answer data 42A, which is bone density per unit volume.

[0064] By training a neural network using the training data 40A shown in Figure 17, it is possible to construct a trained neural network 23A that, when a plain radiographic image G0 is input, outputs bone density per unit volume as an estimated result related to bone density.

[0065] In the above embodiments, bone density per unit area or per unit volume of the plain radiographic image G0 is estimated as information related to bone density, but this is not limiting. For example, an evaluation value of fracture risk may be derived as an estimation result related to bone density. This will be described below as a second embodiment.

[0066] Fig. 18 is a diagram showing the functional configuration of an information derivation device according to the second embodiment. In Fig. 18, the same components as those in Fig. 7 are given the same reference numerals, and detailed description thereof will be omitted. In the second embodiment of the present disclosure, instead of deriving bone density, an evaluation value of fracture risk is derived as correct answer data 42. For this purpose, as shown in Fig. 18, an information derivation device 50A according to the second embodiment further includes a muscle density derivation unit 65, a statistical value derivation unit 66, and an evaluation value derivation unit 67 in addition to the components of the information derivation device 50 according to the first embodiment.

[0067] The muscle density deriving unit 65 identifies muscle regions based on the CT values ​​in the CT image V0. Specifically, a region consisting of pixels with a CT value of 60 to 70 is identified as a muscle region by threshold processing. Note that instead of threshold processing, muscle regions may be detected using a trained neural network that has been trained to detect muscle regions from the CT image V0. Alternatively, the CT image V0 may be displayed on the display 54, and muscle regions may be identified by manually specifying muscle regions in the displayed CT image V0.

[0068] Furthermore, the muscle density deriving unit 65 calculates the muscle attenuation coefficient μm using the above formula (2), and then obtains the muscle mass attenuation coefficient by referring to the table shown in Fig. 11. Then, the muscle density per unit volume ρm is derived using the above formula (4).

[0069] The statistical value derivation unit 66 calculates a statistical value for the subject H based on the bone density derived by the bone density derivation unit 63 and the muscle density derived by the muscle density derivation unit 65. As will be described later, the statistical value is used to calculate a fracture risk evaluation value for evaluating the fracture risk. Specifically, the statistical value derivation unit 66 derives a statistical value Q based on a bone density distribution index value Bd related to the spatial distribution of bone density and a muscle mass distribution index value Md related to the spatial distribution of muscle mass, as shown in the following formula (5). Q = W1 × Bd + W2 × Md (5)

[0070] W1 and W2 in equation (5) are weighting coefficients, which are determined by collecting a large number of bone density distribution index values ​​and muscle density distribution index values ​​and performing regression analysis.

[0071] The bone density distribution index value is a value that represents how the bone density values ​​are spread out. Examples of the bone density distribution index value include the value per unit area or unit volume of bone density, the average value, the median value, the maximum value, and the minimum value. The muscle density distribution index value is a value that represents how the muscle density values ​​are spread out. Examples of the muscle density distribution index value include the value per unit area or unit volume of muscle density, the average value, the median value, the maximum value, and the minimum value.

[0072] Furthermore, the statistical value derivation unit 66 may determine the statistical value Q based on at least one of the subject's height, weight, age, fracture history, etc., in addition to the bone density and muscle density. For example, when determining the statistical value based on the bone density, muscle density, and age, the statistical value Q is calculated using the following formula (6) based on the bone density distribution index value Bd, muscle mass distribution index value Md, and age Y. Q = W1 × Bd + W2 × Md + W3 × Y (6)

[0073] In equation (6), W1, W2, and W3 are weighting coefficients, and the weighting coefficients W1, W2, and W3 are determined by collecting a large amount of data on bone density distribution index values, muscle density distribution index values, and the subject's ages corresponding to these index values, and performing regression analysis based on the data. Even when calculating statistical values ​​by adding the subject's height, weight, fracture history, and other factors in addition to age, it is preferable to multiply and add the weighting coefficients.

[0074] The evaluation value derivation unit 67 calculates a fracture risk evaluation value for evaluating the fracture risk of the subject H based on the statistical value Q. Since the relationship between the statistical value Q and the fracture risk evaluation value is obtained from a large amount of diagnostic data, the evaluation value derivation unit 67 calculates the fracture risk evaluation value using this relationship. The relationship between the statistical value Q and the fracture risk evaluation value may be derived in advance and stored in the storage 53 as a table.

[0075] For example, the fracture risk evaluation value is the probability of fracture occurring within 10 years from the time of diagnosis of subject H (the time of acquisition of plain radiographic image G0). When equation (6) is used to calculate statistical value Q as described above, the relationship between the "probability of fracture occurring within 10 years" and "statistical value Q" is expressed as shown in FIG. 19, where the larger the statistical value Q, the lower the fracture probability.

[0076] In the second embodiment, the fracture risk assessment value derived by the information derivation device 50A is used as the correct answer data of the training data. Fig. 20 is a diagram showing the training data derived in the second embodiment. As shown in Fig. 20, the training data 40B consists of learning data 41 including a composite two-dimensional image C0 and correct answer data 42B, which is the fracture risk assessment value.

[0077] By training a neural network using the training data 40B shown in Figure 20, a trained neural network 23A can be constructed that, when a simple radiographic image G0 is input, outputs a fracture risk assessment value as an estimated result related to bone density.

[0078] Next, a third embodiment of the present disclosure will be described. FIG. 21 is a diagram showing the functional configuration of an information derivation device according to the third embodiment. Note that in FIG. 21, the same components as those in FIG. 7 are assigned the same reference numerals, and detailed description thereof will be omitted. In the third embodiment of the present disclosure, instead of deriving bone density, information indicating the healing state of a bone portion after treatment is derived as correct answer data 42. For this purpose, as shown in FIG. 21, an information derivation device 50B according to the third embodiment further includes a healing information derivation unit 68 in addition to the information derivation device 50 according to the first embodiment. Note that in the third embodiment, it is assumed that surgery to implant an artificial object such as an artificial bone into the bone portion has been performed as the treatment of the bone portion.

[0079] The healing information derivation unit 68 derives, as healing information, information indicating the state of the bones of the subject after the artificial object has been implanted in the bones of the subject H, based on the bone density around the artificial object, such as an artificial bone, implanted in the bones of the subject H. Artificial objects, such as artificial bones, are implanted in a living body by surgery to replace bone lost due to a comminuted fracture, a tumor, or the like.

[0080] Fig. 22 is a diagram showing an example of an artificial bone embedded in the bone of a subject. Fig. 22 shows an example of the bone of subject H who has undergone total hip replacement surgery, with an artificial joint stem 81 embedded in the femur 80 of subject H.

[0081] Known methods for fixing the stem 81 include direct fixation (cementless fixation) and indirect fixation (cement fixation). In the direct fixation method, the stem 81 is inserted into the cavity inside the femur 80 without using cement. The cavity inside the femur 80 is shaped in advance so that the stem 81 can fit into it. The surface of the stem 81 is roughened, and bone tissue grows by penetrating into the interior of the stem 81. In other words, immediately after the stem 81 is embedded in the femur 80, a cavity exists between the stem 81 and the femur 80, but as the femur 80 recovers, the cavity shrinks and disappears as the bone tissue grows. Therefore, by obtaining the bone density around the stem 81, it is possible to understand the degree of recovery of the femur 80 after surgery.

[0082] FIG. 23 is a graph showing an example of the relationship between the distance from the stem 81 inside the femur 80 and bone density at each postoperative stage. The horizontal axis of the graph shown in FIG. 23 represents the position along the line L in FIG. 22. In FIG. 23, the solid line corresponds to the initial stage immediately after the stem 81 is implanted in the femur 80, the dotted line corresponds to the intermediate recovery stage, and the dashed-dotted line corresponds to the complete recovery stage. As shown in FIG. 23, in the initial postoperative stage, the femur 80 and the stem 81 are not in close contact, and bone density near the stem 81 is extremely low. As recovery progresses, bone tissue grows and penetrates toward the inside of the stem 81, increasing bone density near the stem 81. Meanwhile, bone density at positions distal to the stem 81 remains substantially constant at each postoperative stage. At the complete recovery stage, bone density near the stem 81 and bone density at the distal position are substantially equal.

[0083] Hereinafter, a manner in which the healing information deriving unit 68 derives healing information will be described using an example in which a total hip replacement surgery shown in Fig. 22 is performed. The healing information deriving unit 68 derives healing information from a position L A Bone mineral density in A and the position X, which is relatively far from the stem 81. B Bone mineral density in BFor example, the healing information deriving unit 68 derives a value ΔB corresponding to the difference between the bone densities (ΔB=B B -B A ) may be derived as the healing information. In this case, the value derived as the healing information decreases with recovery and approaches 0. The healing information deriving unit 68 may also derive the ratio of bone densities (ΔB=B A / B B ) may be derived as healing information. In this case, the value ΔB derived as healing information increases as the bone recovers, approaching 1. That is, the bone density B A and B B The value ΔB corresponding to the difference between the values ​​is a value that indicates the degree of recovery of the bone portion after surgery. Therefore, by deriving the value ΔB as healing information, it is possible to quantitatively grasp the degree of recovery of the femur 80 after surgery.

[0084] The healing information derivation unit 68 may derive healing information using the bone density per unit area of ​​each pixel of the composite two-dimensional image C0 derived by the bone density derivation unit 63, or may derive healing information using the bone density per unit volume of each pixel of the CT image V0. Also, in the composite two-dimensional image C0, the pixel values ​​of the stem 81 are significantly different from the pixel values ​​in the bone region, so it is possible to identify the region in the composite two-dimensional image C0 where the stem 81 exists. Therefore, the healing information derivation unit 68 can identify the distance from the stem 81 based on the composite two-dimensional image C0.

[0085] FIG. 24 is a cross-sectional view showing an example of the cross-sectional structure of a human bone. As shown in FIG. 24, a human bone is composed of cancellous bone 90 and cortical bone 91, which covers the cancellous bone 90. The cortical bone 91 is harder and denser than the cancellous bone 90. The cancellous bone 90 is a collection of small bone pillars called trabeculae that extend within the marrow cavity. The trabeculae have a plate-like or rod-like structure and are interconnected. Because the bone density of the cancellous bone 90 is significantly different from that of the cortical bone 91, it is possible to distinguish the cortical bone 91 from the cancellous bone 90 in the CT image V0. In particular, when the derived bone density is bone density per unit volume, the bone density of the cancellous bone 90 can be clearly distinguished from the bone density of the cortical bone 91 compared to bone density per unit area.

[0086] When an artificial object is embedded in the cancellous bone 90, the healing information deriving unit 68 may identify the area of ​​the cancellous bone 90 based on the CT value of each pixel of the CT image V0, and derive healing information based on the bone density of the cancellous bone 90 around the artificial object. Specifically, the healing information deriving unit 68 derives healing information based on the bone density of the cancellous bone 90 around the artificial object. A Bone mineral density in A and position X in the cancellous bone 90, which is relatively far from the artificial object. B Bone mineral density in B A numerical value ΔB corresponding to the difference between the values ​​may be derived as the cure information.

[0087] On the other hand, when an artificial object is embedded in the cortical bone 91, it is preferable that the healing information deriving unit 68 specifies the area of ​​the cortical bone 91 based on the CT value of each pixel of the CT image V0, and derives healing information based on the bone density of the cortical bone 91 around the artificial object. Specifically, the healing information deriving unit 68 derives healing information based on the bone density of the cortical bone 91 around the artificial object. A Bone mineral density in A and position X in the cortical bone 91, which is relatively far from the artificial object. B Bone mineral density in B A numerical value ΔB corresponding to the difference between the values ​​may be derived as the cure information.

[0088] Furthermore, when an artificial object embedded in the bone portion of the subject H extends to both the cancellous bone 90 and the cortical bone 91, the areas of the cancellous bone 90 and the cortical bone 91 may be identified based on the CT value of each pixel of the CT image V0, and healing information may be derived based on the bone densities of both the cancellous bone 90 and the cortical bone 91 around the artificial object. Specifically, the healing information derivation unit 68 determines the position L in the cancellous bone 90 that is relatively close to the artificial object. A1 Bone mineral density in A1 and position L in the cancellous bone 90, which is relatively far from the artificial object. B1 Bone mineral density in B1 A numerical value ΔB1 corresponding to the difference between the artificial object and the cortical bone 91 is derived as healing information, and a position L A2 Bone mineral density in A2 and a position L in the cortical bone 91 that is relatively far from the artificial object. B2 Bone mineral density in B2 A numerical value ΔB2 corresponding to the difference between the values ​​ΔB1 and ΔB2 may be derived as the healing information. Note that, when the artificial object embedded in the bone portion of the subject H extends to both the cancellous bone 90 and the cortical bone 91, the healing information may be derived based on the bone density of one of the cancellous bone 90 and the cortical bone 91 around the artificial object. In other words, one of the numerical values ​​ΔB1 and ΔB2 may be derived as the healing information.

[0089] In the third embodiment, the cure information derived by the information derivation device 50B is used as the correct answer data of the training data. Fig. 25 is a diagram showing the training data derived in the third embodiment. As shown in Fig. 25, the training data 40C consists of learning data 41 including a composite two-dimensional image C0 and correct answer data 42C, which is the numerical value of the cure information.

[0090] By training a neural network using the training data 40C shown in FIG. 25, it is possible to construct a trained neural network 23A that outputs information representing the healing state as healing information when a plain radiographic image G0 is input.

[0091] In each of the above embodiments, a bone density image having pixel values ​​representing bone density per unit area or per unit volume derived by the bone density deriving unit 63 may be used as the correct answer data 42 of the training data 40. In this case, the estimation unit 23 of the estimation device 10 derives the bone density image from the plain radiographic image G0 as an estimation result related to bone density. When the bone density image is derived in this way, the bone density image may be displayed on the display screen.

[0092] FIG. 26 is a diagram showing another example of a display screen for the estimation results. As shown in FIG. 26, the display screen 70A has an image display area 71 similar to the display screen 70 shown in FIG. 14. The image display area 71 displays a bone density image Gd, which is an estimation result of bone density in the plain radiographic image G0 of the subject H. In the bone density image Gd, patterns are applied to the bone regions according to bone density. Note that, for simplicity of explanation, in FIG. 26, patterns representing bone mineral content are applied only to the femur. A reference 73 indicating the magnitude of the bone mineral content for the applied pattern is displayed below the image display area 71. The operator can easily recognize the patient's bone mineral content by interpreting the bone mineral content image Gd while referring to the reference 73. Note that, instead of patterns, different colors may be applied to the bone mineral content image Gd according to bone mineral content.

[0093] Although the above embodiments estimate information related to bone density of the femur near the hip joint, the bone to be estimated is not limited to the femur. The technology disclosed herein can also be applied to estimate information related to bone density of any bone, such as the femur and tibia near the knee joint, vertebrae such as the lumbar vertebrae, the calcaneus, and metacarpals.

[0094] Furthermore, although the above-described embodiments derive bone density-related estimation results from a simple radiographic image G0, this is not limiting. For example, the technology of the present disclosure can also be applied to deriving bone density-related estimation results from DXA scan images acquired by imaging a subject using a DXA imaging device such as those described in Japanese Patent Application Laid-Open Nos. 9-108206 and 2006-271437. A DXA scan image refers to a radiographic image captured by a radiation detector after irradiating a subject with a narrowly collimated high-energy radiation beam and a low-energy radiation beam while switching and scanning the beam. A narrowly collimated radiation beam is, for example, a radiation beam formed into a pencil beam, a narrow fan beam, a wide fan beam, or the like using a collimator located between the radiation source and the subject. Low-energy radiation refers to radiation with energy relatively lower than that of high-energy radiation.

[0095] In this case, an image simulating the DXA scan image may be generated from the composite two-dimensional image C0 according to various conditions such as the pixel size of the detector that captures the DXA scan image, the scanning direction and scanning speed during imaging, the mutual distances between the X-ray source, the subject, and the detector, and the radiation energy distribution (determined by the tube voltage, target, and filter), and the generated image simulating the DXA scan image may be used as training data 41 to construct the trained neural network 23A.

[0096] The image simulating the DXA scan image may be generated, for example, by performing a process on the composite two-dimensional image C0 to reduce the resolution according to the pixel size, scanning direction, scanning speed, etc. of the detector used to capture the DXA scan image. The image simulating the DXA scan image is an example of a low-resolution composite two-dimensional image.

[0097] Specifically, an image simulating a DXA scan image is generated as follows. Assume that M×M pixels in the composite 2D image C0 correspond to the actual size of subject H (L mm×L mm), where L, M, and N are natural numbers, and that N×N pixels in the training image of the DXA scan image correspond to the actual size of subject H. In this case, the resolution of the composite 2D image C0 is reduced by taking the average pixel values ​​of the (M / N)×(M / N) pixels in the composite 2D image C0 as the value of all (M / N)×(M / N) pixels in the composite 2D image C0, so that (M / N)×(M / N) pixels in the composite 2D image C0, i.e., multiple adjacent pixels, correspond to a single pixel in the training image of the DXA scan image. Then, by performing this resolution reduction process on all regions of the composite 2D image C0 that correspond to the DXA scan image, an image simulating a DXA scan image is generated. If M / N is not a natural number, the positions of corresponding pixels in the composite 2D image C0 and the training image of the DXA scan image can be adjusted appropriately using natural numbers before and after M / N to generate an image that simulates the DXA scan image from the composite 2D image C0.

[0098] Furthermore, as a resolution reduction process for simulating blurring due to scanning, an image simulating a DXA scan image may be generated by performing a moving average process in one direction corresponding to the scanning direction.

[0099] Furthermore, a simulated DXA scan image may be generated by performing a moving average process on the composite 2D image C0. The moving average process involves appropriately determining the filter size and intensity distribution used to calculate the moving average based on the scanning direction and scanning speed when capturing the DXA scan image, the pixel size of the detector, and the distances between the X-ray source, the subject, and the detector. For example, the faster the scanning speed, the lower the resolution, so the larger the filter size should be set. In this case, if L=10, then M=200 and N=5.

[0100] In the above embodiments, bone density, fracture risk, and healing information are used as correct answer data included in the training data for training the neural network. Therefore, the bone density-related information estimated by the estimation unit 23 from the plain radiographic image G0 is, but is not limited to, the bone density, fracture risk, and healing information in the plain radiographic image G0. The trained neural network 23A may be constructed using the YAM, T-score, or Z-score as correct answer data, and the YAM, T-score, and Z-score may be estimated from the plain radiographic image G0 as bone density-related information. The estimation unit 23 may also use the detection results of the presence or absence of a fracture, tumor, or implant, or the osteoporosis assessment result, as the bone density-related information to be estimated. Furthermore, bone diseases related to bone density, such as multiple myeloma, rheumatism, arthropathy, and cartilage sclerosis, may be estimated as bone density-related information. In this case, the trained neural network 23A may be constructed using training data including such bone density-related information as correct answer data.

[0101] In each of the above embodiments, the trained neural network 23A is constructed by training the neural network in the estimation device 10, but this is not limiting. A trained neural network 23A constructed in a device other than the estimation device 10 may be used in the estimation unit 23 of the estimation device 10 in this embodiment.

[0102] Furthermore, in each of the above embodiments, the process of estimating information related to bone density is performed using a radiological image acquired in a system that uses a radiation detector 5 to image the subject H. However, the technology of the present disclosure can also be applied to cases where a radiological image is acquired using a stimulable phosphor sheet instead of a radiation detector.

[0103] Furthermore, the radiation in the above embodiment is not particularly limited, and in addition to X-rays, α rays, γ rays, etc. can be used.

[0104] In the above embodiment, the following various processors can be used as the hardware structure of processing units that perform various processes, such as the image acquisition unit 21, information acquisition unit 22, estimation unit 23, learning unit 24, and display control unit 25 of the estimation device 10, and the image acquisition unit 61, synthesis unit 62, and bone density derivation unit 63 of the information derivation device 50, etc. As described 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 dedicated electrical circuits that are processors having a circuit configuration specifically designed to perform specific processes, such as a programmable logic device (PLD), a processor whose circuit configuration can be changed after manufacture, such as an FPGA (Field Programmable Gate Array), and an ASIC (Application Specific Integrated Circuit).

[0105] A single processing unit may be configured with one of these various processors, or may be configured with a combination of two or more processors of the same or different types (for example, a combination of multiple FPGAs or a combination of a CPU and an FPGA). Also, multiple processing units may be configured with a single processor.

[0106] Examples of configuring multiple processing units with a single processor include, first, a form in which one processor is configured with a combination of one or more CPUs and software, and this processor functions as multiple processing units, as typified by computers such as client and server. Second, a form in which a processor is used to realize the functions of an entire system including multiple processing units with a single IC (Integrated Circuit) chip, as typified by systems on chips (SoCs). In this way, various processing units are configured using one or more of the above-mentioned various processors as a hardware structure.

[0107] Furthermore, more specifically, the hardware structure of these various processors can be an electric circuit that combines circuit elements such as semiconductor elements. [Explanation of symbols]

[0108] 1. Imaging device 3 Radiation source 5. Radiation detectors 7 CT device 9. Image Storage System 10 Estimation device 11, 51 CPUs 12 Estimation Processing Program 12B Study Program 13, 53 Storage 14, 54 display 15, 55 Input devices 16, 56 memory 17, 57 Network I / F Buses 18 and 58 21 Image acquisition unit 22 Information Acquisition Department 23 Estimation part 23A Trained Neural Network 24 Learning Department 25 Display control unit 30 Neural Networks 31 Input layer 32 Middle Class 33 Output layer 35 convolutional layers 36 Pooling Layer 37 Fully connected layer 40, 40A, 40B, 40C training data 41 Training data 42, 42A, 42B, 42C Correct data 47 Output Data 48 parameters 50, 50A, 50B Information derivation device 52 Information derivation program 61 Image acquisition unit 62 Synthesis section 63 Bone density derivation section 64 plane 65 Muscle density extraction section 66 Statistical Value Derivation Section 67 Evaluation value derivation part 68 Cure Information Derivation Unit 70, 70A display screen 71 Image display area 72 Bone density display area 73 References 80 Femur 81 Stem 90 Cancellous bone 91 Cortical bone C0 Synthetic 2D image G0 plain radiographic image Gd bone density imaging L distance Q Statistic V0 CT image

Claims

1. at least one processor; The processor: the network functions as a trained neural network that derives an estimation result related to the bone density of a bone portion from a simple radiographic image obtained by simply photographing a subject including the bone portion, or as a trained neural network that derives an estimation result related to the bone density of the bone portion from a DXA scanned image obtained by photographing the subject using a DXA method; the trained neural network is trained using, as training data, a composite two-dimensional image representing the subject derived by combining three-dimensional CT images of the subject and information related to the bone density of the subject; The information related to the bone density is derived based on the bone density at each position in the bone region, which is derived based on the attenuation coefficient of the radiation and the bone mass attenuation coefficient, by identifying a bone region in the CT image and deriving the information related to the bone density.

2. The estimation device according to claim 1, wherein the composite two-dimensional image is derived by deriving a radiation attenuation coefficient for a composition at each position in three-dimensional space and projecting the CT image in a predetermined direction based on the attenuation coefficient.

3. 3. The estimation device according to claim 1, wherein the information relating to the bone density is derived by projecting the bone density at each position in the bone region in a predetermined direction.

4. 4. The estimation device according to claim 1, wherein the information related to bone density includes at least one of bone density per unit area, bone density per unit volume, an evaluation value of the subject's fracture risk, and information representing the healing state of the bone portion after treatment.

5. the processor functions as a trained neural network to derive an inference related to bone mineral density of the bone portion from the DXA scan image; 5. The estimation device according to claim 1, wherein the trained neural network is trained using a low-resolution composite two-dimensional image obtained by performing a process to reduce the resolution of the composite two-dimensional image and information related to the bone density of the subject as training data.

6. 6. The estimation device according to claim 5, wherein the low-resolution composite two-dimensional image is an image in which the pixel value of the adjacent pixels is an average value of the pixel values ​​of adjacent pixels of the composite two-dimensional image, and the size of the adjacent pixels corresponds to the size of one pixel of the DXA scanned image.

7. 6. The estimation device according to claim 5, wherein the low-resolution composite two-dimensional image is an image obtained by performing a moving average process on one direction of the composite two-dimensional image, the one direction corresponding to the scanning direction of the DXA scanned image.

8. 6. The estimation device according to claim 5, wherein the low-resolution composite two-dimensional image is an image generated by generating a first low-resolution image in which the pixel value of a plurality of adjacent pixels in the composite two-dimensional image is an average value of the pixel values ​​of the adjacent plurality of pixels, and performing a moving average process on the first low-resolution image in one direction, wherein the size of the adjacent plurality of pixels corresponds to the size of one pixel in the DXA scanned image, and the one direction corresponds to the scanning direction of the DXA scanned image.

9. An estimation method for deriving an estimation result related to bone density from a plain radiographic image or a DXA scanned image by using a trained neural network that derives an estimation result related to bone density of a bone portion from a plain radiographic image obtained by simply radiographing a subject including the bone portion, or that derives an estimation result related to bone density of the bone portion from a DXA scanned image obtained by radiographing the subject using a DXA method, comprising: the trained neural network is trained using, as training data, a composite two-dimensional image representing the subject derived by combining three-dimensional CT images of the subject and information related to the bone density of the subject; The information related to the bone density is derived based on bone density at each position in the bone region, which is derived based on the attenuation coefficient of the radiation and the bone mass attenuation coefficient, by identifying a bone region in the CT image and deriving the information related to the bone density.

10. An estimation program for causing a computer to execute a procedure for deriving an estimation result related to bone density from a plain radiographic image or a DXA scanned image obtained by simply photographing a subject including a bone, using a trained neural network that derives an estimation result related to bone density of the bone from the plain radiographic image or the DXA scanned image obtained by photographing the subject using a DXA method, the trained neural network is trained using, as training data, a composite two-dimensional image representing the subject derived by combining three-dimensional CT images of the subject and information related to the bone density of the subject; An estimation program that identifies a bone region in the CT image, derives the attenuation coefficient of radiation in the bone region, and derives the information related to the bone density based on the bone density at each position in the bone region, which is derived based on the attenuation coefficient of the radiation and the bone mass attenuation coefficient.

Images