Trained model generation program, image generation program, trained model generation device, image generation device, trained model generation method, and image generation method

Abstract

A trained model generation program causes a computer to function as a training execution function that executes machine training by inputting first input image data and second input image data to a machine training device to execute machine training, and causes the machine training device to generate a trained model, wherein the first input image data shows a first input image generated by a first reconstruction method, the second input image data shows a second input image generated by a second reconstruction method, the first reconstruction method uses compressed sensing, and the second reconstruction method is a different reconstruction method from the first reconstruction method and is an analytical reconstruction method; and a trained model acquisition function that acquires trained model data showing the trained model. Furthermore, a reconstructed image with improved image quality is generated by inputting input image data showing an input image to the trained model.

Description

TECHNICAL FIELD

[0001] The present application relates to a trained model generation program, an image generation program, a trained model generation device, an image generation device, a trained model generation method, and an image generation method. BACKGROUND

[0002] In order to reduce the dose of radiation to the subject, for example, an X-ray CT apparatus that performs sparse view CT in which the number of projection directions is reduced to scan the subject or low dose CT in which the subject is scanned in a state in which the current flowing through the X-ray tube is suppressed is used. A reconstruction method using compressed sensing is used as a main reconstruction method of generating a high-definition reconstructed image when these scans are performed. This reconstruction method generates a reconstructed image by using a model in which the concentration change of the reconstructed image is piecewise the same. In addition, even in such a reconstruction method, a reconstruction method called total variation (TV) is frequently used in a medical X-ray CT apparatus.

[0003] As one example of such a technique, a medical imaging apparatus disclosed in Patent Literature 1 can be cited. The medical imaging apparatus has an image reconstruction section that reconstructs an image by iterative optimization calculation of compressed sensing and a basis selection section that selects a basis conversion for optimization at each iteration. The basis selection section selects a basis in the order of a basis set in advance. Alternatively, the basis selection section selects a basis using a weight coefficient set in advance for the basis.

[0004] PRIOR ART DOCUMENTS

[0005] PATENT LITERATURE

[0006] Patent Literature 1: Japanese Patent Application Publication No. 2018-134199 SUMMARY

[0007] (PROBLEMS TO BE SOLVED BY THE INVENTION)

[0008] However, in the case of the above-described medical imaging apparatus, since it adopts compressed sensing, for example, in the case of performing sparse view CT or low dose CT, problems such as trapezoidal artifacts, disappearance of smooth concentration change, disappearance of texture, and the like occur on the reconstructed image. It is known that such a degradation in image quality is significantly exhibited depending on the measurement conditions of sparse view CT or low dose CT.

[0009] The present application has been achieved in view of the above-described problems, and provides a trained model generation program, an image generation program, a trained model generation device, an image generation device, a trained model generation method, and an image generation method capable of mitigating quality degradation occurring on a reconstructed image generated by a reconstruction method using compressed sensing.

[0010] (Technical means for solving the technical problem)

[0011] One embodiment of the present application is a trained model generation program characterized by causing a computer to function as: a training execution function that executes machine training by inputting first input image data and second input image data to a machine training device, and causing the machine training device to generate a trained model, wherein the first input image data shows a first input image generated by a first reconstruction method, the second input image data shows a second input image generated by a second reconstruction method, the first reconstruction method uses compressed sensing, and the second reconstruction method is a different reconstruction method from the first reconstruction method and is an analytical reconstruction method; and a trained model acquisition function that acquires trained model data showing the trained model.

[0012] One embodiment of the present application is the above-described trained model generation program, which can cause the machine training device to generate the trained model by exploring a smoothing parameter of the first reconstruction method, a weight coefficient used in a convolutional neural network serving as the trained model, and a bias used in the convolutional neural network so as to minimize a mean square error represented by the following expression (1) as much as possible,

[0013] [Mathematical expression 1]

[0014]

[0015] wherein:

[0016] w: a vector in which weight coefficients used in a convolutional neural network are arranged in a column,

[0017] b: a vector in which biases used in a convolutional neural network are arranged in a column,

[0018] β: a smoothing parameter of a first reconstruction method,

[0019] x i : a vector in which values showing concentrations represented by respective pixels of an output image are arranged in a column,

[0020] y i : a vector in which values showing concentrations represented by respective pixels of a second input image are arranged in a column,

[0021] zi : A vector in which values showing concentrations represented by respective pixels of a first input image are arranged in a column.

[0022] One embodiment of the present application is the above-described trained model generation program, and the training execution function can input, to the machine training device, at least the first input image data showing the first input image generated by using the first reconstruction method of the compressed sensing in which the first smoothing parameter is set, the first input image data showing the first input image generated by using the first reconstruction method of the compressed sensing in which the second smoothing parameter different from the value of the first smoothing parameter is set, to perform machine training, and cause the machine training device to generate the trained model.

[0023] One embodiment of the present application is the above-described trained model generation program, and the training execution function can input, to the machine training device, at least the first input image data showing the first input image generated by using the first reconstruction method of the compressed sensing in which the first smoothing parameter is set, the first input image data showing the first input image generated by using the first reconstruction method of the compressed sensing in which the second smoothing parameter different from the value of the first smoothing parameter is set, to perform machine training, and cause the machine training device to generate the trained model.

[0024] One embodiment of the present application is an image generation program characterized by causing a computer to realize an image generation function of generating a reconstructed image by inputting input image data showing an input image to a trained model generated by any one of the above-described trained model generation programs.

[0025] One embodiment of the present application is a trained model generation device characterized by including: a training execution unit that performs machine training by inputting first input image data and second input image data to a machine training device, and causes the machine training device to generate a trained model, wherein the first input image data shows a first input image generated by a first reconstruction method, the second input image data shows a second input image generated by a second reconstruction method, the first reconstruction method uses compressed sensing, and the second reconstruction method is a reconstruction method different from the first reconstruction method and is an analytical reconstruction method; and a trained model acquisition unit that acquires trained model data showing the trained model.

[0026] One embodiment of the present application is an image generation device characterized by including an image generation unit that generates a reconstructed image by inputting input image data showing an input image to a trained model generated by the above-described trained model generation device.

[0027] One embodiment of the present invention is a method for generating a trained model, characterized by comprising the following steps: a training execution step, which involves inputting first input image data and second input image data into a machine training device to perform machine training, and causing the machine training device to generate a trained model, wherein the first input image data shows a first input image generated by a first reconstruction method, and the second input image data shows a second input image generated by a second reconstruction method, wherein the first reconstruction method uses compressed sensing, and the second reconstruction method is a different reconstruction method from the first reconstruction method, being an analytical reconstruction method; and a trained model acquisition step, which acquires trained model data showing the trained model.

[0028] One embodiment of the present invention is an image generation method, characterized by comprising an image generation step, wherein the image generation step generates a reconstructed image by inputting input image data showing an input image into the trained model generated by the above-described trained model generation method.

[0029] (The effect of the invention)

[0030] According to the present invention, it is possible to alleviate the image quality degradation that occurs in reconstructed images generated by using a reconstruction method based on compressed sensing. Attached Figure Description

[0031] Figure 1 A diagram illustrating an example of an image generation system according to an embodiment.

[0032] Figure 2 A diagram illustrating an example of an X-ray CT apparatus according to an embodiment.

[0033] Figure 3 A diagram illustrating an example of machine training performed by the machine training apparatus according to the embodiments.

[0034] Figure 4 A diagram illustrating an example of machine training performed by the machine training apparatus according to the embodiments.

[0035] Figure 5 A diagram illustrating an example of machine training performed by the machine training apparatus according to the embodiments.

[0036] Figure 6 This is a flowchart illustrating an example of the processing performed by the trained model generation procedure according to the implementation method.

[0037] Figure 7 This is a diagram illustrating an example of a reconstructed image generated from projection data obtained through scans with a pre-set normal dose and number of projection data directions.

[0038] Figure 8 FIG. 1 is a diagram to show one example of a reconstructed image generated by reconstructing projection data acquired by sparse view CT using a reconstruction method of total variation.

[0039] Figure 9 FIG. 2 is a diagram to show one example of a reconstructed image generated using an existing trained model.

[0040] Figure 10 FIG. 3 is a diagram to show one example of a reconstructed image generated using a trained model generated by inputting first input image data and second input image data related to the embodiment to a machine training device.

[0041] Figure 11 FIG. 4 is a diagram to show one example of a reconstructed image generated using a trained model generated by inputting first input image data related to the embodiment to a machine training device.

[0042] Figure 12 FIG. 5 is a diagram to show one example of a reconstructed image generated by reconstructing projection data acquired by a scan in which a normal dose and a number of projection data directions have been set.

[0043] Figure 13 FIG. 6 is a diagram to show one example of a reconstructed image generated by reconstructing projection data acquired by sparse view CT using a reconstruction method of total variation.

[0044] Figure 14 FIG. 7 is a diagram to show one example of a reconstructed image generated using an existing trained model.

[0045] Figure 15 FIG. 8 is a diagram to show one example of a reconstructed image generated using a trained model generated by inputting first input image data and second input image data related to the embodiment to a machine training device.

[0046] Figure 16 FIG. 9 is a diagram to show one example of a reconstructed image generated using a trained model generated by inputting first input image data related to the embodiment to a machine training device.

[0047] Figure 17 FIG. 10 is a diagram to show one example of a reconstructed image generated by reconstructing projection data acquired by a scan in which a normal dose and a number of projection data directions have been set.

[0048] Figure 18FIG. 1 is a diagram showing one example of a reconstructed image generated by reconstructing projection data acquired by sparse view CT using a reconstruction method using total variation.

[0049] Figure 19 FIG. 2 is a diagram showing one example of a reconstructed image generated using an existing trained model.

[0050] Figure 20 FIG. 3 is a diagram showing one example of a reconstructed image generated using a trained model generated by inputting first input image data and second input image data related to the embodiment to a machine training device.

[0051] Figure 21 FIG. 4 is a diagram showing one example of a reconstructed image generated using a trained model generated by inputting first input image data related to the embodiment to a machine training device a plurality of times and inputting second input image data to the machine training device.

[0052] BRIEF DESCRIPTION OF DRAWINGS

[0053] 1... image generation system

[0054] 10... X-ray CT device

[0055] 20... storage device

[0056] 30... trained model generation device

[0057] 300... trained model generation program

[0058] 310... training execution function

[0059] 320... trained model acquisition function

[0060] 40... image generation device

[0061] 400... image generation program

[0062] 410... image generation function

[0063] 50... machine training device DETAILED DESCRIPTION

[0064] [EMBODIMENT]

[0065] An embodiment of the present application will be described with reference to the accompanying drawings. Figure 1 FIG. 1 is a diagram showing one example of an image generation system related to the embodiment. As shown in FIG. 1, the image generation system 1 includes an X-ray CT device 10, a storage device 20, a trained model generation device 30, and an image generation device 40. Figure 1As shown, the image generation system 1 includes an X-ray CT apparatus 10, a storage apparatus 20, a trained model generation apparatus 30, an image generation apparatus 40, and a machine training apparatus 50. The X-ray CT apparatus 10, the storage apparatus 20, the trained model generation apparatus 30, the image generation apparatus 40, and the machine training apparatus 50 are connected to a network NW. The network NW is, for example, the Internet, an intranet, a WAN (Wide Area Network), or a LAN (Local Area Network).

[0066] Figure 2 A diagram showing one example of the X-ray CT apparatus according to the embodiment. As shown, the X-ray CT apparatus 10 includes a top plate 111, a top plate drive section 112, an X-ray tube 121, an X-ray high voltage section 122, an X-ray detector 123, a scan execution section 124, and a CT image generation section 130. Figure 2

[0067] The top plate 111 is a plate-shaped member on which a subject is placed. The top plate drive section 112 moves the top plate 111 relative to the X-ray tube 121 and the X-ray detector 123. The X-ray tube 121 generates X-rays that are irradiated to the subject. The X-ray high voltage section 122 applies high voltage to the X-ray tube 121. The X-ray detector 123 includes a detection element that detects X-rays irradiated by the X-ray tube 121.

[0068] The scan execution section 124 performs scanning of the subject by controlling the top plate drive section 112, the X-ray tube 121, the X-ray high voltage section 122, and the X-ray detector 123, and acquires a plurality of projection data. When scanning of the subject is performed, the X-ray tube 121 and the X-ray detector 123 rotate around the periphery of the subject in a state of facing each other. In addition, the position of the X-ray tube 121 at the time of acquisition of each of the plurality of projection data is referred to as a field of view. The CT image generation section 130 generates a CT image by reconstructing the plurality of projection data generated by scanning of the subject, and stores CT image data showing the CT image in the storage apparatus 20.

[0069] Further, the X-ray CT apparatus 10 can perform not only normal CT photography but also sparse view CT photography and low dose CT photography. In the sparse view CT photography, the subject is scanned by reducing the number of projection directions. In the low dose CT photography, the subject is scanned in a state of suppressing the current flowing through the X-ray tube.

[0070] Figure 1 The trained model generation apparatus 30 and the image generation apparatus 40 shown are, for example, computers that include a storage medium and a hardware processor.

[0071] ​Storage media include, for example, hard disk drives (HDDs), solid-state drives (SSDs), flash memory, and ROM (read-only memory). The trained model generation device 30 stores the following information on its storage media: Figure 1 The trained model generation program 300 is shown. The image generation device 40 has a storage medium storing... Figure 1 The image generation program 400 shown.

[0072] The hardware processor is, for example, a CPU (Central Processing Unit). The hardware processor in the trained model generation device 30 reads and executes the trained model generation program 300, thereby achieving... Figure 1 The training execution function 310 and the trained model acquisition function 320 are shown. Additionally, the hardware processor of the image generation apparatus 40 stores the image generation program 40 by reading and executing it. Figure 1 The image generation function 410 is shown. After the model generation program 300 and the image generation program 400 are trained, they perform the first, second, or third processing described below.

[0073] [First Processing]

[0074] The training execution function 310 inputs a set of first input image data and output image data to the machine training device 50 to perform machine training and generate a trained model. The first input image data is data showing a first input image generated using a first reconstruction method based on compressed sensing. The output image data is data showing an output image corresponding to the first input image, with less image quality degradation compared to the first input image, and representing the correct solution of the machine training performed by the machine training device 50. Specifically, the output image is generated using a third reconstruction method different from the first reconstruction method. Furthermore, the conditions for generating the output image are normal dosage and a certain number of projection directions.

[0075] In this scenario, the multiple first input images and multiple output images can be all one-to-one associated with each other, all many-to-one associated with each other, only some one-to-one associated, only some many-to-one associated, or none associated at all. Furthermore, the number of first input images and the number of output images can be the same or different.

[0076] The compressed sensing used in the first reconstruction method is, for example, total variation. The process of generating the first input image using total variation is to explore a vector z that minimizes the value represented by the following equation (2) (which includes vector p, matrix A, vector z, and the smoothing parameter β of the first reconstruction method). Vector p is a vector that arranges the values ​​representing the intensity of X-rays (represented by each pixel of the projection data acquired in each field of view) into a column. Matrix A is a matrix representing the length of each tiny region contained within the area scanned by each element of the X-rays. Vector z is a vector that arranges the values ​​representing the concentration represented by each pixel of the first input image into a column. The smoothing parameter β of the first reconstruction method is a parameter indicating the intensity of smoothing the first input image. TV is the total variational norm of the vector z.

[0077] [Mathematical Expression 2]

[0078]

[0079] The first input image is a CT image generated by first reconstructing projection data generated by performing normal CT imaging, sparse view CT imaging, or low-dose CT imaging on the X-ray CT device 10. Alternatively, the first input image is a CT image generated by first reconstructing projection data generated by another X-ray CT device.

[0080] Machine training device 50, for example, generates a trained model using a convolutional neural network (CNN). Figure 3 This diagram illustrates an example of machine training performed by the machine training apparatus according to an embodiment. Figure 3 As shown, the machine training device 50 receives a vector z associated with the first input image data. i and the vector x associated with the output image data i The input is taken and machine training is performed. The process of performing the machine training and generating the trained model is to explore how to minimize the number of samples N of the first input image used for machine training and the output image corresponding to the first input image, vector w, vector b, smoothing parameter β of the first reconstruction method, and vector x. i and vector z i The vectors w and b represent the mean square error (MSE), and the smoothing parameter β of the first reconstruction method. Vector w is a vector that arranges the weight coefficients used in the convolutional neural network into a column. Vector b is a vector that arranges the biases used in the convolutional neural network into a column. Vector x iis a vector that arranges values showing the density represented by each pixel of the output image in a row. The vector z i is a vector that arranges values showing the density represented by each pixel of the first input image in a row. Further, the smoothing parameter β of the first reconstruction method is the same as described above. In addition, "CNN" included in the formula (3) indicates a convolutional neural network.

[0081] [Math. 3]

[0082]

[0083] The trained model acquisition function 320 acquires trained model data showing the trained model. Also, the trained model acquisition function 320 stores the trained model data in the storage device 20.

[0084] The image generation function 410 inputs input image data to the trained model shown by the trained model data stored in the storage device 20 to generate a reconstructed image. The input image data is data showing an input image. The input image is an image generated by the same compressive sensing as the first input image used in the training described above.

[0085] The processing of the image generation function 410 to generate the reconstructed image is represented by the following formula (4) that includes the vector x, the vector w, the vector b, the vector z, and the smoothing parameter β of the first reconstruction method. Further, the vector w, the vector b, the vector z, and the smoothing parameter β of the first reconstruction method are the same as described above. In addition, "CNN" included in the formula (4) indicates a convolutional neural network. The vector w, the vector b, and the smoothing parameter β of the first reconstruction method are determined automatically in the training process to generate the trained model.

[0086] [Math. 4]

[0087] x = CNN(w, b)z(β) … (4)

[0088] [Second Processing]

[0089] The training execution function 310 inputs a group of the first input image data, the second input image data, and the output image data to the machine training device 50 to execute machine training, and causes the machine training device 50 to generate a trained model. The first input image data and the output image data are the same as in the case of the first processing described above.

[0090] In this scenario, the multiple first input images, multiple second input images, and multiple output images can all be associated one-to-one with each other, all can be associated many-to-one with each other, only some can be associated one-to-one with each other, only some can be associated many-to-one with each other, or they can be unassociated. Furthermore, the number of first input images, the number of second input images, and the number of output images can be the same or different from each other.

[0091] Furthermore, in this case, the multiple first input images and multiple second input images can be all one-to-one associated with each other, all many-to-one associated with each other, only some one-to-one associated, only some many-to-one associated, or not associated at all. Additionally, the number of first input images and the number of second input images can be the same or different. Furthermore, these apply to the relationship between the first input images and the output image, as well as the relationship between the second input images and the output image.

[0092] The second input image data is data showing the second input image generated by a second reconstruction method, which is an analytical reconstruction method. The second reconstruction method is, for example, a filtered back-projection (FBP) algorithm. The second input image is a CT image generated by reconstructing projection data generated by the X-ray CT apparatus 10 performing normal CT imaging, sparse-view CT imaging, or low-dose CT imaging. Alternatively, the second input image is a CT image generated by reconstructing projection data generated by another X-ray CT apparatus.

[0093] Machine training device 50, for example, generates a trained model using a convolutional neural network. Figure 4 This diagram illustrates an example of machine training performed by the machine training apparatus according to an embodiment. Figure 4 As shown, the machine training device 50 receives a vector z associated with the first input image data. i The vector y associated with the second input image data i and the vector x associated with the output image data i The input is taken and machine training is performed. The process of performing machine training and generating the trained model is to explore how to minimize the number of samples N in the group consisting of the first input image, the second input image and the output image, vector w, vector b, the smoothing parameter β of the first reconstruction method, and vector x. i Vector y i and vector z i The vectors w and b, representing the mean square error, and the smoothing parameter β of the first reconstruction method, are processed. Vector y iis a vector in which values showing the concentration represented by each pixel of the second input image are arranged in a row. Further, the vector x, the vector w, the vector b, the vector z, and the smoothing parameter β of the first reconstruction method are the same as in the case of the first processing. In addition, "CNN" included in the formula (6) indicates a convolutional neural network. i and the vector z i The same as in the case of the first processing.

[0094] [Formula 5]

[0095]

[0096] The trained model acquisition function 320 acquires trained model data showing the trained model. Also, the trained model acquisition function 320 stores the trained model data in the storage device 20.

[0097] The image generation function 410 inputs input image data to the trained model shown by the trained model data stored in the storage device 20 to generate a reconstructed image. The input image data is data showing an input image. The input image is, for example, an image generated by the same compressive sensing as the first input image described above or the same analytical reconstruction method as the second input image described above.

[0098] The processing of the image generation function 410 to generate a reconstructed image is represented by the following formula (6) (which includes the vector x, the vector w, the vector b, the vector y, the vector z, and the smoothing parameter β of the first reconstruction method). The vector y is a vector in which values showing the concentration represented by each pixel of the second input image are arranged in a row. Further, the vector x, the vector w, the vector b, the vector z, and the smoothing parameter β of the first reconstruction method are the same as in the case of the first processing. In addition, "CNN" included in the formula (6) indicates a convolutional neural network.

[0099] [Formula 6]

[0100]

[0101] [Third Processing]

[0102] The training execution function 310 inputs at least two kinds of first input image data and output image data generated by the first reconstruction method in which different smoothing parameters have been set to the machine training device 50 to execute machine training, and causes the machine training device 50 to generate a trained model. The first kind of first input image data is first input image data showing a first input image generated by the first reconstruction method using compressive sensing in which a first smoothing parameter has been set. The second kind of first input image data is first input image data showing a first input image generated by the first reconstruction method using compressive sensing in which a second smoothing parameter different from the value of the first smoothing parameter has been set.

[0103] Similarly, in addition to the first type of first input image data and the second type of first input image data, the training execution function 310 can also input a group of first input image data containing the third type up to the Mth type (M: a natural number three or more) of first input image data to the machine training device 50 to perform machine training and cause the machine training device 50 to generate a trained model. In this case, the training execution function 310 inputs a group containing three or more types of first input image data to the machine training device 50 to perform machine training and causes the machine training device 50 to generate a trained model. The Mth type of first input image data is the first input image data showing the first input image generated by using a first reconstruction method that uses compressed sensing (which has a Mth smoothing parameter whose value is different from the other smoothing parameters).

[0104] In other words, the training execution function 310 inputs a group of multiple first input images and output images generated by a first reconstruction method of compressed sensing with different smoothing parameters to the machine training device 50 to perform machine training and enable the machine training device 50 to generate a trained model.

[0105] In this scenario, the first input image groups and multiple output images contained within the multiple first input images can be all one-to-one associated with each other, all many-to-one associated with each other, only some of them one-to-one associated, only some of them many-to-one associated, or they can be unassociated. Furthermore, the number of first input image groups and the number of output images can be the same or different.

[0106] Machine training device 50, for example, generates a trained model using a convolutional neural network. Figure 5 This diagram illustrates an example of machine training performed by the machine training apparatus according to an embodiment. Figure 5 As shown, the machine training device 50 receives vectors z that are respectively associated with multiple first input image data. i (1) Vector z i (2), ... and vector z i (M) and the vector x associated with the output image data i The input is used to perform machine training. The process of performing this machine training and generating the trained model is to explore how to minimize the number of samples N, vector w, vector b, and vector x of the first input image and the corresponding output image used during training. i Vector z i (1) Vector z i (2) ... and vector z i(M) (M: a natural number of 2 or more). The vectors z i (1), the vector z i (2),..., and the vector z i (M) are vectors in which values showing the concentration represented by each pixel of the first input image are arranged in a column. Further, the vector x, the vector w, and the vector b are the same as in the case of the first processing. i The same as in the case of the first processing.

[0107] [mathematical expression 7]

[0108]

[0109] The trained model acquisition function 320 acquires trained model data showing the trained model. Also, the trained model acquisition function 320 stores the trained model data in the storage device 20.

[0110] The image generation function 410 inputs the input image data to the trained model shown by the trained model data stored in the storage device 20 to generate a reconstructed image. The input image data is a vector z i (1), the vector z i (2),..., and the vector z i (M'). The input image is an image generated by compressive sensing with the same method as the plurality of first input images described above, that is, with the smoothing parameter set to be different from each other. Further, M' can be the same as M described above or different from M described above.

[0111] The processing of the image generation function 410 to generate the reconstructed image is represented by the following expression (8) (which includes the vector x, the vector w, the vector b, the vector z(1), the vector z(2),..., and the vector z(M) ). The vector z(1), the vector z(2),..., and the vector z(M) are vectors in which values showing the concentration represented by each pixel of the first input image are arranged in a column. Further, the vector x, the vector w, and the vector b are the same as in the case of the first processing. In addition, "CNN" included in expression (8) means a convolutional neural network.

[0112] [mathematical expression 8]

[0113]

[0114] Next, one example of the processing performed by the trained model generation program 300 according to the embodiment will be described with reference to Figure 6 One example of the processing performed by the trained model generation program 300 according to the embodiment will be described with reference to Figure 6 One example of the processing performed by the trained model generation program 300 according to the embodiment will be described with reference to

[0115] In step S10, the training execution function 310 inputs at least the first input image data to the machine training device 50 to execute machine training, and causes the machine training device 50 to generate a trained model.

[0116] In step S20, the trained model acquisition function 320 acquires trained model data showing the trained model generated in step S10.

[0117] The trained model generation program 300 and the image generation program 400 related to the embodiment are described above.

[0118] The trained model generation program 300 inputs the first input image data showing the first input image generated by using the first reconstruction method of compressed sensing to the machine training device 50 to execute machine training, and causes the machine training device 50 to generate a trained model. Also, the trained model generation program 300 acquires trained model data showing the trained model.

[0119] Thus, the trained model generation program 300 can generate a trained model that can generate a reconstructed image in which the stair-step artifact occurring on the first input image due to compressed sensing, disappearance of smooth concentration change, disappearance of texture, and the like are mitigated. Also, even in a case where the optimal value of the smoothing parameter β of the first reconstruction method is unknown, the trained model generation program 300 can obtain such an effect only by inputting the first input image data to the machine training device 50.

[0120] In addition to the first input image data, the trained model generation program 300 inputs second input image data showing a second input image generated by a second reconstruction method to the machine training device 50 to execute machine training, and causes the machine training device 50 to generate a trained model. The second reconstruction method is a reconstruction method different from the first reconstruction method, and is an analytical reconstruction method. Also, the analytical reconstruction method has an advantage that a reconstructed image in which a smooth concentration change is reproduced with relatively high accuracy can be generated.

[0121] Thus, the trained model generation program 300 can generate a trained model that can compensate for the stair-step artifact occurring on the first input image due to compressed sensing, disappearance of smooth concentration change, disappearance of texture, and the like with the advantage. Also, even in a case where the optimal value of the smoothing parameter β of the first reconstruction method is unknown, the trained model generation program 300 can obtain such an effect only by inputting the first input image data and the second input image data to the machine training device 50.

[0122] The trained model generation program 300 generates a trained model by exploring the smoothing parameter β of the first reconstruction method that minimizes the mean square error represented by Equation (5) above, the weight coefficient w used in the convolutional neural network used in the trained model, and the bias b used in the convolutional neural network.

[0123] Therefore, since the trained model generation program 300 generates the trained model by exploring the optimal value of the smoothing parameter β of the first reconstruction method during the training process, it saves the personnel who want the image generation program 400 to generate reconstructed images from the workload of trial and error in determining the specific value of the smoothing parameter β of the first reconstruction method. Furthermore, this effect is particularly effective because it is often difficult to determine the optimal value of the smoothing parameter β of the first reconstruction method based on experience.

[0124] The trained model generation program 300 inputs at least two first input image data generated by a first reconstruction method with different smoothing parameters to the machine training device 50 to perform machine training and generate a trained model.

[0125] Therefore, the trained model generation program 300 can generate a trained model by combining features from multiple reconstructed images with varying smoothing intensities. This trained model generates a reconstructed image that mitigates trapezoidal artifacts, loss of smooth density variations, and texture loss that occur in the first input image due to compressed sensing. Furthermore, even when the optimal value of the smoothing parameter β in the first reconstruction method is unknown, the trained model generation program 300 can achieve this effect simply by inputting at least two types of first input image data into the machine training device 50.

[0126] In addition to the first input image data, the trained model generation program 300 inputs the output image data, which shows the output image generated under normal dose and number of projection directions, to the machine training device 50 to perform machine training and cause the machine training device 50 to generate the trained model.

[0127] Thus, the trained model generation program 300 enables the machine training device 50 to perform teacher-led training and enables the machine learning device 50 to generate a trained model with higher accuracy.

[0128] Image generation program 400 inputs input image data, which shows the input image, into the trained model generated by trained model generation program 300 to generate a reconstructed image.

[0129] Thus, the image generation program 400 can mitigate the image quality degradation that occurs in the reconstructed image generated by the reconstruction method using compressed sensing.

[0130] Next, referring to Figures 7 to 11 A first specific example of the above effect will be described. Figure 7 A graph showing one example of a reconstructed image generated by reconstructing projection data acquired by a scan with a normal dose and a number of projection data directions set. The closer the reconstructed image is to Figure 7 the reconstructed image P7, the less the image quality degradation due to compressed sensing.

[0131] Figure 8 A graph showing one example of a reconstructed image generated by reconstructing projection data acquired by sparse view CT using a reconstruction method using total variation. Figure 8 The reconstructed image P8 shown was generated from 64-direction projection data acquired by sparse view CT photography. In addition, as Figure 8 As shown in the lower left of FIG. 8, the mean square error of the reconstructed image P8 was 82.19, and the image evaluation index (SSIM: Structural similarity) was 0.89.

[0132] Here, the mean square error is calculated by the following equation (9) (which includes a function g(i, j) representing the reconstructed image P7 of a normal dose and a number of projection directions, a function f(i, j) representing the reconstructed image that is the target, and the total number of pixels L of each of the two reconstructed images). The indices i and j are indices showing the positions of the pixels of each of the two reconstructed images.

[0133] [Mathematical Formula 9]

[0134]

[0135] In addition, the image evaluation index is calculated by the following equation (10), and the greater the value, the closer the reconstructed image is to the reconstructed image P7. Further, details of equation (10) are described in the submitted paper "Z. Wang, A. C. Bovik, H. R. Sheikh and E. P. Simoncelli, "Image quality assessment: From error visibility to structural similarity," IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, Apr. 2004."

[0136] [Mathematical Formula 10]

[0137]

[0138] In Figure 8The low-contrast structures such as blood vessels in the illustrated reconstructed image P8 are not depicted. Figure 7 The smooth density change and the fine texture depicted in the illustrated reconstructed image P7 are also depicted in the reconstructed image P9. However, the part of the low-contrast structures depicted in the reconstructed image P7 in white cannot be depicted in the reconstructed image P9.

[0139] Figure 9 A graph showing one example of the reconstructed image generated using the trained model. The trained model mentioned here is a trained model generated by inputting only the reconstructed image generated by the filtered back projection method and the output image generated under the condition of the normal dose and the number of projection directions to the machine training device using the convolutional neural network to train the machine training device. Figure 9 The illustrated reconstructed image P9 is generated from the projection data of 64 directions acquired by sparse view CT photography. In addition, as Figure 9 As shown in the lower left of FIG. 7, the mean square error of the reconstructed image P9 is 79.06, and the image evaluation index is 0.88.

[0140] Figure 9 The illustrated reconstructed image P9 reproduces the low-contrast structures in the illustrated reconstructed image P7 in the whole. Figure 7 The smooth density change and the fine texture depicted in the illustrated reconstructed image P7 are also depicted in the reconstructed image P9. However, the part of the low-contrast structures depicted in the reconstructed image P7 in white cannot be depicted in the reconstructed image P9.

[0141] Figure 10 A graph showing one example of the reconstructed image generated using the trained model, the trained model being generated by inputting the first input image data and the second input image data related to the embodiments to the machine training device. Figure 10 The illustrated reconstructed image P10 is generated from the projection data of 64 directions acquired by sparse view CT photography. In addition, as Figure 10 As shown in the lower left of FIG. 7, the mean square error of the reconstructed image P9 is 79.06, and the image evaluation index is 0.88.

[0142] Figure 10 The illustrated reconstructed image P10 reproduces the low-contrast structures in the illustrated reconstructed image P7 in the whole. Figure 7 The smooth density change and the fine texture depicted in the illustrated reconstructed image P7 are also depicted in the reconstructed image P10. In addition, the reconstructed image P10 is Figure 7 The low-contrast structures in the illustrated reconstructed image P7 are also depicted in the reconstructed image P10 in white. The reconstructed image P10 is Figures 8 to 11 the reconstructed image P10 is the closest to the Figure 7 the reconstructed image P10 is the closest to the

[0143] Figure 11The diagram illustrates an example of a reconstructed image generated using a trained model, which is generated by inputting first input image data, as described in the implementation, into a machine training device. Figure 11 The reconstructed image P11 shown is generated based on 64-direction projection data acquired through sparse-view CT photography. Additionally, as... Figure 11 As shown in the lower left corner, the mean square error of the reconstructed image P11 is 37.15, and the image evaluation index is 0.91.

[0144] Figure 11 The reconstructed image P11 shown reproduces more than a certain amount of data overall. Figure 7 The reconstructed image P7 depicts smooth density variations and subtle textures. Additionally, the reconstructed image P11... Figure 7 The reconstructed image P7 shown similarly depicts low-contrast structures in white. However, because the reconstructed image P11 does not incorporate the advantages of this analytical reconstruction method, which can generate reconstructed images that reproduce smooth concentration variations with high accuracy, some aspects, such as... Figure 8 The reconstructed image P8, like the one shown, has been smoothed. Compared to the reconstructed image P10, the smooth density variations and subtle textures depicted in the reconstructed image P7 disappear across the entire image. The reconstructed image P11 is... Figures 8 to 11 The second closest in the reconstructed image shown Figure 7 The reconstructed image shown is a reconstructed image of the image.

[0145] Next, refer to Figures 12 to 16 A second specific example illustrating the above effect will be provided. Figure 12 This diagram illustrates an example of a reconstructed image generated by reconstructing projection data, acquired through scans with a pre-set normal dose and number of projection data directions. The closer the reconstructed image is to the desired image... Figure 12 The less image quality degradation is shown in the reconstructed image P12, the better.

[0146] Figure 13 The figure illustrates an example of a reconstructed image generated by reconstructing projection data using a total variational reconstruction method, the projection data being acquired via a sparse view CT. Figure 13 The reconstructed image P13 shown is generated based on 64-direction projection data acquired through sparse-view CT photography. Additionally, as... Figure 13 As shown in the lower left corner, the mean square error of the reconstructed image P13 is 98.56, and the image evaluation index is 0.86.

[0147] exist Figure 13 In the reconstructed image P13 shown, structures with low contrast, such as blood vessels, are contrasted with... Figure 12The reconstructed image P12 shown is also depicted in light gray. However, since the reconstructed image P13 is smoothed by total variation, the smooth concentration change and fine texture depicted in the reconstructed image P12 disappear over the entire reconstructed image P13.

[0148] Figure 14 A graph showing one example of a reconstructed image generated using a trained model. Figure 14 The reconstructed image P14 shown is generated from 64-direction projection data acquired by sparse-view CT photography. In addition, as Figure 14 As shown in the lower left of FIG. 8, the mean square error of the reconstructed image P14 is 76.25, and the image evaluation index is 0.84.

[0149] Figure 14 The reconstructed image P14 shown reproduces the structure of the contrast object in the reconstructed image P12 in light gray as a whole with high accuracy. Figure 12 The smooth concentration change and fine texture depicted in the reconstructed image P12. However, the reconstructed image P14 cannot depict the portion of the contrast object in the reconstructed image P12 that is depicted in light gray with low contrast.

[0150] Figure 15 A graph showing one example of a reconstructed image generated using a trained model, which is generated by inputting the first input image data and the second input image data related to the embodiments to a machine training device. Figure 15 The reconstructed image P15 shown is generated from 64-direction projection data acquired by sparse-view CT photography. In addition, as Figure 15 As shown in the lower left of FIG. 9, the mean square error of the reconstructed image P15 is 43.55, and the image evaluation index is 0.88.

[0151] Figure 15 The reconstructed image P15 shown reproduces the structure of the contrast object in the reconstructed image P12 in light gray as a whole with high accuracy. Figure 12 The smooth concentration change and fine texture depicted in the reconstructed image P12. In addition, in the reconstructed image P15, the portion of the contrast object that is depicted in light gray with low contrast in the reconstructed image P12 is depicted in light gray as well. Figure 12 The reconstructed image P12 shown is also depicted in light gray. The reconstructed image P15 is Figures 13 to 16 The reconstructed image P15 shown is the closest to Figure 12 The reconstructed image P16 shown is the reconstructed image closest to

[0152] Figure 16 A graph showing one example of a reconstructed image generated using a trained model, which is generated by inputting the first input image data related to the embodiments to a machine training device. Figure 16 The reconstructed image P16 shown is generated from 64-direction projection data acquired by sparse-view CT photography. In addition, asFigure 16 As shown in the lower left corner, the mean square error of the reconstructed image P16 is 59.01, and the image evaluation index is 0.87.

[0153] Figure 16 The reconstructed image P16 shown reproduces more than a certain amount of data overall. Figure 12 The reconstructed image P12 depicts smooth density variations and subtle textures. Additionally, in the reconstructed image P16, compared to... Figure 12 The reconstructed image P12 shown similarly depicts low-contrast structures in light gray. However, because the reconstructed image P16 does not incorporate the advantages of this analytical reconstruction method, which can generate reconstructed images that reproduce smooth concentration variations with high accuracy, some parts, such as... Figure 13 The reconstructed image P13, as shown, was also smoothed. Compared to reconstructed image P15, the smooth density variations and subtle textures depicted in reconstructed image P12 disappeared throughout reconstructed image P16. Reconstructed image P16 is... Figures 13 to 16 The second closest in the reconstructed image shown Figure 12 The reconstructed image shown is a reconstructed image of the image.

[0154] Next, refer to Figures 17 to 21 A third specific example illustrating the above effect will be provided. Figure 17 This diagram illustrates an example of a reconstructed image generated by reconstructing projection data, acquired through scans with a pre-set normal dose and number of projection data directions. The closer the reconstructed image is to the desired image... Figure 17 The less image quality degradation is shown in the reconstructed image P17, the better.

[0155] Figure 18 The figure illustrates an example of a reconstructed image generated by reconstructing projection data using a total variational reconstruction method, wherein the projection data is acquired via a sparse view CT. Figure 18 The reconstructed image P18 shown is generated based on 64-direction projection data acquired through sparse-view CT photography. Additionally, as... Figure 18 As shown in the lower left corner, the image evaluation index of reconstructed image P18 is 0.586. Furthermore, this image evaluation index is the average of the image evaluation indices of reconstructed images primarily depicting the liver, generated using a trained model produced by machine training device 50 for 100 cases. Additionally, Figures 19 to 21 The image rating index shown in the lower left corner is the same.

[0156] exist Figure 18 In the reconstructed image P18 shown, structures with low contrast, such as blood vessels, are contrasted with... Figure 17The illustrated reconstructed image P12 is also depicted in light gray. However, since the reconstructed image P18 is smoothed by total variation, the smooth concentration change and fine texture depicted in the reconstructed image P17 disappear over the entire reconstructed image P18.

[0157] Figure 19 A graph showing one example of a reconstructed image generated using a trained model. Figure 19 The illustrated reconstructed image P19 is generated from 64-directional projection data acquired by sparse-view CT photography. In addition, as Figure 19 As shown in the lower left of FIG. 17, the image evaluation index of the reconstructed image P19 is 0.823.

[0158] Figure 19 The illustrated reconstructed image P19 reproduces the structure of the contrast material in the reconstructed image P17 in high accuracy. Figure 17 The smooth concentration change and fine texture depicted in the illustrated reconstructed image P17. In addition, the reconstructed image P20 depicts the contrast material in light gray, like the reconstructed image P17. The reconstructed image P20 is

[0159] Figure 20 A graph showing one example of a reconstructed image generated using a trained model. Figure 20 The illustrated reconstructed image P20 is generated from 64-directional projection data acquired by sparse-view CT photography. In addition, as Figure 20 As shown in the lower left of FIG. 18, the image evaluation index of the reconstructed image P20 is 0.902.

[0160] Figure 20 The illustrated reconstructed image P20 reproduces the structure of the contrast material in high accuracy. Figure 17 The smooth concentration change and fine texture depicted in the illustrated reconstructed image P17. In addition, the reconstructed image P20 depicts the contrast material in light gray, like the reconstructed image P17. The reconstructed image P20 is Figure 17 The illustrated reconstructed image P17 depicts the contrast material in light gray. The reconstructed image P20 is Figures 18 to 21 the reconstructed image shown in FIG. 16. Figure 17 The reconstructed image of the illustrated reconstructed image P17.

[0161] Figure 21 A graph showing one example of a reconstructed image generated using a trained model. Figure 21 The illustrated reconstructed image P21 is generated from 64-directional projection data acquired by sparse-view CT photography. In addition, asFigure 21 The image evaluation index of the reconstructed image P21 is 0.910 as shown in the lower left.

[0162] Figure 21 The reconstructed image P21 as shown in the figure as a whole reproduces the smooth concentration change and the fine texture depicted in the reconstructed image P17 with high accuracy. Figure 17 The reconstructed image P21 as shown in the figure as a whole reproduces the smooth concentration change and the fine texture depicted in the reconstructed image P17 with high accuracy. Figure 17 The reconstructed image P21 as shown in the figure as a whole reproduces the smooth concentration change and the fine texture depicted in the reconstructed image P17 with high accuracy. Figures 18 to 21 The reconstructed image P21 is the closest to the reconstructed image P17 as shown in the figure among the reconstructed images. Figure 17 The reconstructed image P21 is the closest to the reconstructed image P17 as shown in the figure among the reconstructed images.

[0163] The reconstructed image P21 is the closest to the reconstructed image P17 as shown in the figure among the reconstructed images. Figure 17 The reconstructed image P21 is the closest to the reconstructed image P17 as shown in the figure among the reconstructed images. The reason for this is that the optimal value of the smoothing parameter β differs depending on the position of the reconstructed image, and the first input image data having the optimal smoothing parameter β is selected for each position of the reconstructed image generated by the trained model, and the reconstructed image is generated by combining these smoothing parameters β. That is, unlike the reconstructed image P20, the reconstructed image P21 is generated using the trained model generated by inputting a plurality of first input image data having mutually different smoothing parameters β set thereto, and thus, the reconstructed image P21 becomes a reconstructed image closer to the reconstructed image P17 as shown in the figure than the reconstructed image P20. Figure 17 The reconstructed image P21 is the closest to the reconstructed image P17 as shown in the figure among the reconstructed images.

[0164] Further, in the above-described embodiment, a case in which the trained model generation device 30 and the image generation device 40 are separate is exemplified, but the present application is not limited thereto. That is, the trained model generation device 30 and the image generation device 40 can be formed as an integrated device.

[0165] Further, in the above-described embodiment, a case in which the machine training device 50 generates a trained model using a convolutional neural network is exemplified, but the present application is not limited thereto. The machine training device 50 can also generate, for example, a trained model using a recurrent neural network (RNN: Recurrent Neural Network).

[0166] Further, in the above-described embodiment, a case in which the trained model generation program 300 does not input the second input image data to the machine training device 50 in the third process is exemplified, but the present application is not limited thereto. The trained model generation program 300 can input the second input image data to the machine training device 50 together with at least two types of first input image data and output image data in the third process.

[0167] Further, in the above-described embodiment, a case where the output image data is input to the machine training device 50 to cause the machine training device 50 to perform the teacher-present training in a case where the training-finished model generation program 300 executes any one of the first process, the second process, and the third process has been described as an example, but the present application is not limited to this. That is, in a case where the training-finished model generation program 300 executes the first process, the second process, or the third process, the machine training device 50 can also be caused to perform the teacher-absent training without inputting the output image data to the machine training device 50.

[0168] Further, in the above-described embodiment, a case where each function illustrated in FIG. 8 is implemented by a hardware processor that reads and executes the training-finished model generation program 300 has been described as an example, but the present application is not limited to this. Figure 1 At least a part of the functions illustrated in FIG. 8 can be implemented by a hardware processor including a circuit (circuitry) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), and a GPU (Graphics Processing Unit), or the like. Alternatively, at least a part of the functions illustrated in FIG. 8 can be implemented by a combination of software and hardware. Figure 1 At least a part of the functions illustrated in FIG. 8 can be implemented by a hardware processor including a circuit (circuitry) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), and a GPU (Graphics Processing Unit), or the like. Alternatively, at least a part of the functions illustrated in FIG. 8 can be implemented by a combination of software and hardware. Figure 1 At least a part of the functions illustrated in FIG. 8 can be implemented by a combination of software and hardware. Further, these hardware can be integrated into one or divided into a plurality of hardware.

[0169] Further, in the above-described embodiment, a case where the training-finished model generation program 300 and the image generation program 400 are applied to a reconstructed image generated by an X-ray CT device has been described as an example, but the present application is not limited to this. The training-finished model generation program 300 and the image generation program 400 can also be applied to a reconstructed image generated by, for example, a Positron Emission Tomography (PET) device, a Single Photon Emission Computed Tomography (SPECT) device, or a Magnetic Resonance Imaging (MRI) device.

[0170] The embodiments of the present application have been described in detail with reference to the accompanying drawings. However, the specific configuration of the embodiments of the present application is not limited to the above-described embodiments, and at least one of various combinations, modifications, substitutions, and design changes can be added to the above-described embodiments without departing from the spirit of the present application.

Claims

1. A storage medium storing a trained model generation program, characterized in that, Enable the computer to perform the following functions: The training execution function performs machine training by inputting first input image data and second input image data into a machine training device, and causes the machine training device to generate a trained model. The first input image data shows a first input image generated by a first reconstruction method, and the second input image data shows a second input image generated by a second reconstruction method. The first reconstruction method uses compressed sensing, and the second reconstruction method is a different reconstruction method from the first reconstruction method, which is an analytical reconstruction method. The function to obtain the trained model retrieves the trained model data, which shows the trained model. The training execution function enables the machine training device to generate the trained model by exploring the smoothing parameters of the first reconstruction method that minimize the mean square error represented by equation (1), the weight coefficients used in the convolutional neural network used as the trained model, and the bias used in the convolutional neural network. [Mathematical Expression 1] in: w: A vector that arranges the weights used in the convolutional neural network into a single column. b: Arrange the biases used in the convolutional neural network into a vector. β: Smoothing parameter of the first reconstruction method x i This will be a vector showing the concentration values ​​represented by each pixel in the output image, arranged in a column. y i This will display a vector showing the concentration values ​​represented by each pixel of the second input image, arranged in a column. z i : A vector that shows the values ​​of the concentration represented by each pixel of the first input image arranged in a column.

2. The storage medium storing the trained model generation program according to claim 1, characterized in that, The training execution function will at least input the first input image data, which shows the first input image generated by the first reconstruction method of compressed sensing with a first smoothing parameter set, and the first input image data, which shows the first input image generated by the first reconstruction method of compressed sensing with a second smoothing parameter set to a value different from the first smoothing parameter, to the machine training device to perform machine training and cause the machine training device to generate the trained model.

3. The storage medium storing the trained model generation program according to any one of claims 1 to 2, characterized in that, In addition to the first input image data, the training execution function also inputs output image data to the machine training device to perform machine training and causes the machine training device to generate the trained model. The output image data shows the output image generated by a third reconstruction method different from the first reconstruction method.

4. A storage medium storing an image generation program, characterized in that, The computer is made capable of image generation, which generates a reconstructed image by inputting input image data showing an input image to the trained model generated by a storage medium storing a trained model generation program according to any one of claims 1 to 3.

5. A training-completed model generation device, characterized in that, have: The training execution unit performs machine training by inputting first input image data and second input image data into a machine training device, and causes the machine training device to generate a trained model. The first input image data shows a first input image generated by a first reconstruction method, and the second input image data shows a second input image generated by a second reconstruction method. The first reconstruction method uses compressed sensing, and the second reconstruction method is a different reconstruction method from the first reconstruction method, which is an analytical reconstruction method. The training completed model acquisition unit acquires the training completed model data that shows the trained completed model; The training execution unit generates the trained model by exploring the smoothing parameters of the first reconstruction method that minimize the mean square error represented by equation (1), the weight coefficients used in the convolutional neural network used as the trained model, and the bias used in the convolutional neural network. [Mathematical Expression 1] in: w: A vector that arranges the weights used in the convolutional neural network into a single column. b: Arrange the biases used in the convolutional neural network into a vector. β: Smoothing parameter of the first reconstruction method x i This will be a vector showing the concentration values ​​represented by each pixel in the output image, arranged in a column. y i This will display a vector showing the concentration values ​​represented by each pixel of the second input image, arranged in a column. z i : A vector that shows the values ​​of the concentration represented by each pixel of the first input image arranged in a column.

6. An image generation apparatus, characterized in that, The device includes an image generation unit that generates a reconstructed image by inputting input image data showing an input image to the trained model generated by the trained model generation apparatus according to claim 5.

7. A method for generating a trained model, characterized in that, Includes the following steps: The training execution step involves inputting first input image data and second input image data into a machine training device to perform machine training, and causing the machine training device to generate a trained model. The first input image data shows a first input image generated by a first reconstruction method, and the second input image data shows a second input image generated by a second reconstruction method. The first reconstruction method uses compressed sensing, and the second reconstruction method is a different reconstruction method from the first reconstruction method, namely, an analytical reconstruction method. The step of obtaining the trained model involves acquiring the trained model data that shows the trained model. The training execution step enables the machine training device to generate the trained model by exploring the smoothing parameters of the first reconstruction method that minimize the mean squared error represented by equation (1), the weight coefficients used in the convolutional neural network used as the trained model, and the bias used in the convolutional neural network. [Mathematical Expression 1] in: w: A vector that arranges the weights used in the convolutional neural network into a single column. b: Arrange the biases used in the convolutional neural network into a vector. β: Smoothing parameter of the first reconstruction method x i This will be a vector showing the concentration values ​​represented by each pixel in the output image, arranged in a column. y i This will display a vector showing the concentration values ​​represented by each pixel of the second input image, arranged in a column. z i : A vector that shows the values ​​of the concentration represented by each pixel of the first input image arranged in a column.

8. An image generation method, characterized in that, The method includes an image generation step, which generates a reconstructed image by inputting input image data showing the input image into the trained model generated by the trained model generation method according to claim 7.

Images