Estimation device, method, and program

Abstract

To highly accurately estimate the composition of a soft tissue of a subject in an estimation device, method and program.SOLUTION: An estimation device includes at least one processor. The processor functions as a learned neural network which derives an estimation result related to the composition of a soft tissue of a subject from a simple radiation image acquired by simply imaging the subject or a DXA scan image acquired by imaging the subject by a DXA method. A learned neural network is learned by using the two radiation images acquired by imaging the subject with the radiations having the different energy distributions, the radiation image of the subject and the soft image expressing the soft tissue of the subject or the composite two-dimensional image expressing the subject derived by composing the three-dimensional CT image of the subject, and information related to the composition of the soft tissue of the subject as teacher data.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] This disclosure relates to estimation apparatus, method, and program. [Background technology]

[0002] In bone diseases such as osteoporosis, DXA (Dual X-ray Absorptiometry) is known as one of the representative bone mineral density quantification methods used for diagnosing bone mineral density. DXA involves measuring the mass attenuation coefficient μ(cm²) of radiation incident on and transmitted through the human body, which depends on the materials that make up the body (e.g., bone). 2 (g) and its density ρ (g / cm³) 3 This method utilizes the fact that bone undergoes attenuation characterized by its energy and thickness t (cm), and calculates bone mineral density from the pixel values ​​of radiation images obtained by imaging with two different energy levels of radiation.

[0003] Furthermore, techniques are known for deriving the proportions of adipose tissue and red tissue in each pixel of a DXA image (see Patent Document 1). The technique described in Patent Document 1 measures the proportion of adipose tissue and red tissue (muscle, non-fat, and non-mineral tissue) by analyzing a combination of low-energy and high-energy DXA images.

[0004] Furthermore, various methods have been proposed for evaluating bone mineral density using radiographic images obtained by photographing a subject. For example, Patent Documents 2 and 3 propose a method for estimating information about bone mineral density from images of bones by using a trained neural network constructed by training a neural network. In the method described in Patent Document 2, images of bones obtained by simple radiography and bone mineral density are used as training data to train the neural network. In the method described in Patent Document 3, images of bones obtained by simple radiography, bone mineral density, and information related to bone mineral density (e.g., age, sex, weight, drinking habits, smoking habits, fracture history, body fat percentage, and subcutaneous fat percentage) are used as training data to train the neural network.

[0005] Simple radiography is a method of imaging in which a subject is irradiated with radiation once, and a single two-dimensional image, which is the transmitted image of the subject, is obtained. In the following explanation, the radiation image obtained by simple radiography will be referred to as a simple radiation image. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2019-63499 [Patent Document 2] U.S. Patent No. 6064716 [Patent Document 3] International Publication No. 2020 / 054738 [Overview of the project] [Problems that the invention aims to solve]

[0007] On the other hand, there is a need to estimate the composition of soft tissues such as muscles and fat in the subject with high accuracy.

[0008] This disclosure is made in view of the above circumstances and aims to enable high-precision estimation of the composition of the soft tissue of the subject. [Means for solving the problem]

[0009] The estimation device described herein comprises at least one processor, The processor functions as a trained neural network that derives estimation results related to the composition of the subject's soft tissue from simple radiographic images obtained by simply photographing the subject, or from DXA scan images obtained by photographing the subject using the DXA method. A trained neural network is trained using the following as training data: (i) two radiation images obtained by photographing the subject with radiation of different energy distributions, (ii) a radiation image of the subject and a soft tissue image representing the subject's soft tissues, or (iii) a composite two-dimensional image representing the subject derived by combining a three-dimensional CT image of the subject, and information related to the composition of the subject's soft tissues.

[0010] In addition, in the estimation device according to this disclosure, information related to the composition of soft tissue may be derived based on the pixel values ​​of the soft tissue region in a soft tissue image derived from two radiation images obtained by photographing the subject with radiation having different energy distributions.

[0011] Furthermore, in the estimation device according to this disclosure, information related to the composition of soft tissue may be derived by identifying the soft tissue region in the CT image, deriving the radiation attenuation coefficient in the soft tissue region, and deriving the density of the soft tissue composition based on the radiation attenuation coefficient and the mass attenuation coefficient in the soft tissue region.

[0012] Furthermore, in the estimation device according to this disclosure, information related to the composition of soft tissue may be derived by projecting the density of the composition at each location in the soft tissue region in a predetermined direction.

[0013] In addition, in the estimation device according to the present disclosure, the information related to the composition of soft tissue may include at least one of the muscle mass per unit area, the muscle mass per unit volume, the fat mass per unit area, the fat mass per unit volume, the disease risk of a predetermined disease or the disease information representing the disease level of a predetermined disease, and the fall rate.

[0014] In addition, in the estimation device according to the present disclosure, the disease information derives the muscle mass of a predetermined site of the subject, and may be derived by representing the correspondence information representing the correspondence between the disease information representing the disease risk of a predetermined disease or the disease level of a predetermined disease and the muscle mass of a predetermined site, and specifying the disease information based on the muscle mass.

[0015] In addition, in the estimation device according to the present disclosure, the disease information derives the muscle mass and bone mineral content of a predetermined site of the subject, and may be derived by representing the correspondence information representing the correspondence between the disease information representing the disease risk of a predetermined disease or the disease level of a predetermined disease and the muscle mass of a predetermined site, the muscle mass, and specifying the disease information based on the muscle mass and the bone mineral content.

[0016] In addition, in the estimation device according to the present disclosure, the predetermined site is the lower limb, the predetermined disease is diabetes, and the disease information may be the disease risk of diabetes.

[0017] In addition, in the estimation device according to the present disclosure, the predetermined site is the extremities or the whole body, the predetermined disease is sarcopenia, and the disease information may be the disease level of sarcopenia.

[0018] In addition, in the estimation device according to the present disclosure, the fall rate derives the muscle mass of a predetermined site of the subject, It may also be derived by identifying the fall rate based on muscle information, which is the muscle mass of a predetermined body part or muscle strength corresponding to the muscle mass of a predetermined body part, and the subject's fall rate, as well as correspondence information that represents the correspondence between the subject's fall rate and muscle information, or muscle mass.

[0019] Furthermore, in the estimation device described herein, the fall rate is derived by deriving the muscle mass and bone mineral density of predetermined parts of the subject. The fall rate may also be derived by identifying the fall rate based on muscle information, which represents the relationship between muscle mass in a predetermined area or muscle strength corresponding to the muscle mass in a predetermined area, and the fall rate of the subject, as well as muscle mass and bone mineral density.

[0020] Furthermore, in the estimation device according to this disclosure, the predetermined area may be at least one of the lower limbs and the buttocks.

[0021] Furthermore, in the estimation device described herein, the processor functions as a trained neural network that derives estimation results related to the composition of soft tissue from DXA scan images. A trained neural network may be one that has been trained using low-resolution synthetic two-dimensional images, obtained by processing synthetic two-dimensional images to reduce their resolution, and information related to the composition of the soft tissues of the subject, as training data.

[0022] Furthermore, in the estimation device according to this disclosure, the low-resolution composite two-dimensional image is an image in which the pixel values ​​of adjacent pixels in the composite two-dimensional image are the average of the pixel values ​​of adjacent pixels, and the size of adjacent pixels may correspond to the size of one pixel in the DXA scan image.

[0023] Furthermore, in the estimation apparatus according to this disclosure, the low-resolution composite two-dimensional image is an image obtained by applying a moving average process to one direction of the composite two-dimensional image, and this one direction may correspond to the scanning direction of the DXA scan image.

[0024] Furthermore, in the estimation device according to this disclosure, the low-resolution composite two-dimensional image is generated by creating a first low-resolution image in which the average value of the pixel values ​​of multiple adjacent pixels in the composite two-dimensional image is used as the pixel value of multiple adjacent pixels, and then applying a moving average process in one direction to the first low-resolution image to generate the image, wherein the size of multiple adjacent pixels corresponds to the size of one pixel in the DXA scan image, and the one direction corresponds to the scanning direction of the DXA scan image.

[0025] The estimation method according to this disclosure is an estimation method that derives estimation results related to the composition of soft tissue from simple radiographic images obtained by simply photographing the subject, or from DXA scan images obtained by photographing the subject using the DXA method, using a trained neural network that derives estimation results related to the composition of soft tissue from simple radiographic images or DXA scan images, A trained neural network is trained using the following as training data: (i) two radiation images obtained by photographing the subject with radiation of different energy distributions, (ii) a radiation image of the subject and a soft tissue image representing the subject's soft tissues, or (iii) a composite two-dimensional image representing the subject derived by combining a three-dimensional CT image of the subject, and information related to the composition of the subject's soft tissues.

[0026] Furthermore, the estimation method described herein may be provided as a program for a computer to execute. [Effects of the Invention]

[0027] According to this disclosure, the composition of the soft tissue of the subject can be estimated with high accuracy. [Brief explanation of the drawing]

[0028] [Figure 1] A schematic block diagram showing the configuration of a radiographic imaging system to which the estimation device according to the first embodiment of this disclosure is applied. [Figure 2] This figure shows the schematic configuration of the estimation device according to the first embodiment. [Figure 3]This figure shows the functional configuration of the estimation device according to the first embodiment. [Figure 4] This figure shows the schematic configuration of the neural network used in this embodiment. [Figure 5] Diagram showing training data [Figure 6] This figure shows the schematic configuration of the information extraction device according to the first embodiment. [Figure 7] This figure shows the functional configuration of the information extraction device according to the first embodiment. [Figure 8] Figure showing soft tissue images [Figure 9] A diagram showing an example of the energy spectra of radiation after it has passed through muscle tissue and radiation after it has passed through adipose tissue. [Figure 10] A diagram to explain neural network learning. [Figure 11] Conceptual diagram of the processing performed by a trained neural network. [Figure 12] A diagram showing the display screen for the estimation results. [Figure 13] Flowchart of the learning process performed in the first embodiment [Figure 14] Flowchart of the estimation process performed in the first embodiment [Figure 15] This figure shows the functional configuration of the information extraction device according to the second embodiment. [Figure 16] Figure showing the training data derived in the second embodiment. [Figure 17] This figure shows the functional configuration of the information extraction device according to the third embodiment. [Figure 18] Figure showing bone images [Figure 19] A diagram showing the relationship between the contrast between bony and soft tissues in relation to the body thickness of the subject. [Figure 20] A diagram showing an example of a lookup table for obtaining correction coefficients. [Figure 21] A diagram showing correspondence information representing the correlation between muscle mass and fall rate in the third embodiment. [Figure 22] Figure showing the training data derived in the third embodiment. [Figure 23] This figure shows the functional configuration of the information extraction device according to the fourth embodiment. [Figure 24] A diagram illustrating the derivation of a composite 2D image. [Figure 25] A diagram illustrating the derivation of a composite 2D image. [Figure 26] Diagram to explain CT values [Figure 27] A diagram showing the relationship between radiation energy and the mass attenuation coefficient. [Figure 28] Figure showing the training data derived in the fourth embodiment. [Figure 29] This figure shows the functional configuration of the information extraction device according to the fifth embodiment. [Figure 30] Figure showing the training data derived in the fifth embodiment. [Figure 31] This figure shows the functional configuration of the information extraction device according to the sixth embodiment. [Figure 32] Figure showing the training data derived in the sixth embodiment. [Figure 33] This figure shows the functional configuration of the information extraction device according to the seventh embodiment. [Figure 34] Figure showing the training data derived in the seventh embodiment. [Figure 35] This figure shows the functional configuration of the information extraction device according to the eighth embodiment. [Figure 36] Figure showing the training data derived in the eighth embodiment. [Figure 37] This figure shows training data using soft tissue images as learning data. [Modes for carrying out the invention]

[0029] Embodiments of this disclosure will be described below with reference to the drawings. Figure 1 is a schematic block diagram showing the configuration of a radiographic imaging system to which the estimation device according to the first embodiment of this disclosure is applied. As shown in Figure 1, the radiographic imaging system according to the first embodiment comprises an imaging device 1, a CT (Computed Tomography) device 4, an image storage system 9, an estimation device 10 according to the first embodiment, and an information extraction device 50. The imaging device 1, the CT device 4, the estimation device 10, and the information extraction device 50 are connected to the image storage system 9 via a network (not shown).

[0030] The imaging device 1 is capable of performing energy subtraction using a so-called one-shot method, in which radiation such as X-rays emitted from the radiation source 3 and transmitted through the subject H is irradiated onto the first radiation detector 5 and the second radiation detector 6 with varying energies. During imaging, as shown in Figure 1, the first radiation detector 5, a radiation energy conversion filter 7 made of a copper plate or the like, and the second radiation detector 6 are arranged in order from the side closest to the radiation source 3, and the radiation source 3 is driven. The first and second radiation detectors 5 and 6 and the radiation energy conversion filter 7 are in close contact.

[0031] As a result, the first radiation detector 5 acquires a first radiation image G1 of subject H using low-energy radiation, including so-called soft rays. The second radiation detector 6 acquires a second radiation image G2 of subject H using high-energy radiation, from which soft rays have been removed. Therefore, the first radiation image G1 and the second radiation image G2 are obtained by photographing subject H with radiation that has a different energy distribution. The first and second radiation images G1 and G2 are transmitted to the image storage system 9. Both the first radiation image G1 and the second radiation image G2 are frontal images of subject H, including the buttocks and upper lower limbs.

[0032] The first and second radiation detectors 5 and 6 are capable of repeatedly recording and reading radiation images. They may be so-called direct-type radiation detectors that generate an electric charge by directly receiving radiation, or they may be so-called indirect-type radiation detectors that first convert radiation into visible light and then convert that visible light into an electric charge signal. Furthermore, as a method for reading out the radiation image signal, it is desirable to use a so-called TFT (thin film transistor) readout method, in which the radiation image signal is read out by turning a TFT switch on and off, or a so-called optical readout method, in which the radiation image signal is read out by irradiating it with reading light. However, other methods may also be used.

[0033] Furthermore, the imaging device 1 can also acquire a simple two-dimensional image of the subject H, which is a simple radiation image G0, by performing simple imaging of the subject H using only the first radiation detector 5. To distinguish the imaging that acquires the first and second radiation images G1 and G2 described above from simple imaging, it is called energy subtraction imaging. In this embodiment, the first and second radiation images G1 and G2 acquired by energy subtraction imaging are used to derive training data for training the neural network described later. In addition, the simple radiation image G0 acquired by simple imaging is used to derive estimation results related to the composition of soft tissue, as described later.

[0034] The CT device 4 acquires multiple tomographic images representing multiple 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 composition that makes up the human body. The CT value will be explained later. In this embodiment, the CT device 4 is also used to derive the training data, which will be explained later.

[0035] The image storage system 9 is a system that stores image data of radiographic images acquired by the imaging device 1 and image data of CT images acquired by the CT device 4. The image storage system 9 retrieves images from the stored radiographic and CT images in accordance with 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 System). In this embodiment, the image storage system 9 stores a large amount of training data for training the neural network described later.

[0036] Next, an estimation device according to the first embodiment will be described. First, with reference to Figure 2, the hardware configuration of the estimation device according to the first embodiment will be described. As shown in Figure 2, the estimation device 10 is a computer such as a workstation, server computer, or 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, 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, storage 13, display 14, input devices 15, memory 16, and network I / F 17 are connected to a bus 18. Note that the CPU 11 is an example of a processor in this disclosure.

[0037] The storage 13 is implemented using an HDD (Hard Disk Drive), an SSD (Solid State Drive), and flash memory, etc. The storage 13, as a storage medium, stores the estimation program 12A and the learning program 12B installed on the estimation device 10. The CPU 11 reads the estimation program 12A and the learning program 12B from the storage 13, expands them into memory 16, and executes the expanded estimation program 12A and the learning program 12B.

[0038] The estimation program 12A and the learning program 12B are stored in a memory device of a server computer connected to the network, or in network storage, in a state that allows external access, and are downloaded and installed on the computers constituting the estimation device 10 upon request. Alternatively, they are recorded on a recording medium such as a DVD (Digital Versatile Disc) or CD-ROM (Compact Disc Read Only Memory) and distributed, and then installed from that recording medium onto the computers constituting the estimation device 10.

[0039] Next, the functional configuration of the estimation device according to the first embodiment will be described. Figure 3 is a diagram showing the functional configuration of the estimation device according to the first embodiment. As shown in Figure 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 functions as the image acquisition unit 21, the information acquisition unit 22, the estimation unit 23, and the display control unit 25 by executing the estimation program 12A. The CPU 11 also functions as the learning unit 24 by executing the learning program 12B.

[0040] The image acquisition unit 21 causes the imaging device 1 to perform energy subtraction imaging of the subject H, thereby acquiring a first radiation image G1 and a second radiation image G2, which are, for example, frontal images of the area around the groin of the subject H, from the first and second radiation detectors 5 and 6. When acquiring the first radiation image G1 and the second radiation image G2, imaging conditions such as the imaging dose, beam quality, tube voltage, SID (Source Image receptor Distance), which is the distance between the radiation source 3 and the surfaces of the first and second radiation detectors 5 and 6, 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 a scatter removal grid are set.

[0041] SOD and SID are used to calculate body thickness distribution, as described later. For SOD, it is preferable to acquire it using a TOF (Time Of Flight) camera, for example. For SID, it is preferable to acquire it using a potentiometer, ultrasonic rangefinder, or laser rangefinder, for example.

[0042] The imaging conditions can be set by the operator via input device 15. The set imaging conditions, along with the first and second radiographic images G1 and G2 acquired by energy subtraction imaging, are transmitted to the image storage system 9 and stored.

[0043] Furthermore, the image acquisition unit 21 uses only the first radiation detector 5 to cause the imaging device 1 to perform a simple radiographic image of the subject H, thereby acquiring a simple radiographic image G0, which is a frontal view of the area around the groin of the subject H. The simple radiographic image G0 is stored in the storage 13.

[0044] In this embodiment, the first and second radiation images G1 and G2, as well as the simple radiation image G0, may be acquired by a program separate from the estimation program 12A.

[0045] The information acquisition unit 22 acquires training data for the neural network described later from the image storage system 9 via the network interface 17.

[0046] The estimation unit 23 derives estimation results related to the soft tissue composition of the subject H from the simple radiographic image G0. In this embodiment, the estimation results related to the soft tissue composition are derived from predetermined muscle mass estimates for soft tissue regions included in the simple radiographic image G0. To this end, the estimation unit 23 derives estimation results related to the soft tissue composition using a trained neural network 23A that outputs muscle mass when the simple radiographic image G0 is input.

[0047] The learning unit 24 constructs a trained neural network 23A by machine learning the neural network using training data. Examples of neural networks include simple perceptrons, multilayer perceptrons, deep neural networks, convolutional neural networks, deep belief networks, recurrent neural networks, and stochastic neural networks. In this embodiment, a convolutional neural network is used as the neural network.

[0048] Figure 4 shows a neural network used in this embodiment. As shown in Figure 4, the neural network 30 comprises an input layer 31, a hidden layer 32, and an output layer 33. The hidden layer 32 comprises, for example, multiple convolutional layers 35, multiple 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, convolutional layers 35 and pooling layers 36 are alternately arranged between the input layer 31 and the fully connected layer 37.

[0049] Note that the configuration of the neural network 30 is not limited to the example in Figure 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.

[0050] Figure 5 shows an example of training data used to train a neural network. As shown in Figure 5, the training data 40 consists of training data 41 and ground truth data 42. In this embodiment, the data input to the trained neural network 23A to obtain estimation results related to the composition of soft tissue is a simple radiographic image G0, while the training data 41 includes two radiographic images, a first radiographic image G1 and a second radiographic image G2, obtained by the energy subtraction imaging described above.

[0051] The correct data 42 is the muscle mass of predetermined parts of the subject from which the training data 41 was acquired. In this embodiment, the predetermined parts are the buttocks and the upper lower limbs. In this embodiment, since the muscle mass per unit area is estimated from a two-dimensional simple radiographic image G0, the unit of muscle mass is (g / cm²). 2 ) In this embodiment, the muscle mass per unit volume may be estimated. In this case, the unit of muscle mass is (g / cm³). 3 The correct answer data 42, muscle mass, is derived by the information derivation device 50. Note that the muscle mass that becomes the correct answer data 42 is an example of information related to the composition of the soft tissue of the subject. The information derivation device 50 will be described below.

[0052] Figure 6 is a schematic block diagram showing the configuration of the information extraction device according to the first embodiment. As shown in Figure 6, the information extraction device 50 according to the first embodiment is a computer such as a workstation, server computer, or personal computer, and includes a CPU 51, non-volatile storage 53, and memory 56 as a temporary storage area. The information extraction device 50 also includes a display 54 such as a liquid crystal display, input devices 55 such as a keyboard and pointing devices such as a mouse, and a network I / F 57 connected to a network (not shown). The CPU 51, storage 53, display 54, input devices 55, memory 56, and network I / F 57 are connected to a bus 58.

[0053] Storage 53, like storage 13, is implemented using an HDD, SSD, flash memory, etc. The information extraction program 52 is stored in storage 53 as a storage medium. The CPU 51 reads the information extraction program 52 from storage 53, expands it into memory 56, and executes the expanded information extraction program 52.

[0054] Next, the functional configuration of the information extraction device according to the first embodiment will be described. Figure 7 is a diagram showing the functional configuration of the information extraction device according to the first embodiment. As shown in Figure 7, the information extraction device 50 according to the first embodiment includes an image acquisition unit 61, a scattered radiation removal unit 62, a subtraction unit 63, and a muscle mass extraction unit 64. When the CPU 51 executes the information extraction program 52, the CPU 51 functions as the image acquisition unit 61, the scattered radiation removal unit 62, the subtraction unit 63, and the muscle mass extraction unit 64.

[0055] The image acquisition unit 61 acquires the first radiation image G1 and the second radiation image G2, which will be used as training data 41 stored in the image storage system 9. The image acquisition unit 61 may acquire the first radiation image G1 and the second radiation image G2 by having the imaging device 1 take an image of the subject H, similar to the image acquisition unit 21 of the estimation device 10.

[0056] The image acquisition unit 61 also acquires the shooting conditions used when acquiring the first radiation image G1 and the second radiation image G2, which are stored in the image storage system 9. The shooting conditions include the shooting dose, tube voltage, SID, SOD, and the presence or absence of a scatter removal grid when acquiring the first radiation image G1 and the second radiation image G2.

[0057] Here, each of the first radiation image G1 and the second radiation image G2 includes a scattered radiation component based on radiation scattered within the subject H, in addition to the primary radiation component of the radiation transmitted through the subject H. Therefore, the scattered radiation removal unit 62 removes the scattered radiation component from the first radiation image G1 and the second radiation image G2. For example, the scattered radiation removal unit 62 may remove the scattered radiation component from the first radiation image G1 and the second radiation image G2 by applying the method described in Japanese Patent Application Publication No. 2015-043959. When using the method described in Japanese Patent Application Publication No. 2015-043959, the derivation of the thickness distribution of the subject H and the derivation of the scattered radiation component for removal are performed simultaneously.

[0058] The following describes the removal of scattered radiation components from the first radiation image G1, but the removal of scattered radiation components from the second radiation image G2 can be performed in the same manner. First, the scattered radiation removal unit 62 acquires a virtual model of the subject H having an initial thickness distribution T0(x,y). The virtual model is data that virtually represents the subject H, in which the thickness according to the initial thickness distribution T0(x,y) is associated with the coordinate position of each pixel in the first radiation image G1. Note that the virtual model of the subject H having an initial thickness distribution T0(x,y) may be stored in advance in the storage 53 of the information derivation device 50. Alternatively, the scattered radiation removal unit 62 may calculate the thickness distribution T(x,y) of the subject H based on the SID and SOD included in the imaging conditions. In this case, the initial thickness distribution T0(x,y) can be obtained by subtracting SOD from SID.

[0059] Next, the scattered radiation removal unit 62 generates an estimated image by combining an estimated primary radiation image, which is obtained by estimating the primary radiation image obtained by photographing the virtual model, and an estimated scattered radiation image, which is obtained by estimating the scattered radiation image obtained by photographing the virtual model, as an estimated image obtained by estimating the first radiation image G1 obtained by photographing the subject H.

[0060] Next, the scattered radiation removal unit 62 modifies the initial thickness distribution T0(x,y) of the virtual model so that the difference between the estimated image and the first radiation image G1 is small. The scattered radiation removal unit 62 repeatedly generates the estimated image and modifies the thickness distribution until the difference between the estimated image and the first radiation image G1 satisfies a predetermined termination condition. The scattered radiation removal unit 62 derives the thickness distribution when the termination condition is met as the thickness distribution T(x,y) of the subject H. Furthermore, the scattered radiation removal unit 62 removes the scattered radiation component contained in the first radiation image G1 by subtracting the scattered radiation component when the termination condition is met from the first radiation image G1.

[0061] The subtraction unit 63 derives a soft tissue image Gs from the first and second radiation images G1 and G2 by performing energy subtraction processing, thereby extracting the soft tissue of the subject H. In subsequent processing, the first and second radiation images G1 and G2 have the scattered radiation component removed. When deriving the soft tissue image Gs, the subtraction unit 63 performs weighted subtraction between corresponding pixels on the first and second radiation images G1 and G2, as shown in equation (1) below, thereby generating a soft tissue image Gs from which the soft tissue of the subject H contained in each radiation image G1 and G2 is extracted, as shown in Figure 8. In equation (1), α is the weighting coefficient. Gs(x, y)=G1(x, y)-α×G2(x, y) (1)

[0062] The muscle mass deriving unit 64 derives the muscle mass for each pixel in the soft tissue region of the soft tissue image Gs based on the pixel value. Soft tissue includes muscle tissue, adipose tissue, blood, and water. In the muscle mass deriving unit 64 of the first embodiment, tissue other than adipose tissue in the soft tissue is considered to be muscle tissue. That is, in the muscle mass deriving unit 64 of the first embodiment, non-adipose tissue, including blood and water, is treated as muscle tissue.

[0063] The muscle mass extraction unit 64 separates muscle and fat from the soft tissue image Gs by utilizing the difference in energy characteristics between muscle tissue and adipose tissue. As shown in Figure 9, the dose of radiation after passing through the subject H is lower than the dose of radiation before it enters the subject H, which is the human body. Also, since muscle tissue and adipose tissue absorb different energies and have different attenuation coefficients, the energy spectra of the radiation after passing through muscle tissue and the radiation after passing through adipose tissue are different. As shown in Figure 9, the energy spectra of the radiation that passes through the subject H and irradiates the first radiation detector 5 and the second radiation detector 6, respectively, depend on the body composition of the subject H, specifically the ratio of muscle tissue to adipose tissue. Since adipose tissue is more permeable to radiation than muscle tissue, a higher proportion of muscle tissue compared to adipose tissue results in a lower dose of radiation after passing through the human body.

[0064] Therefore, the muscle mass derivation unit 64 separates muscle and fat from the soft tissue image Gs by utilizing the difference in energy characteristics between muscle tissue and adipose tissue as described above. In other words, the muscle mass derivation unit 64 generates a muscle image from the soft tissue image Gs. The muscle mass derivation unit 64 also derives the muscle mass of each pixel based on the pixel values ​​of the muscle image.

[0065] The specific method by which the muscle mass derivation unit 64 separates muscle and fat from the soft tissue image Gs is not limited, but as an example, the muscle mass derivation unit 64 of the first embodiment generates a muscle image from the soft tissue image Gs using the following equations (2) and (3). Specifically, first, the muscle mass derivation unit 64 derives the muscle ratio rm(x,y) at each pixel position (x,y) in the soft tissue image Gs using equation (2). In equation (2), μm 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. Also, Δ(x,y) represents the density difference distribution. The density difference distribution is the distribution of density changes on the image as seen from the density obtained when radiation reaches the first radiation detector 5 and the second radiation detector 6 without passing through the subject H. The distribution of density changes in the image is calculated by subtracting the density of each pixel in the subject H region from the density in the pass-through region obtained by directly irradiating the first radiation detector 5 and the second radiation detector 6 with radiation in the soft tissue image Gs. rm(x,y)={μf-Δ(x,y) / T(x,y)} / (μf-μm) (2)

[0066] Furthermore, the muscle mass derivation unit 64 generates a muscle image Gm from the soft tissue image Gs using the following equation (3). In equation (3), x and y are the coordinates of each pixel in the muscle image Gm. Gm(x,y) = rm(x,y) × Gs(x,y) (3)

[0067] The muscle mass derivation unit 64 can derive the pixel values ​​of the muscle image Gm as the muscle mass for each pixel. Alternatively, the muscle ratio rm for each pixel may be derived as the muscle mass. Furthermore, as shown in equation (4) below, the muscle mass derivation unit 64 multiplies each pixel (x,y) of the muscle image Gm by a predetermined coefficient K1(x,y) that represents the relationship between the pixel value and the muscle mass, thereby obtaining the muscle mass M(x,y)(g / cm) for each pixel of the muscle image Gm. 2 You may derive ) from this. M(x,y) = K1(x,y) × Gm(x,y) (4)

[0068] Furthermore, in the first embodiment, the muscle mass derivation unit 64 derives the muscle mass of a predetermined area. In this embodiment, the muscle mass of a predetermined area refers to the representative value of the muscle mass for each pixel in the region corresponding to the predetermined area in the soft tissue image Gs. In this embodiment, since the predetermined area is the buttocks and the upper part of the lower limbs, the representative value of the muscle mass for each pixel in the region corresponding to the buttocks and the upper part of the lower limbs in the soft tissue image Gs is treated as the muscle mass of the buttocks and the upper part of the lower limbs. As the representative value, the mean, median, minimum, and maximum can be used. In this embodiment, the representative value of the muscle mass of the predetermined area is used as the correct answer data 42.

[0069] The muscle mass used as the ground truth data 42 is derived at the same time as the training data 41 is acquired and transmitted to the image storage system 9. In the image storage system 9, the training data 41 and the ground truth data 42 are associated and stored as training data 40. In order to improve the robustness of learning, additional training data 40 may be created and stored, which includes images of the same image that have undergone at least one of the following actions: scaling, contrast adjustment, translation, rotation within the plane, inversion, and noise addition, as training data 41.

[0070] Returning to the estimation device 10, the learning unit 24 trains the neural network using a large amount of training data 40. Figure 10 is a diagram illustrating the training of the neural network 30. When training the neural network 30, the learning unit 24 inputs the training data 41, i.e., the first and second radiographic images G1 and G2, into the input layer 31 of the neural network 30. The learning unit 24 then causes the output layer 33 of the neural network 30 to output muscle mass of predetermined areas as output data 47. The learning unit 24 then derives the difference between the output data 47 and the ground truth data 42 as the loss L0.

[0071] The learning unit 24 learns the neural network 30 based on the loss L0. Specifically, the learning unit 24 adjusts the kernel coefficients in the convolutional layer 35, the weights of the connections between each layer, and the weights of the connections in the fully connected layer 37 (hereinafter referred to as parameters 48) to reduce the loss L0. For example, backpropagation can be used to adjust the parameters 48. The learning unit 24 repeats the adjustment of parameters 48 until the loss L0 falls below a predetermined threshold. As a result, the parameters 48 are adjusted so that when a simple radiographic image G0 is input, the muscle mass of a predetermined area is output, and the trained neural network 23A is constructed. The constructed trained neural network 23A is stored in the storage 13.

[0072] Figure 11 is a conceptual diagram of the processing performed by the trained neural network 23A. As shown in Figure 11, when a simple radiographic image G0 of a patient is input to the trained neural network 23A constructed as described above, the trained neural network 23A will output the muscle mass of predetermined areas contained in the input simple radiographic image G0.

[0073] The display control unit 25 displays the muscle mass estimation results estimated by the estimation unit 23 on the display 14. Figure 12 shows the display screen of the estimation results. As shown in Figure 12, the display screen 70 has an image display area 71 and a muscle mass display area 72. The image display area 71 displays a simple radiographic image G0 of the subject H. The muscle mass display area 72 displays a representative value of the muscle mass of a predetermined area, derived from the muscle mass of each pixel of the simple radiographic image G0 estimated by the estimation unit 23.

[0074] Next, the process performed in the first embodiment will be described. Figure 13 is a flowchart of the learning process 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 the training data 41 contained in the training data 40 into the neural network 30 to output muscle mass, learns the neural network 30 using the loss L0 based on the difference with the correct answer data 42 (step ST2), and returns to step ST1. The learning unit 24 then repeats the processes of steps ST1 and ST2 until the loss L0 reaches a predetermined threshold, and terminates the learning process. The learning unit 24 may also terminate the learning process by repeating the learning a predetermined number of times. As a result, the learning unit 24 constructs a trained neural network 23A.

[0075] Next, the estimation process in the first embodiment will be described. Figure 14 is a flowchart showing the estimation process performed in the first embodiment. The simple radiographic image G0 is assumed to be acquired by imaging and stored in the storage 13. When an instruction to start processing 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 estimation results related to the composition of soft tissue from the simple radiographic image G0 (step ST12). Then, the display control unit 25 displays the estimation results related to the composition of soft tissue derived by the estimation unit 23 together with the simple radiographic image G0 on the display 14 (step ST13), and the process ends.

[0076] Thus, in this embodiment, a trained neural network 23A, constructed by training with the first and second radiographic images G1 and G2 as training data, is used to derive estimation results related to the soft tissue composition of subject H contained in the simple radiographic image G0. Here, in this embodiment, two radiographic images, the first and second radiographic images G1 and G2, are used to train the neural network. Therefore, compared to the case where one radiographic image and information related to the soft tissue composition are used as training data, the trained neural network 23A is able to derive estimation results related to the soft tissue composition from the simple radiographic image G0 with greater accuracy. Accordingly, according to this embodiment, estimation results related to the soft tissue composition can be derived with greater accuracy.

[0077] In the first embodiment described above, muscle mass is derived as the ground truth data 42, but this is not the only way to do so. Disease information representing the risk of developing a predetermined disease or the disease level of a predetermined disease may also be derived as the ground truth data 42. This will be described below as the second embodiment.

[0078] Figure 15 shows the functional configuration of the information extraction device according to the second embodiment. In Figure 15, the same reference numerals are used for components identical to those in Figure 7, and detailed explanations are omitted. As shown in Figure 15, the information extraction device 50A according to the second embodiment is further equipped with a specific unit 65 compared to the information extraction device 50 according to the first embodiment.

[0079] The identification unit 65 identifies disease information that represents the risk of contracting a predetermined disease or the disease level of a predetermined disease, based on the muscle mass of a predetermined area derived by the muscle mass derivation unit 64 and the corresponding relationship information.

[0080] Correspondence information refers to information that represents the correspondence between disease information, which indicates the risk of developing a predetermined disease or the disease level of a predetermined disease, and muscle mass in a predetermined body part.

[0081] In general, there are diseases for which a relationship between muscle mass and the risk of developing the disease is known. For example, muscles are known to take up some of the glucose in the blood and regulate blood glucose levels. Therefore, when muscle mass decreases, blood glucose levels rise, and the risk of developing diabetes increases. In particular, a decrease in muscle mass in the lower limbs tends to correspond to an increased risk of developing diabetes. For this reason, for diabetes, information representing the relationship between muscle mass in the upper part of the lower limbs, specifically the thigh, and the risk of developing diabetes is used as correspondence information. A specific example is correspondence information in which the average value of thigh muscle mass for each age is used as a reference value, and the lower the muscle mass is from the reference value and the greater the deviation from the reference value, the higher the risk of developing the disease.

[0082] For example, in the case of sarcopenia, it is known that as the disease progresses, that is, as the disease level increases, muscle mass, especially in the limbs, tends to decrease. Therefore, for sarcopenia, information representing the correspondence between muscle mass in the limbs (a predetermined area) and the disease level (degree of progression) of sarcopenia is used as correspondence information. Specifically, for each age group, the average value of muscle mass in the limbs is used as the reference value, and correspondence information is associated with a higher disease level as the value is lower than the reference value and the deviation from the reference value is greater.

[0083] The information extraction device 50A may store correspondence information for multiple types of diseases in the storage 53, or it may store correspondence information for a specific disease in the storage 53.

[0084] Thus, the identification unit 65 of the second embodiment uses correspondence information corresponding to the parts of the subject (imaging areas) shown in the first radiographic image G1 and the second radiographic image G2 to identify disease information representing the risk of contracting a predetermined disease or the disease level of a predetermined disease.

[0085] In the second embodiment, the disease information derived by the information derivation device 50A is used as the correct answer data for the training data. Figure 16 shows the training data derived in the second embodiment. As shown in Figure 16, the training data 40A consists of training data 41A including the first and second radiographic images G1 and G2, and correct answer data 42A which is disease information.

[0086] By training a neural network using the training data 40A shown in Figure 16, a trained neural network 23A can be constructed that, when a simple radiographic image G0 is input, outputs disease information as an estimated result related to the composition of soft tissue.

[0087] In the first embodiment described above, muscle mass is derived as the correct answer data 42, but this is not the only way to do so. The fall rate may also be derived as the correct answer data 42. This will be described below as the third embodiment.

[0088] Figure 17 shows the functional configuration of the information extraction device according to the third embodiment. In Figure 17, the same reference numerals are used for components identical to those in Figure 7, and detailed explanations are omitted. As shown in Figure 17, the information extraction device 50B according to the third embodiment further includes a bone mineral content extraction unit 66 and a specific unit 67 compared to the information extraction device 50 according to the first embodiment.

[0089] In the third embodiment, the subtraction unit 63 derives a bone image Gb from the first and second radiation images G1 and G2 by performing energy subtraction processing, thereby extracting the bone portion of the subject H. In subsequent processing, the first and second radiation images G1 and G2 have the scattered radiation component removed. To derive the bone image Gb, the subtraction unit 63 performs weighted subtraction between corresponding pixels on the first and second radiation images G1 and G2, as shown in equation (5) below, thereby generating a bone image Gb from which the bone portion of the subject H contained in each radiation image G1 and G2 is extracted, as shown in Figure 18. In equation (5), β is the weighting coefficient. Gb(x,y)=β·G2(x,y)-G1(x,y) (5)

[0090] The bone mineral density derive unit 66 derives the bone mineral density for each pixel of the bone image Gb. In the third embodiment, the bone mineral density derive unit 66 derives the bone mineral density B by converting each pixel value of the bone image Gb to the pixel value of the bone image acquired under standard imaging conditions. Specifically, the bone mineral density derive unit 66 derives the bone mineral density by correcting each pixel value of the bone image Gb using a correction coefficient obtained from a lookup table described later.

[0091] Here, the higher the tube voltage at radiation source 3 and the higher the energy of the radiation emitted from radiation source 3, the smaller the contrast between soft tissue and bone in the radiographic image. Also, during the process of radiation passing through subject H, the low-energy component of the radiation is absorbed by subject H, resulting in beam hardening, which increases the energy of the radiation. The increase in radiation energy due to beam hardening increases with the thickness of subject H.

[0092] Figure 19 shows the relationship between the contrast between bone and soft tissue and the thickness of subject H. Figure 19 shows the relationship between the contrast between bone and soft tissue and the thickness of subject H at three tube voltages: 80kV, 90kV, and 100kV. As shown in Figure 19, the higher the tube voltage, the lower the contrast. Furthermore, beyond a certain thickness of subject H, the greater the thickness, the lower the contrast. Note that the larger the pixel value in the bone region of the bone image Gb, the greater the contrast between bone and soft tissue. Therefore, the relationship shown in Figure 19 shifts towards higher contrast as the pixel value in the bone region of the bone image Gb increases.

[0093] In the third embodiment, a lookup table for obtaining correction coefficients to compensate for the difference in contrast in the bone image Gb according to the tube voltage during acquisition, and for the decrease in contrast due to the effect of beam hardening, is stored in the storage 53 of the information derivation device 50. The correction coefficient is a coefficient for correcting each pixel value of the bone image Gb.

[0094] Figure 20 shows an example of a lookup table for obtaining a correction coefficient. In Figure 20, a lookup table (hereinafter simply referred to as "table") LUT1 is shown as an example, with the reference imaging conditions set to a tube voltage of 90kV. As shown in Figure 20, in table LUT1, a larger correction coefficient is set as the tube voltage increases and the body thickness of the subject H increases. In the example shown in Figure 20, since the reference imaging conditions are a tube voltage of 90kV, the correction coefficient is 1 when the tube voltage is 90kV and the body thickness is 0. Note that although table LUT1 is shown in two dimensions in Figure 20, the correction coefficient differs depending on the pixel value of the bone region. Therefore, table LUT1 is actually a three-dimensional table with an axis representing the pixel value of the bone region added.

[0095] The bone mineral density extraction unit 66 extracts a correction coefficient K0(x,y) for each pixel according to the shooting conditions, including the body thickness distribution T(x,y) of the subject H and the tube voltage setting value stored in the storage 13, from the table LUT1. The body thickness distribution T(x,y) can be the body thickness distribution when the termination condition is met in the scattered radiation removal process performed by the scattered radiation removal unit 62. Then, as shown in equation (6) below, the bone mineral density extraction unit 66 multiplies each pixel (x,y) in the bone region of the bone image Gb by the correction coefficient K0(x,y) to obtain the bone mineral density B(x,y) (g / cm³) for each pixel of the bone image Gb. 2 The bone mineral density B(x,y) derived in this way represents the pixel value of the bone region included in the radiographic image obtained by imaging the subject H with a tube voltage of 90kV, which is the standard imaging condition, and from which the effects of beam hardening have been removed. For this reason, the bone mineral density image, in which the derived bone mineral density is the pixel value of each pixel, is derived by the bone mineral density derivation unit 66. B(x,y) = K0(x,y) × Gb(x,y) (6)

[0096] In the third embodiment, the identification unit 67 receives information representing muscle mass from the muscle mass extraction unit 64 and information representing bone mineral content from the bone mineral content extraction unit 66. Based on the muscle mass of predetermined areas extracted by the muscle mass extraction unit 64, the bone mineral content extracted by the bone mineral content extraction unit 66, and the corresponding relationship information, the identification unit 67 determines the fall rate (fall incidence rate) of the subject.

[0097] It is generally known that there is a correlation between the probability of a person falling and their muscles or muscle strength. In particular, it is known that there is a correlation with the amount of muscle in the calf (soleus muscle) and the gluteal muscles that support the pelvis (gluteus maximus and gluteus medius), or the muscle strength of these muscles.

[0098] As an example, Figure 21 shows correspondence information 90 representing the correlation between muscle mass and fall rate for each age group. In the correspondence information 90 shown in Figure 21, the lower the muscle mass, the higher the fall rate at any age. Furthermore, even with the same muscle mass, the lower the bone mineral density, the higher the fall rate and the higher the probability of fracture in the event of a fall. For this reason, the information output device 50B of the third embodiment has age-specific correspondence information 90 for each bone mineral density, as shown in Figure 21, and multiple sets of correspondence information 90 are stored in the storage 53.

[0099] Furthermore, although it depends on the quality of the muscle, there is a tendency for greater muscle strength to be associated with greater muscle mass. For this reason, muscle strength may be used as a parameter to determine the fall rate instead of muscle mass.

[0100] Thus, the identification unit 67 of the third embodiment determines the degree of the subject falling over using correspondence information corresponding to the parts (imaging areas) of the subject H included in the first radiographic image G1 and the second radiographic image G2.

[0101] In the third embodiment, the fall rate derived by the information derivation device 50B is used as the correct answer data for the training data. Figure 22 shows the training data derived in the third embodiment. As shown in Figure 22, the training data 40B consists of training data 41B including the first and second radiation images G1 and G2, and correct answer data 42B which is the fall rate.

[0102] By training a neural network using the training data 40B shown in Figure 22, a trained neural network 23A can be constructed that, when a simple radiographic image G0 is input, outputs the fall rate as an estimated result related to the composition of soft tissue.

[0103] It should be noted that the invention is not limited to the first to third embodiments described above, and for example, the first to third embodiments may be combined. For example, the identification unit 67 according to the third embodiment may identify the risk of contracting a predetermined disease or the disease level of a predetermined disease as disease information based on the muscle mass of a predetermined area derived by the muscle mass derivation unit 64, the bone mineral density derived by the bone mineral density derivation unit 66, and the correspondence relationship information, and use this as training data. Alternatively, the identification unit 65 according to the second embodiment may identify the fall rate based on the muscle mass of a predetermined area derived by the muscle mass derivation unit 64 and the correspondence relationship information 90, and use this as training data.

[0104] Alternatively, muscle mass, disease information, and fall rate, or a combination of these, may be identified and used as training data. A trained neural network constructed by learning using such training data will, upon input of a simple radiographic image G0, output muscle mass, disease information, and fall rate, or a combination of these, as estimated results related to the composition of soft tissues.

[0105] Furthermore, the images used for deriving muscle mass in the first to third embodiments and for deriving bone mineral density in the third embodiment may be reduced images. For example, the muscle mass deriving unit 64 may derive muscle mass for each pixel of a reduced image obtained by reducing the soft tissue image Gs. Similarly, the bone mineral density deriving unit 66 may derive bone mineral density for each pixel of a reduced image obtained by reducing the bone image Gb. When reduced images are used in this way, noise can be reduced and the signal-to-noise ratio (SNR) can be improved, thereby improving the deriving accuracy.

[0106] Furthermore, in the first to third embodiments described above, training data is derived from the first and second radiographic images G1 and G2, but training data may also be derived from CT images acquired by the CT device 4. This will be described below as the fourth embodiment.

[0107] Figure 23 shows the functional configuration of the information extraction device according to the fourth embodiment. In Figure 23, the same reference numerals are used for components identical to those in Figure 7, and detailed explanations are omitted. As shown in Figure 23, the information extraction device 50C according to the fourth embodiment includes a synthesis unit 68 and a muscle mass extraction unit 69, replacing the scattered radiation removal unit 62, subtraction unit 63, and muscle mass extraction unit 64 of the information extraction device 50 according to the first embodiment.

[0108] In the fourth embodiment, the image acquisition unit 61 acquires a CT image V0 for deriving training data from the image storage system 9. The image acquisition unit 61 may acquire the CT image V0 by having the CT device 4 take a picture of the subject H, similar to the image acquisition unit 21 of the estimation device 10.

[0109] The synthesis unit 68 derives a synthesized two-dimensional image C0 representing the subject H by synthesizing the CT image V0. Figure 24 is a diagram illustrating the derivation of the synthesized two-dimensional image C0. Note that in Figure 24, the three-dimensional CT image V0 is shown in two dimensions for illustrative purposes. As shown in Figure 24, the subject H is contained within the three-dimensional space represented by the CT image V0. The subject H consists of multiple components: bone, fat, muscle, and internal organs.

[0110] Here, the CT value V0(x,y,z) at each pixel of the CT image V0 can be expressed by the following equation (7), using the attenuation coefficient μi of the composition at that pixel and the attenuation coefficient μw of water. (x,y,z) are coordinates representing the pixel position in the CT image V0. In the following explanation, unless otherwise specified, the attenuation coefficient refers to the linear attenuation coefficient. The attenuation coefficient represents the degree (percentage) to which radiation is attenuated by absorption or scattering, etc. The attenuation coefficient varies 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 (7)

[0111] The attenuation coefficient μw for water is known. Therefore, by solving equation (7) for μi, the attenuation coefficient μi for each composition can be calculated as shown in equation (8) below. μi=V0(x,y,z)×μw / 1000+μw (8)

[0112] As shown in Figure 24, the synthesis unit 68 virtually irradiates the subject H with radiation dose I0 and derives a synthesized two-dimensional image C0 by virtually detecting the radiation that has passed through the subject H using a radiation detector (not shown) installed on a virtual plane 80. The virtual radiation dose I0 and radiation energy are set according to predetermined shooting conditions. Specifically, the radiation dose I0 can be set by referring to a table corresponding to shooting conditions such as tube voltage, mAs value, and SID. Similarly, the radiation energy can be set by referring to a table corresponding to the tube voltage. In this case, the attained dose I1(x,y) for each pixel of the synthesized two-dimensional image C0 is the amount of radiation that has passed through one or more components in the subject H. Therefore, the attained dose I1(x,y) can be derived using the following equation (9) with the attenuation coefficient μi of one or more components through which the radiation of dose I0 passes. The attained dose I1(x,y) becomes the pixel value of each pixel of the synthesized two-dimensional image C0. I1(x,y)=I0×exp(-∫μi·dt) (9)

[0113] If the radiation source is assumed to be a surface light source, the attenuation coefficient μi used in equation (9) can be the one derived from equation (8) using the CT values ​​of each pixel arranged vertically as shown in Figure 24. Alternatively, if the surface light source is assumed to be a point light source, as shown in Figure 25, pixels on the path of radiation reaching each pixel can be identified based on the geometric positional relationship between the point light source and each position on the virtual plane 80, and the attenuation coefficient μi derived from equation (8) using the CT values ​​of the identified pixels can be used.

[0114] The muscle mass derivation unit 69 uses the CT image V0 to derive the muscle mass of the subject H for each pixel of the composite two-dimensional image C0. Here, we will explain the CT value. Figure 26 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 Figure 26, the CT value is set according to the composition of the human body, with water having a CT value of 0 and air having a CT value of -1000 (in HU).

[0115] The muscle mass derivation unit 69 first identifies muscle regions in the CT image V0 based on the CT values ​​of the CT image V0. Specifically, it identifies muscle regions by thresholding, which are areas consisting of pixels with CT values ​​between 60 and 70. Alternatively, instead of thresholding, a trained neural network that has been trained to detect muscle regions from the CT image V0 may be used to identify muscle regions. Furthermore, the CT image V0 may be displayed on the display 54, and muscle regions may be identified by accepting manual specification of muscle regions on the displayed CT image V0.

[0116] Here, the density per unit volume of the composition in each pixel of the CT image is ρ[g / cm³]. 3 ] is the attenuation coefficient μi[1 / cm] of the composition and the mass attenuation coefficient μe[cm] of the composition. 2 It can be derived from [ / g] by the following equation (10). ρ = μi / μe (10)

[0117] FIG. 27 is a diagram showing the relationship between radiation energy and mass attenuation coefficient in various components of the human body. FIG. 27 shows the relationship between radiation energy and mass attenuation coefficient for bone, muscle, etc. and fat. Note that muscle, etc. means muscle, blood, and water. In the present embodiment, the relationship between the radiation energy and the mass attenuation coefficient shown in FIG. 27 is stored as a table in the storage 53. In the present embodiment, since the mass attenuation coefficient of muscle is required, the mass attenuation coefficient of muscle is obtained by referring to the relationship for muscle, etc. in the table shown in FIG. 27 based on virtual radiation energy. Also, the attenuation coefficient μm at each pixel in the muscle region is derived by the above formula (8). Then, by the above formula (10), the muscle density ρm per unit volume at each pixel in the muscle region included in the CT image V0 is derived as the muscle mass.

[0118] Note that since the CT image V0 is a three-dimensional image, the unit of the muscle mass per unit volume derived by the above formula (10) is [g / cm 3 . In the present embodiment, the muscle mass derivation unit 69 derives the muscle mass per unit area for each pixel of the synthesized two-dimensional image C0. For this reason, the muscle mass derivation unit 69 projects the muscle density ρm per unit volume derived by the above formula (10) onto the virtual plane 80 in the same manner as when the synthesized two-dimensional image C0 is derived, so as to derive the muscle mass M [g / cm 2 per unit area for each pixel of the synthesized two-dimensional image C0.

[0119] During projection, a representative value of the muscle mass of each pixel in the CT image V0, which is located along the path from a virtual radiation source to each pixel of the composite 2D image C0, can be derived. The representative value can be the cumulative value, mean, maximum, median, or minimum. Furthermore, in this embodiment, the muscle mass derivation unit 69 can derive a representative value of the muscle mass for predetermined areas. For example, if the predetermined areas are the buttocks and upper lower limbs, the muscle mass derivation unit 69 derives a representative value of the muscle mass of each pixel in the region of the buttocks and upper lower limbs in the composite 2D image C0. The representative value can be the mean, median, minimum, or maximum. In this embodiment, the representative value of the muscle mass of the predetermined areas, the buttocks and upper lower limbs, is used as the ground truth data.

[0120] In the fourth embodiment, the muscle mass derived by the information derivation device 50C is used as the correct answer data for the training data. Figure 28 shows the training data derived in the fourth embodiment. As shown in Figure 28, the training data 40C consists of training data 41C including a synthesized two-dimensional image C0 and correct answer data 42C which is muscle mass.

[0121] By training a neural network using the training data 40C shown in Figure 28, a trained neural network 23A can be constructed that, similar to the first embodiment, outputs muscle mass as an estimated result related to the composition of soft tissue when a simple radiographic image G0 is input.

[0122] In the fourth embodiment described above, muscle mass is derived from CT image V0 as the ground truth data, but this is not the only way to do so. Similar to the second embodiment, disease information representing the risk of developing a predetermined disease or the disease level of a predetermined disease may be derived as the ground truth data. This will be described below as the fifth embodiment.

[0123] Figure 29 shows the functional configuration of the information extraction device according to the fifth embodiment. In Figure 29, the same reference numerals are used for components identical to those in Figure 23, and detailed explanations are omitted. As shown in Figure 29, the information extraction device 50D according to the fifth embodiment further includes a specific unit 65A that performs the same processing as the specific unit 65 of the information extraction device 50A according to the second embodiment, in addition to the information extraction device 50C according to the fourth embodiment.

[0124] In the fifth embodiment, the identification unit 65A identifies disease information representing the risk of contracting a predetermined disease or the disease level of a predetermined disease, based on the muscle mass of a predetermined area derived from the CT image V0 by the muscle mass derivation unit 69 and the corresponding relationship information.

[0125] In the fifth embodiment, the disease information derived by the information derivation device 50D is used as the correct answer data for the training data. Figure 30 shows the training data derived in the fifth embodiment. As shown in Figure 30, the training data 40D consists of training data 41D including a synthesized two-dimensional image C0 and correct answer data 42D which is disease information.

[0126] By training a neural network using the training data 40D shown in Figure 30, a trained neural network 23A can be constructed that, similar to the second embodiment, outputs disease information as an estimated result related to the composition of soft tissue when a simple radiographic image G0 is input.

[0127] In the fourth embodiment described above, muscle mass is derived from CT image V0 as the ground truth data, but this is not the only way to do so. Similar to the third embodiment, the fall rate may be derived as the ground truth data. This will be described below as the sixth embodiment.

[0128] Figure 31 shows the functional configuration of the information extraction device according to the sixth embodiment. In Figure 31, the same reference numerals are used for components identical to those in Figure 23, and detailed explanations are omitted. As shown in Figure 31, the information extraction device 50E according to the sixth embodiment further includes a bone mineral content extraction unit 81 and a specific unit 67A that performs the same processing as the specific unit 67 of the information extraction device 50B according to the third embodiment, compared to the information extraction device 50C according to the fourth embodiment.

[0129] The bone mineral density deriving unit 81 uses the CT image V0 to derive the bone mineral density of the subject H for each pixel in the synthesized two-dimensional image C0. Specifically, the bone mineral density deriving unit 81 first identifies the bone region in the CT image V0 based on the CT values ​​of the CT image V0. More precisely, it identifies the bone region as a region consisting of pixels with CT values ​​between 100 and 1000 by thresholding. Alternatively, instead of thresholding, the bone region may be identified using a trained neural network that has been trained to detect the bone region from the CT image V0. Furthermore, the CT image V0 may be displayed on the display 54, and the bone region may be identified by accepting manual specification of the bone region in the displayed CT image V0.

[0130] In the sixth embodiment, since the mass attenuation coefficient of the bone region is required, the mass attenuation coefficient of the bone region is obtained by referring to the relationship for the bone region in the table shown in Figure 27 based on a hypothetical radiation energy. The attenuation coefficient μb for each pixel in the bone region is derived using equation (8) above. Then, the bone mineral content ρb per unit volume for each pixel in the bone region included in the CT image V0 is derived using equation (10) above.

[0131] Since CT image V0 is a three-dimensional image, the unit of bone mineral content per unit volume derived by the above formula (10) is [g / cm³]. 3In the sixth embodiment, the bone mineral content extraction unit 81 extracts the bone mineral content per unit area for each pixel of the composite two-dimensional image C0. To this end, the bone mineral content extraction unit 81 projects the bone mineral content ρb per unit volume extracted by the above formula (10) onto a virtual plane 80, similar to when the composite two-dimensional image C0 was extracted, thereby determining the bone mineral content B [g / cm³] per unit area for each pixel of the composite two-dimensional image C0. 2 Derive ].

[0132] Furthermore, during projection, a representative value of the bone mineral content of each pixel in the CT image V0, which is located along the path from a virtual radiation source to each pixel of the composite two-dimensional image C0, should be derived. As representative values, the cumulative value, mean, maximum, median, and minimum can be used. Moreover, in the sixth embodiment, the bone mineral content derivation unit 81 only needs to derive representative values ​​of the bone mineral content for predetermined areas. In this embodiment, the predetermined areas are the buttocks and the upper part of the lower limbs, so the bone mineral content derivation unit 81 derives representative values ​​of the bone mineral content of each pixel in the region of the buttocks and the upper part of the lower limbs in the composite two-dimensional image C0. As representative values, the mean, median, minimum, and maximum can be used.

[0133] In the sixth embodiment, the identification unit 67A determines the fall rate based on the muscle mass of a predetermined area derived from the CT image V0 by the muscle mass derivation unit 69, the bone mineral density of a predetermined area derived from the bone mineral density derivation unit 81, and the correspondence relationship information 90 shown in Figure 21.

[0134] In the sixth embodiment, the fall rate derived by the information derivation device 50E is used as the correct answer data for the training data. Figure 32 shows the training data derived in the sixth embodiment. As shown in Figure 32, the training data 40E consists of training data 41E including a synthesized two-dimensional image C0 and correct answer data 42E which is the fall rate.

[0135] By training a neural network using the training data 40E shown in Figure 32, a trained neural network 23A can be constructed that, when a simple radiographic image G0 is input, outputs the fall rate as an estimated result related to the composition of soft tissue.

[0136] It should be noted that the embodiments are not limited to the fourth to sixth embodiments described above, and for example, the embodiments of the fourth to sixth embodiments may be combined. For example, the identification unit 67A according to the sixth embodiment may be configured to identify disease information representing the risk of contracting a predetermined disease or the disease level of a predetermined disease, based on the muscle mass of a predetermined area derived by the muscle mass derivation unit 69, the bone mineral density of a predetermined area derived by the bone mineral density derivation unit 81, and correspondence relationship information, and use this as training data. Alternatively, the identification unit 65A according to the fifth embodiment may be configured to identify the fall rate, based on the muscle mass of a predetermined area derived by the muscle mass derivation unit 69 and correspondence relationship information 90, and use this as training data.

[0137] Alternatively, muscle mass, disease information, and fall rate, or a combination of these, may be identified and used as training data. A trained neural network constructed by learning using such training data will, upon input of a simple radiographic image G0, output muscle mass, disease information, and fall rate, or a combination of these, as estimated results related to the composition of soft tissues.

[0138] Furthermore, while muscle mass is derived as the ground truth data in the first embodiment described above, the method is not limited to this. Fat mass may also be derived as the ground truth data. This will be described below as the seventh embodiment.

[0139] Figure 33 shows the functional configuration of the information extraction device according to the seventh embodiment. In Figure 33, the same reference numerals are used for components identical to those in Figure 7, and detailed explanations are omitted. As shown in Figure 33, the information extraction device 50F according to the seventh embodiment is equipped with a fat mass extraction unit 82 in place of the muscle mass extraction unit 64 included in the information extraction device 50 according to the first embodiment.

[0140] The fat mass extraction unit 82 separates muscle and fat from the soft tissue image Gs by utilizing the difference in energy characteristics between muscle tissue and fat tissue as described above. That is, while the muscle mass extraction unit 64 in the first embodiment extracts a muscle image from the soft tissue image Gs, the fat mass extraction unit 82 generates a fat image from the soft tissue image Gs. Furthermore, the fat mass extraction unit 82 extracts the fat mass of each pixel based on the pixel values ​​of the fat image.

[0141] The fat mass deriving unit 82 first derives the muscle ratio rm(x,y) at each pixel position (x,y) in the soft tissue image Gs using the above formula (2), similar to the muscle mass deriving unit 64 in the first embodiment. Then, by subtracting the muscle ratio rm(x,y) from 1, the fat ratio rf(x,y) (=1-rm(x,y)) is derived.

[0142] Furthermore, the fat volume extraction unit 82 generates a fat image Gf from the soft tissue image Gs using the following equation (11). In equation (11), x and y are the coordinates of each pixel in the fat image Gf. Gf(x,y) = rf(x,y) × Gs(x,y) (11)

[0143] The fat amount derivation unit 82 can derive the pixel values ​​of the fat image Gf as the fat amount for each pixel. Alternatively, the fat percentage rf for each pixel may be derived as the fat amount. Furthermore, as shown in equation (12) below, the fat image Gf can be obtained by multiplying each pixel (x,y) of the fat image Gf by a predetermined coefficient K2(x,y) that represents the relationship between the pixel value and the fat amount, thereby obtaining the fat amount F(x,y) (g / cm³) for each pixel of the fat image Gf. 2Alternatively, the following can be derived: In this case, the amount of fat will be the amount of fat per unit area, but it may also be used to derive the amount of fat per unit volume. F(x,y) = K²(x,y) × Gf(x,y) (12)

[0144] In the seventh embodiment, the amount of fat derived by the information derivation device 50F is used as the correct answer data for the training data. Figure 34 shows the training data derived in the seventh embodiment. As shown in Figure 34, the training data 40F consists of training data 41F including the first and second radiation images G1 and G2, and correct answer data 42F which is the amount of fat.

[0145] By training a neural network using the training data 40F shown in Figure 34, a trained neural network 23A can be constructed that, when a simple radiographic image G0 is input, outputs fat mass as an estimated result related to the composition of soft tissue.

[0146] Furthermore, while muscle mass is derived as the ground truth data in the fourth embodiment described above, the method is not limited to this. Fat mass may also be derived as the ground truth data. This will be described below as the eighth embodiment.

[0147] Figure 35 shows the functional configuration of the information extraction device according to the eighth embodiment. In Figure 35, the same reference numerals are used for components identical to those in Figure 7, and detailed explanations are omitted. As shown in Figure 35, the information extraction device 50G according to the eighth embodiment is equipped with a fat mass extraction unit 83 in place of the muscle mass extraction unit 69 included in the information extraction device 50C according to the fourth embodiment.

[0148] The fat mass deriving unit 83 uses the CT image V0 to derive the fat mass of the subject H for each pixel in the synthesized two-dimensional image C0. Specifically, the fat mass deriving unit 83 first identifies fat regions in the CT image V0 based on the CT values ​​of the CT image V0. More precisely, threshold processing identifies regions consisting of pixels with a CT value of approximately -100, for example, -100±10, as fat regions. Alternatively, instead of threshold processing, a trained neural network trained to detect fat regions from the CT image V0 may be used to identify fat regions. Furthermore, the CT image V0 may be displayed on the display 54, and fat regions may be identified by accepting manual specification of fat regions in the displayed CT image V0.

[0149] The fat volume derivation unit 83 then obtains the fat mass attenuation coefficient by referring to the relationship for fat shown in Figure 27. It also derives the attenuation coefficient μf for each pixel in the fat region using equation (8) above. Then, using equation (10) above, it derives the fat volume, which is the fat density ρf per unit volume for each pixel in the fat region included in the CT image V0.

[0150] Since CT image V0 is a three-dimensional image, the unit of the amount of fat per unit volume derived by the above formula (10) is [g / cm³]. 3 In this embodiment, the fat amount derivation unit 83 derives the amount of fat per unit area for each pixel of the composite two-dimensional image C0. To this end, the fat amount derivation unit 83 projects the fat density ρf per unit volume derived by the above formula (10) onto a virtual plane 80, similar to when the composite two-dimensional image C0 was derived, thereby determining the amount of fat F [g / cm³] per unit area for each pixel of the composite two-dimensional image C0. 2 Derive ].

[0151] During projection, a representative value of the fat content of each pixel in the CT image V0, which is located along the path from a virtual radiation source to each pixel of the composite two-dimensional image C0, can be derived. The representative value can be the cumulative value, mean, maximum, median, or minimum. Furthermore, in this embodiment, the fat content deriving unit 83 can derive a representative value of the fat content for a predetermined area. For example, if the predetermined area is the abdomen, the fat content deriving unit 83 derives a representative value of the fat content of each pixel in the abdominal region of the composite two-dimensional image C0. The representative value can be the mean, median, minimum, or maximum.

[0152] In the eighth embodiment, the amount of fat derived by the information derivation device 50G is used as the correct answer data for the training data. Figure 36 shows the training data derived in the eighth embodiment. As shown in Figure 36, the training data 40G consists of training data 41G including a synthesized two-dimensional image C0 and correct answer data 42G which is the amount of fat.

[0153] By training a neural network using the training data 40G shown in Figure 36, a trained neural network 23A can be constructed that, similar to the seventh embodiment, outputs fat mass as an estimated result related to the composition of soft tissue when a simple radiographic image G0 is input.

[0154] Furthermore, while the above embodiments derive estimation results related to the composition of soft tissue from simple radiation images G0, the invention is not limited to this. For example, the technology of this disclosure can also be applied when deriving estimation results related to the composition of soft tissue from DXA scan images obtained by photographing a subject with a DXA imaging device described in Japanese Patent Publication No. 9-108206 and Japanese Patent Publication No. 2006-271437, etc. A DXA scan image refers to a radiation image captured by a radiation detector when a subject is irradiated with a narrowly collimated high-energy radiation beam and a low-energy radiation beam while switching and scanning. A narrowly collimated radiation beam is, for example, a radiation beam formed into a pencil beam, narrow fan beam, or wide fan beam using a collimator between the radiation source and the subject. Low-energy radiation refers to radiation with energy relatively lower than high-energy radiation.

[0155] In this case, an image simulating a DXA scan image may be generated from a synthesized 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 relative distances between the X-ray source, the subject, and the detector, and the energy distribution of the radiation (determined by the tube voltage, target, and filter). The generated image simulating a DXA scan image may then be used as training data 41 to construct a trained neural network 23A.

[0156] An image simulating a DXA scan image can be generated, for example, by applying a process to the composite two-dimensional image C0 derived by the composite unit 68 in the fourth to sixth embodiments to reduce its resolution according to the pixel size, scanning direction, and scanning speed of the detector used to capture the DXA scan image. An image simulating a DXA scan image is an example of a low-resolution composite two-dimensional image.

[0157] Specifically, an image simulating a DXA scan image is generated as follows: Let L, M, and N be natural numbers, and assume that M × M pixels of the synthesized 2D image C0 correspond to the actual size L mm × L mm of the subject H, and that N × N pixels of the training image of the DXA scan image correspond to each of these. In this case, the synthesized 2D image C0 is reduced in resolution by setting the average value of all (M / N) × (M / N) pixels of the synthesized 2D image C0 to the average value of all (M / N) × (M / N) pixels of the synthesized 2D image C0, so that (M / N) × (M / N) pixels of the synthesized 2D image C0, i.e., multiple adjacent pixels, correspond to one pixel of the training image of the DXA scan image. Then, by performing this resolution reduction process on all regions of the synthesized 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 the corresponding pixels in the synthesized 2D image C0 and the training image of the DXA scan can be adjusted appropriately using natural numbers before and after M / N to generate an image that simulates the DXA scan from the synthesized 2D image C0.

[0158] Furthermore, as a low-resolution process to simulate blurring caused by 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.

[0159] Alternatively, an image simulating a DXA scan image may be generated by performing a moving average operation on the synthesized 2D image C0. For the moving average operation, the size of the filter used to calculate the moving average and the intensity distribution of the filter should be appropriately determined based on the scanning direction, scanning speed, detector pixel size, and the relative distances between the X-ray source, the subject, and the detector during DXA scan image acquisition. For example, the faster the scanning speed, the lower the resolution, so the filter size should be set to be relatively larger. In this case, if L=10, M=200 and N=5 would be approximately.

[0160] In the embodiments described above, the predetermined areas are the buttocks, lower limbs, and limbs, but are not limited to these. In these embodiments, it is possible to derive estimation results related to the composition of soft tissue for any area.

[0161] Furthermore, in the first to third and seventh embodiments described above, the first and second radiographic images G1 and G2 are used as training data 41 for the training data 40, but the system is not limited to this. As shown in the training data 40H in Figure 37, instead of the second radiographic image G2, the soft tissue image Gs derived by the subtraction unit 63 may be used as training data 41H. In Figure 37, the correct answer data 42H is muscle mass.

[0162] Furthermore, in the first to third and seventh embodiments described above, the first radiation image G1 and the second radiation image G2 themselves are used to derive muscle mass as ground truth data, but the method is not limited to this. Muscle mass may also be derived using the first radiation image G1 and the second radiation image G2, in which a moving average is calculated for each pixel of the first radiation image G1 and the second radiation image G2 with respect to surrounding pixels, and the moving average is used as the pixel value for each pixel. In this case, the pixels to be used for the moving average can be appropriately determined from information on the relative distances between the radiation source 3, the subject H, and the radiation detectors 5 and 6, as well as information on the pixel sizes of the radiation detectors 5 and 6.

[0163] In each of the above embodiments, the training of the neural network is performed in the estimation device 10 to construct the trained neural network 23A, but this is not the only way to do so. The 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.

[0164] Furthermore, in the first to third and seventh embodiments described above, the first and second radiographic images G1 and G2 are acquired by the one-shot method when performing energy subtraction processing to derive muscle mass, but the method is not limited to this. The first and second radiographic images G1 and G2 may also be acquired by the so-called two-shot method, in which imaging is performed twice using only one radiation detector. In the case of the two-shot method, the position of subject H included in the first radiographic image G1 and the second radiographic image G2 may shift due to the movement of the subject H. For this reason, it is preferable to perform positional alignment of the subject in the first radiographic image G1 and the second radiographic image G2 before performing the processing of this embodiment.

[0165] For the alignment process, for example, the method described in Japanese Patent Publication No. 2011-255060 can be used. The method described in Japanese Patent Publication No. 2011-255060 generates multiple first band images and multiple second band images representing structures with different frequency bands for each of the first and second radiation images G1 and G2, obtains the amount of positional displacement of corresponding positions in the first and second band images of the corresponding frequency bands, and aligns the first radiation image G1 and the second radiation image G2 based on the amount of positional displacement.

[0166] Furthermore, in each of the above embodiments, information related to the composition of soft tissue as ground truth data for training data is derived using radiation images acquired in a system that photographs subject H using first and second radiation detectors 5 and 6. However, instead of radiation detectors, information related to the composition of soft tissue as ground truth data may be derived from first and second radiation images G1 and G2 acquired using a accumulative phosphor sheet. In this case, two accumulative phosphor sheets can be stacked and irradiated with radiation that has passed through subject H to accumulate and record radiation image information of subject H on each accumulative phosphor sheet, and the first and second radiation images G1 and G2 can be acquired by photoelectrically reading the radiation image information from each accumulative phosphor sheet. Note that the two-shot method may also be used when acquiring the first and second radiation images G1 and G2 using a accumulative phosphor sheet.

[0167] Furthermore, the radiation in the above embodiment is not particularly limited, and in addition to X-rays, alpha rays or gamma rays can be used.

[0168] Furthermore, in the above embodiment, the hardware structure of the 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, scattered radiation removal unit 62, subtraction unit 63, and muscle mass extraction unit 64 of the information extraction device 50, can be the various processors shown below. As mentioned above, the various processors include a CPU, which is a general-purpose processor that executes software (programs) and functions as various processing units, as well as a programmable logic device (PLD), such as an FPGA (Field Programmable Gate Array), whose circuit configuration can be changed after manufacturing, and a dedicated electrical circuit, such as an ASIC (Application Specific Integrated Circuit), which has a circuit configuration specifically designed to perform a particular process.

[0169] A single processing unit may consist of one of these various processors, or it may consist of a combination of two or more processors of the same or different types (for example, a combination of multiple FPGAs or a combination of a CPU and an FPGA). Alternatively, multiple processing units may be composed of a single processor.

[0170] Examples of configuring multiple processing units with a single processor include, firstly, a configuration where one or more CPUs and software combine to form a single processor, as exemplified by client and server computers, and this processor functions as multiple processing units. Secondly, a configuration using a processor that realizes the functions of the entire system, including multiple processing units, on a single IC (Integrated Circuit) chip, as exemplified by System-on-a-Chip (SoC). Thus, various processing units are configured, in terms of hardware structure, using one or more of the above-mentioned processors.

[0171] Furthermore, the hardware structure of these various processors can more specifically utilize electrical circuits (Circuitry) that combine circuit elements such as semiconductor elements. [Explanation of Symbols]

[0172] 1. Imaging device 3 Radiation source 4 CT device 5, 6 Radiation detectors 7. Radiation energy conversion filter 9. Image storage system 10 Estimation device 11.51 CPU 12A Estimation Program 12B Learning Program 13, 53 storage 14.54 displays 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 Pre-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, 40D, 40E, 40F, 40G, 40H Training data 41, 41A, 41B, 41C, 41D, 41E, 41F, 41G, 41H Training data 42, 42A, 42B, 42C, 42D, 42E, 42F, 42G, 42H Correct data 47 Output data 48 parameters 50,50A, 50B, 50C, 50D, 50E, 50F, 50G information derivation device 52 Information Derivation Program 61 Image acquisition unit 62 Scattered radiation removal section 63 Subtraction section 64 Muscle mass extraction unit 65, 65A specific part 66 Bone mineral content extraction section 67, 67A specific part 68 Synthesis section 69 Muscle mass extraction unit 70 display screen 71 Image display area 72 Muscle mass display area 73 References 80 planes 81 Bone mineral content extraction section 82, 83 Fat amount derivation part 90 Correspondence Information C0 Composite 2D Image G0 Simple Radiation Image G1, G2 radiographic images Gb bone image Gs Soft tissue images LUT1 Table

Claims

1. at least one processor; The processor: The neural network functions as a trained neural network that derives an estimation result related to the composition of the soft tissue of a subject from a simple radiographic image obtained by simply photographing the subject or a DXA scanned image obtained by photographing the subject using a DXA method; The trained neural network is trained using, as training data, (i) two radiological images obtained by photographing a subject with radiation having different energy distributions, (ii) a radiological image of the subject and a soft tissue image representing the soft tissue of the subject, or (iii) a composite two-dimensional image representing the subject derived by combining three-dimensional CT images of the subject, and information related to the composition of the soft tissue of the subject.

2. The estimation device according to claim 1, wherein the information relating to the composition of the soft tissue is derived based on pixel values ​​of the soft tissue region in soft tissue images derived from two radiological images obtained by photographing a subject with radiation having different energy distributions.

3. The estimation device of claim 1, wherein the information related to the composition of the soft tissue is derived by identifying a soft tissue region in the CT image, deriving a radiation attenuation coefficient in the soft tissue region, and deriving the density of the composition of the soft tissue based on the radiation attenuation coefficient and the mass attenuation coefficient in the soft tissue region.

4. The estimation device according to claim 3 , wherein the information relating to the composition of the soft tissue is derived by projecting the density of the composition at each position in the soft tissue region in a predetermined direction.

5. 5. The estimation device according to claim 1, wherein the information relating to the composition of the soft tissue includes at least one of muscle mass per unit area, muscle mass per unit volume, fat mass per unit area, fat mass per unit volume, disease information representing the risk of contracting a predetermined disease or the disease level of a predetermined disease, and fall rate.

6. The disease information is obtained by deriving the muscle mass of a predetermined part of the subject; The estimation device according to claim 5, wherein correspondence information represents a correspondence between disease information representing the risk of contracting a predetermined disease or the disease level of a predetermined disease and the muscle mass of the predetermined part, and the disease information is derived by identifying the disease information based on the muscle mass.

7. The disease information is derived by deriving the muscle mass and bone mineral content of a predetermined part of the subject; The estimation device described in claim 5, wherein the disease information is derived by identifying the disease information based on correspondence information representing the correspondence between disease information representing the risk of contracting a predetermined disease or the disease level of a predetermined disease and the muscle mass of the predetermined area, the muscle mass, and the bone mineral content.

8. the predetermined site is a lower leg, the predetermined disease is diabetes; The estimation device according to claim 6 or 7, wherein the disease information is the risk of developing diabetes.

9. the predetermined site is a limb or the whole body, the predetermined disease is sarcopenia; The estimation device according to claim 6 or claim 7, wherein the disease information is the disease level of the sarcopenia.

10. The fall rate is calculated by deriving the muscle mass of a predetermined part of the subject; 6. The estimation device according to claim 5, wherein the fall rate is derived by identifying correspondence information that represents a correspondence relationship between muscle information, which is muscle mass of the predetermined part or muscle strength corresponding to the muscle mass of the predetermined part, and the fall rate of the subject, and the muscle mass.

11. The fall rate is calculated by deriving the muscle mass and bone mineral content of a predetermined part of the subject; 6. The estimation device according to claim 5, wherein the fall rate is derived by identifying the fall rate based on correspondence relationship information that represents a correspondence relationship between muscle information, which is muscle mass of the predetermined part or muscle strength corresponding to the muscle mass of the predetermined part, and the fall rate of the subject, the muscle mass, and the bone mineral content.

12. The estimation device according to claim 10 or 11, wherein the predetermined region is at least one of a lower leg and a buttock.

13. the processor functions as a trained neural network that derives inferences related to the composition of the soft tissue from the DXA scan images; 13. The estimation device according to claim 1, wherein the trained neural network is trained using, as training data, a low-resolution composite two-dimensional image obtained by performing processing to reduce the resolution of the composite two-dimensional image and information related to the composition of the subject's soft tissue.

14. The estimation device according to claim 13, 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.

15. The estimation device according to claim 13, 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.

16. 14. The estimation device according to claim 13, 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 a size of the adjacent plurality of pixels corresponds to a pixel size of one pixel in the DXA scanned image, and the one direction corresponds to a scanning direction of the DXA scanned image.

17. An estimation method for deriving an estimation result related to the composition of soft tissues of a subject from a plain radiographic image obtained by simple radiography of the subject or a DXA scanned image obtained by photographing the subject using a trained neural network that derives an estimation result related to the composition of the soft tissues of the subject from the plain radiographic image or the DXA scanned image, the method comprising: The trained neural network is trained using, as training data, (i) two radiological images obtained by photographing a subject with radiation having different energy distributions, (ii) a radiological image of the subject and a soft-tissue image representing the subject's soft tissue, or (iii) a composite two-dimensional image representing the subject derived by combining three-dimensional CT images of the subject, and information related to the composition of the subject's soft tissue.

18. An estimation program that causes a computer to execute a procedure for deriving an estimation result related to the composition of soft tissues of a subject from a plain radiographic image obtained by plain radiography of the subject or a DXA scanned image obtained by photographing the subject using a trained neural network that derives an estimation result related to the composition of the soft tissues of the subject from the plain radiographic image or the DXA scanned image, The trained neural network is an estimation program that is trained using, as training data, (i) two radiological images obtained by photographing a subject with radiation having different energy distributions, (ii) a radiological image of the subject and a soft tissue image representing the soft tissue of the subject, or (iii) a composite two-dimensional image representing the subject derived by combining three-dimensional CT images of the subject, and information related to the composition of the soft tissue of the subject.

Images