Method for generating absorption coefficient image, nuclear medicine diagnosis apparatus, and method for making learned model

Abstract

The absorption coefficient image generation method comprises the steps of: generating an input image (6); generating an intermediate image (7) comprising an image related to a tissue region on the basis of the input image (6); and generating an absorption coefficient image (9) on the basis of the intermediate image (7) and a known absorption coefficient of the tissue region.

Description

Technical Field

[0001] This invention relates to a method for generating absorption coefficient images, a nuclear medicine diagnostic device, and a method for creating a learned model. Background Technology

[0002] Previously, a method for generating absorption coefficient images for nuclear medicine diagnostic devices was known. For example, such a method is disclosed in U.S. Patent Application Publication No. 2019 / 0130569 (hereinafter referred to as "Patent Document 1").

[0003] Patent Document 1 disclosed a method for generating absorption coefficient images for a positron emission tomography (PET) apparatus (nuclear medicine diagnostic apparatus). In this method, a pre-learned machine learning model is used to generate the absorption coefficient image. Specifically, a PET image generated from PET (Positron Emission Tomography) data is input into the machine learning model. Furthermore, the absorption coefficient image is output from the machine learning model. Thus, without performing CT (Computed Tomography) or MR (Magnetic Resonance) imaging on the subject, the absorption coefficient image is generated from PET data (measurement data) using a machine learning model.

[0004] Existing technical documents

[0005] Patent documents

[0006] Patent Document 1: U.S. Patent Application Publication No. 2019 / 0130569 Summary of the Invention

[0007] The problem the invention aims to solve

[0008] The method described in Patent Document 1 above allows for the generation of absorption coefficient images from PET data (measurement data) using a machine learning model without subjecting the subject to CT or MR imaging. However, when the machine learning model outputs an absorption coefficient image, it does not consider whether the absorption coefficient of the image is within an appropriate range (a generally acceptable value). Therefore, there is a possibility that the absorption coefficient of the image may be outside the appropriate range (a generally unacceptable value). Thus, the following problem exists: it is difficult to generate an absorption coefficient image from PET data (measurement data) using a machine learning model without subjecting the subject to CT or MR imaging, while simultaneously eliminating the possibility that the absorption coefficient of the image may be outside the appropriate range.

[0009] The present invention was made to solve the problems described above. One of the objectives of the present invention is to provide an absorption coefficient image generation method and nuclear medicine diagnostic device that can ensure that the absorption coefficient of the absorption coefficient image is within an appropriate range (typically acceptable value) even when the subject is not subjected to CT or MR imaging, but the absorption coefficient image is generated based on measurement data.

[0010] Solution for solving the problem

[0011] To achieve the above objectives, the first aspect of the present invention provides an absorption coefficient image generation method for a nuclear medicine diagnostic apparatus for generating absorption coefficient images within a subject. The absorption coefficient image generation method includes the following steps: generating an input image by image processing of measurement data obtained from the detection of radiation emitted from the subject; generating an intermediate image based on the input image, including an image related to a tissue region; and generating an absorption coefficient image based on the intermediate image and the known absorption coefficients of the tissue region. Here, "tissue" refers, for example, to the brain, bone, skin, muscle, internal organs, and body cavities.

[0012] Furthermore, the nuclear medicine diagnostic apparatus of the second aspect of the present invention includes: a detection unit that detects radiation generated from a radiopharmaceutical within a subject; and a processing unit that generates a radiation distribution image within the subject based on the detection of radiation by the detection unit, wherein the processing unit is configured to: generate an input image by performing image processing on measurement data obtained based on the detection of radiation emitted from the subject; generate an intermediate image including an image related to a tissue region based on the input image; and generate an absorption coefficient image for generating the radiation distribution image based on the intermediate image and a known absorption coefficient of the tissue region.

[0013] Furthermore, the third aspect of the present invention provides a method for creating a learned-complete model for a nuclear medicine diagnostic device. This method includes the following steps: preparing a tissue label image representing the tissue to which each pixel belongs; creating a pseudo-radioactivity distribution image and a pseudo-absorption coefficient image based on the tissue label image; generating pseudo-measurement data through simulation calculations based on the pseudo-radioactivity distribution image and the pseudo-absorption coefficient image; generating a pseudo-image by image processing the pseudo-measurement data; and using the pseudo-image as learning data to create a learned-complete model.

[0014] The effects of the invention

[0015] In the absorption coefficient image generation method of the first aspect and the nuclear medicine diagnostic apparatus of the second aspect of the present invention, as described above, an input image is generated by image processing of measurement data obtained based on the detection of radiation emitted from a subject; an intermediate image including an image related to a tissue region is generated based on the input image; and an absorption coefficient image is generated based on the intermediate image and the known absorption coefficient of the tissue region. Thus, an absorption coefficient image can be generated based on an intermediate image including an image related to a tissue region. As a result, even when the absorption coefficient image is generated based on measurement data without performing CT or MR imaging on the subject, it is possible to ensure that the absorption coefficient of the absorption coefficient image is within an appropriate range (a generally acceptable value).

[0016] Furthermore, in the method for creating a learned model in the third aspect of the present invention, as described above, the method includes the following steps: preparing a tissue label image representing the tissue to which each pixel belongs; creating a pseudo-radioactivity distribution image and a pseudo-absorption coefficient image based on the tissue label image; creating pseudo-measurement data by performing simulation calculations based on the pseudo-radioactivity distribution image and the pseudo-absorption coefficient image; generating a pseudo-image by performing image processing on the pseudo-measurement data; and creating a learned model using the pseudo-image as learning data. Thus, a learned model can be created using pseudo-images obtained through simulation calculations as learning data. As a result, unlike the case where actual images (clinical images) are used as learning data to create a learned model, it is not necessary to collect a large number of clinical images. Therefore, a learned model can be created without the work involved in collecting a large number of clinical images, which is not easy from the perspective of personal information protection. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the structure of a PET device according to one embodiment.

[0018] Figure 2 This is a schematic perspective view showing the structure of a radiation (gamma ray) detector according to one embodiment.

[0019] Figure 3 This is a flowchart illustrating a radioactive distribution image generation process in one embodiment.

[0020] Figure 4 This is a diagram generated to illustrate the radioactivity distribution image of one embodiment.

[0021] Figure 5 This is a diagram used to illustrate an implementation method for generating an intermediate image based on an input image.

[0022] Figure 6This is a diagram used to illustrate how an absorption coefficient image is generated based on an intermediate image in one implementation method.

[0023] Figure 7 This is a diagram used to illustrate the learning process of a machine learning model in one implementation.

[0024] Figure 8 This is a flowchart illustrating a learning-complete model creation method for one implementation.

[0025] Figure 9 This is a diagram illustrating the detailed learning process of a machine learning model in one implementation.

[0026] Figure 10 This is a diagram used to illustrate the generation of an intermediate image from an input image in a first variation of an implementation.

[0027] Figure 11 This is a diagram illustrating a second variation of a machine learning model used to explain one implementation method.

[0028] Figure 12 This is a diagram used to illustrate a second variation of an implementation method in which an absorption coefficient image is generated based on an intermediate image.

[0029] Figure 13 This is a diagram used to illustrate the generation of an absorption coefficient image based on an intermediate image in a first modification of a second modification of an embodiment.

[0030] Figure 14 This is a diagram illustrating the generation of an absorption coefficient image based on an intermediate image in a second variation of an embodiment.

[0031] Figure 15 This is a diagram used to illustrate the generation of an intermediate image from an input image in a third variation of an implementation.

[0032] Figure 16 This is a diagram illustrating how an intermediate image is generated from an input image in a first variation of a third variation of an implementation.

[0033] Figure 17 This is a diagram illustrating the generation of an intermediate image from an input image in a second variation of a third variation of an implementation.

[0034] Figure 18 This is a diagram illustrating how an intermediate image is generated from an input image in a third variation of an embodiment.

[0035] Figure 19 This is a diagram illustrating the generation of an intermediate image from an input image in a fourth variation of a third variation of an implementation.

[0036] Figure 20 This is a diagram used to illustrate a fourth variation of an implementation method, in which an intermediate image and a reconstructed image with absorption correction are generated from an input image.

[0037] Figure 21 This is a diagram used to illustrate the generation of an intermediate image from an input image in a fifth variation of an implementation.

[0038] Figure 22 This is a diagram used to illustrate the generation of an absorption coefficient image based on an intermediate image in a fifth variation of an embodiment.

[0039] Figure 23 This is a diagram illustrating a machine learning model for a sixth variation of an implementation. Detailed Implementation

[0040] The embodiments embodied in the present invention will now be described with reference to the accompanying drawings.

[0041] (Structure of a PET unit)

[0042] Reference Figure 1 and Figure 2 The structure of a PET (Positron Emission Tomography) device 1 according to one embodiment will be described.

[0043] like Figure 1 As shown, the PET device 1 is a device that photographs a subject 100 by detecting radiation (gamma rays) generated within the subject 100 due to pre-ingestion of a radioactive agent. The subject 100 is a human. The radiation (gamma rays) is annihilation radiation generated within the subject 100 due to the annihilation of electron-electron pairs between positrons generated from the radioactive agent and atoms near those positrons. The PET device 1 is configured to generate a radioactivity distribution image 10 of the subject 100 based on the photographic results of the subject 100 (see reference). Figure 3 Furthermore, the PET device 1 can be configured to take images of the whole body of the subject 100, or it can be configured to take images of a part of the subject 100 (such as the breast and head). Additionally, the PET device 1 is an example of the "nuclear medicine diagnostic device" claimed in the claims.

[0044] The PET device 1 includes a detector ring 2 surrounding the subject 100. The detector ring 2 is arranged in multiple layers stacked along the body axis of the subject 100. Multiple radiation (gamma ray) detectors 3 (see reference) are disposed inside the detector ring 2. Figure 2Thus, detector ring 2 is configured to detect radiation (gamma rays) generated from a radioactive agent within the subject 100. Furthermore, detector ring 2 is an example of the "detection section" in the claims.

[0045] Additionally, the PET unit 1 includes a control unit 4. The control unit 4 includes a simultaneous counting circuit 40 and a processing circuit 41. Furthermore, in Figure 1 The image only shows the radiation from detector 3 (reference 3). Figure 2 Two wirings are connected to the control unit 4 (simultaneous counting circuit 40), but in fact, the control unit 4 (simultaneous counting circuit 40) is connected to the photomultiplier tube (PMT: Photo Multiplier Tube) 33 (see later) of the radiation detector 3. Figure 2 The total number of channels corresponds to the number of wirings. Furthermore, processing circuit 41 is an example of the "processing unit" in the claims. Additionally, sensors other than PMTs, such as SiPM (Silicon Photomultiplier), are sometimes used.

[0046] like Figure 2 As shown, the radiation detector 3 includes a scintillator block 31, a light guide 32, and a photomultiplier tube 33. Alternatively, the light guide 32 may not be used in some cases.

[0047] Scintillator block 31 will remove the radioactive agent from subject 100 (refer to...) Figure 1 The radiation (gamma rays) produced is converted into light. When the subject 100 ingests the radioactive agent, the positrons of the positron-emitting RI annihilate, thereby producing two radiation rays (gamma rays). The scintillator elements constituting the scintillator block 31 emit light as the radiation (gamma rays) are incident, thereby converting the radiation (gamma rays) into light.

[0048] The light guide 32 is optically coupled to the scintillator block 31 and the photomultiplier tube 33, respectively. The light emitted from the scintillator element in the scintillator block 31 diffuses in the scintillator block 31 and is input to the photomultiplier tube 33 via the light guide 32.

[0049] The photomultiplier tube 33 amplifies the light input via the photoconductor 32 and converts it into an electrical signal. This electrical signal is sent to the simultaneous counting circuit 40 (see reference). Figure 1 ).

[0050] Simultaneously counting circuit 40 (refer to) Figure 1 The detection signal data (count value) is generated based on the electrical signal sent from the photomultiplier tube 33.

[0051] Specifically, the simultaneous counting circuit 40 (refer to...) Figure 1The position of the scintillator block 31 and the incident timing of the radiation (gamma rays) are checked. The transmitted electrical signal is determined to be appropriate data only when the radiation (gamma rays) simultaneously incident on the two scintillator blocks 31 located on both sides of the subject 100 (diagonally opposite the center of the subject 100). That is, the simultaneous counting circuit 40 detects, based on the aforementioned electrical signal, whether the radiation (gamma rays) is simultaneously observed (i.e., simultaneously counted) in the two radiation detectors 3 located on both sides of the subject 100 (diagonally opposite the center of the subject 100).

[0052] The detection signal data (count value), consisting of appropriate data determined to be counted simultaneously by the simultaneous counting circuit 40, is sent to the processing circuit 41 (refer to...). Figure 1 The processing circuit 41 generates a radioactivity distribution image 10 within the subject 100 based on the detection of radiation (gamma rays) by the detector ring 2 (see reference). Figure 3 ).

[0053] (Generation of a radioactive distribution image)

[0054] Next, refer to Figure 3 Flowcharts and Figures 4-6 The radioactivity distribution image generation process performed by the PET apparatus 1 in one embodiment will be described below. Furthermore, the radioactivity distribution image generation process is performed by the processing circuit 41 of the control unit 4.

[0055] like Figure 3 and Figure 4 As shown, firstly, in step 101, measurement data 5 is obtained based on the detection of radiation emitted from the subject 100.

[0056] Then, in step 102, the input image 6 is generated by image processing the measurement data 5. Specifically, in step 102, the input image 6 is generated by performing histogram-based image processing, machine learning-based image processing, or processing including back projection. As histogram-based image processing, the following method can be used: based on the TOF (Time of Flight) information contained in the measurement data 5, events are added at the positions with the highest probability, thereby performing image processing. As machine learning-based image processing, the following method can be used: image processing is performed using a machine learning model that converts the measurement data 5 into the input image 6. As processing including back projection, for example, simple back projection processing and reconstruction processing can be used. As reconstruction processing, for example, analytical reconstruction processing and successive approximation reconstruction processing can be used. As analytical reconstruction processing, for example, the FBP (Filtered Back Projection) method can be used. As successive approximation reconstruction processing, for example, the OSEM (Ordered Subsets Expectation Maximization) method can be used. In step 102, for example, a reconstruction process is performed. In this case, input image 6 is the reconstructed image.

[0057] Input image 6 is an image representing the interior of subject 100. Input image 6 includes at least one of the following: a three-dimensional image, an axial cross-sectional image, a coronal cross-sectional image, a sagittal cross-sectional image, a small patch image obtained by cutting out a region from a three-dimensional image, a small patch image obtained by cutting out a region from an axial cross-sectional image, a small patch image obtained by cutting out a region from a coronal cross-sectional image, and a small patch image obtained by cutting out a region from a sagittal cross-sectional image. Here, "cross-sectional image" refers to a two-dimensional image of a slice. Furthermore, an axial cross-sectional image refers to an image of a cross-section orthogonal to the body axis. Furthermore, a coronal cross-sectional image refers to an image of a transverse cross-section parallel to the body axis. Furthermore, a sagittal cross-sectional image refers to an image of a longitudinal cross-section parallel to the body axis. Additionally, input image 6 can be a single slice or multiple consecutive slices.

[0058] Additionally, in step 102, input image 6 is generated without performing at least one of absorption correction processing and scattering correction processing. Absorption correction processing is a process that corrects for the absorption of radiation within the subject 100. Scattering correction processing is a process that corrects for the scattering of radiation within the subject 100. In step 102, an uncorrected input image 6 without performing at least one of absorption correction processing and scattering correction processing is generated based on measurement data 5.

[0059] Furthermore, in step 102, image quality conversion processing can be performed without it, or it can be performed with image quality conversion processing, or it can be performed with region recognition processing. In this embodiment, the input image 6 may include at least one of the following: an image obtained without image quality conversion processing, an image obtained with image quality conversion processing, and an image obtained with region recognition processing. Image quality conversion processing may include, for example, gamma correction processing, histogram flattening processing, smoothing processing, and edge detection processing. Additionally, image quality conversion processing may include adding random noise with uniform, normal, Poisson, or Laplacian distributions. Furthermore, image quality conversion processing may include making the image as a whole or a specific region of the image a constant multiple. Furthermore, region recognition processing may include identifying the contours of the subject 100 in the image.

[0060] Then, in step 103, an intermediate image 7, including images related to the tissue region, is generated based on the input image 6. Specifically, in step 103, the intermediate image 7 is generated by applying a pre-learned machine learning model 8 to the input image 6. The machine learning model 8 is a machine learning model that takes the input image 6 as input and the intermediate image 7 as output. The machine learning model 8 includes at least one of the following: a machine learning model that takes a 3D image as input, a machine learning model that takes an axial section image as input, a machine learning model that takes a coronal section image as input, a machine learning model that takes a sagittal section image as input, a machine learning model that takes a small patch image cut from a 3D image as input, a machine learning model that takes a small patch image cut from an axial section image as input, a machine learning model that takes a small patch image cut from a coronal section image as input, and a machine learning model that takes a small patch image cut from a sagittal section image as input.

[0061] Furthermore, in the accompanying drawings of this embodiment ( Figure 5 , Figure 6 as well as Figure 8 In the example shown in the figure, for ease of explanation, the machine learning model 8 takes the input image 6 of the axial section image as input and the intermediate image 7 corresponding to the axial section image as output.

[0062] The intermediate image 7 is composed of a combination of N (finite number) tissues with known absorption coefficients, such as the brain, bones, skin, muscles, and internal organs. For example, if the measurement data 5 is measurement data of a human head, the elements (tissues) of the image related to the tissue region constituting the intermediate image 7 include at least one of the background (outside the subject), cavities (such as the nasal cavity and oral cavity), soft tissues (such as the brain and skin), and bones (skull). Furthermore, for example, if the measurement data 5 is measurement data of a human breast, the elements (tissues) of the image related to the tissue region constituting the intermediate image 7 include at least one of the background (outside the subject) and soft tissues.

[0063] In this embodiment, such as Figure 5 As shown, the intermediate image 7 includes a tissue composition ratio image 71, representing the proportion of tissue contained in each pixel, as an image related to the tissue region. The tissue composition ratio image 71 is a multi-channel image where the proportions of multiple tissues contained in each pixel are set as pixel values. Figure 5 In the example shown, image 71, representing the tissue composition, is an image of a human head, comprising images in four channels: background, cavities, soft tissue, and bone. The background channel is configured so that the proportion of background contained in each pixel is set as the pixel value. Similarly, the cavity channel is configured so that the proportion of cavities contained in each pixel is set as the pixel value. The soft tissue channel is configured so that the proportion of soft tissue contained in each pixel is set as the pixel value. The bone channel is configured so that the proportion of bone contained in each pixel is set as the pixel value. Furthermore, since the pixel values ​​of each of the four channels represent proportions, the sum of the pixel values ​​of the four channels for a given pixel is 1.

[0064] Then, as Figure 3 and Figure 4 As shown, in step 104, an absorption coefficient image 9 is generated based on the intermediate image 7 and the known absorption coefficients of the tissue region. In this embodiment, as... Figure 6 As shown, in step 104, an absorption coefficient image 9 is generated by assigning absorption coefficients to the tissues in the tissue composition ratio image 71 based on known absorption coefficients. Specifically, in step 104, the absorption coefficient image 9 is generated by performing a linear combination process on the tissue composition ratio images 71 of each tissue with known absorption coefficients as coefficients. More specifically, the linear combination process on the tissue composition ratio images 71 of each tissue with known absorption coefficients as coefficients is performed using the following equation (1).

[0065] [Number 1]

[0066]

[0067] in,

[0068] n: Organization label (organization number)

[0069] j: pixel number

[0070] μ j : Absorption coefficient of pixel j

[0071] μ * n The absorption coefficient of tissue n (a known absorption coefficient).

[0072] r nj : The composition ratio of pixel j to the organization n.

[0073] In addition, r nj It satisfies the following equation (2).

[0074] [Number 2]

[0075]

[0076] For example, in the case where the tissue composition ratio of image 71 is an image of a human head, and includes four channels of images: background, cavities, soft tissue, and bone, the background absorption coefficient μ, which is generally known, is used. * 0. Absorption coefficient μ of voids * 1. Absorption coefficient μ of soft tissue * 2 and the bone resorption coefficient μ * 3. Using the above formula (1), perform linear combination processing of the tissue composition ratio image 71 of each tissue with known absorption coefficients as coefficients.

[0077] Then, as Figure 3 and Figure 4 As shown, in step 105, a radioactivity distribution image 10 is generated by reconstructing the image based on the absorption coefficient image 9 and the measurement data 5. At this time, at least one of absorption correction processing and scattering correction processing is performed based on the absorption coefficient image 9. For example, in step 105, absorption correction processing is performed based on the absorption coefficient image 9, and scattering correction processing is performed based on the scattering distribution data obtained from the absorption coefficient image 9 and the measurement data 5. In step 105, a quantitative radioactivity distribution image 10 obtained by performing absorption correction processing and scattering correction processing is generated.

[0078] (Machine Learning Model)

[0079] Next, refer to Figures 7-9 The machine learning model 8 of a PET device 1 according to one embodiment will be explained. Furthermore, for ease of understanding, the input image 6 and the intermediate image 7, which serve as learning data, will be referred to as input image 6a and intermediate image 7a, respectively.

[0080] like Figure 7 As shown, machine learning model 8 uses multiple pairs of input image 6a and intermediate image 7a as learning data and learns through supervised learning. Specifically, machine learning model 8 takes the pre-prepared input image 6a as input and the pre-prepared intermediate image 7a as training images (correct images) for learning. Furthermore, the details of the learning process of machine learning model 8 will be described later.

[0081] Furthermore, machine learning model 8 includes a deep neural network. The deep neural network of machine learning model 8 also includes convolutional processing. That is, machine learning model 8 includes a deep convolutional neural network. For example, a U-Net with skip connections can be used as the deep convolutional neural network of machine learning model 8. Furthermore, the softmax function can be used as the activation function of the deep convolutional neural network of machine learning model 8.

[0082] Reference Figure 8 Flowcharts and Figure 9 This document describes a method for creating the machine learning model 8 of a PET device 1 according to one embodiment (a method for creating a learned model).

[0083] like Figure 8 and Figure 9 As shown, firstly, in step 111, a tissue label image 11 is prepared, in which each pixel represents the tissue to which it belongs. The tissue label image 11 can be prepared by performing region segmentation processing on medical images such as MR images and CT images. Alternatively, the tissue label image 11 can be prepared by acquiring tissue label images publicly available on the Internet (e.g., BrainWeb).

[0084] Then, in step 112, a pseudo-radioactivity distribution image 12 and a pseudo-absorption coefficient image 13 are created based on the tissue label image 11. Specifically, the pseudo-radioactivity distribution image 12 is created by assigning radioactivity concentrations to each tissue in the tissue label image 11. Additionally, the pseudo-absorption coefficient image 13 is created by assigning absorption coefficients to each tissue in the tissue label image 14, which is created by integrating the labels of each tissue in the tissue label image 11. The tissue label image 14 is an image with a reduced number of labels compared to the tissue label image 11 by integrating the labels.

[0085] Then, in step 113, pseudo-measurement data 15 is generated by performing simulation calculations based on the pseudo-radioactivity distribution image 12 and the pseudo-absorption coefficient image 13. Specifically, in step 113, the pseudo-radioactivity distribution image 12, the pseudo-absorption coefficient image 13, and various simulation conditions are input and simulation calculations are performed to generate pseudo-measurement data 15. For example, Monte Carlo simulation calculations and analytical simulation calculations can be used for the simulation calculations. In this embodiment, the machine learning model 8 uses the pseudo-measurement data 15 generated based on at least one of Monte Carlo simulation calculations and analytical simulation calculations for learning. For example, the machine learning model 8 uses the pseudo-measurement data 15 generated based on analytical simulation calculations, which is a combination of Monte Carlo simulation calculations and analytical simulation calculations, for learning.

[0086] Then, in step 114, a pseudo-reconstructed image 16 is generated by performing a back-projection process (image processing) on ​​the pseudo-measurement data 15. Specifically, in step 114, the pseudo-measurement data 15 and various reconstruction conditions, including pixel size, are input and reconstruction processing is performed, thereby generating the pseudo-reconstructed image 16. In the reconstruction processing, resolution information (pixel size information) is input as a parameter. Furthermore, the pseudo-reconstructed image 16 is an example of the "pseudo-image" of the claims.

[0087] Furthermore, in step 114, a normalized pseudo-reconstructed image 16 is generated by performing a normalization process on the pseudo-reconstructed image 16 to normalize the pixel value range to [0, 1]. Additionally, in step 114, the normalized pseudo-reconstructed image 16 can be multiplied by a coefficient greater than 0 and less than 1, or a specific region of the normalized pseudo-reconstructed image 16 or the pseudo-reconstructed image 16 before normalization can be multiplied by a positive coefficient. In this way, the machine learning model 8 can learn from the pseudo-reconstructed images 16 with various pixel values. In this embodiment, the input image 6a (pseudo-reconstructed image 16) serving as the learning data for the machine learning model 8 includes at least one of the following: a normalized image with a normalized pixel value range, an image obtained by multiplying the normalized image by a coefficient greater than 0 and less than 1, and an image obtained by multiplying a specific region of the normalized image or the image before normalization by a positive coefficient.

[0088] When multiplying the normalized image by a coefficient greater than 0 and less than 1, for example, multiplying by 1 / n (where n is a positive integer). In this case, setting n = 2 to 10 (in increments of 1), n ​​= 20 to 100 (in increments of 10), and n = 200 to 1000 (in increments of 100), etc., a large number of input images 6a with different pixel values ​​(image brightness) can be generated.

[0089] Furthermore, by multiplying a specific region of the normalized image or the unnormalized image by a positive coefficient, regions on a tissue-by-tissue basis can be used as specific regions. For example, if the normalized image is an image of a human head, the gray matter, white matter, cerebellum, skin of the head, and muscles of the head can be used as specific regions. Thus, the machine learning model 8 can be trained considering the diversity of radioactive distribution caused by individual differences, differences in radiopharmaceuticals, etc.

[0090] Furthermore, in step 114, image quality conversion processing can be performed without it, or image quality conversion processing can be performed, or region recognition processing can be performed. In this embodiment, the input image 6a includes at least one of the following: an image obtained without image quality conversion processing, an image obtained with image quality conversion processing, and an image obtained with region recognition processing. Image quality conversion processing can include, for example, gamma correction processing, histogram flattening processing, smoothing processing, and edge detection processing. Additionally, image quality conversion processing can include, for example, adding random noise with a uniform, normal, Poisson, or Laplacian distribution. Furthermore, image quality conversion processing can include making the image as a whole or a specific region of the image a constant multiple. In this case, the machine learning model 8 can learn from the input image 6a with multiple pixel values. Additionally, region recognition processing can include identifying the contours of the subject 100 in the image.

[0091] Furthermore, in step 115, an intermediate image (training image) 7a is created as learning data based on the tissue label image 11. Specifically, in step 115, the intermediate image 7a is created based on the tissue label image 14, which is created by integrating the labels of each tissue from the tissue label image 11. More specifically, assuming the same low-resolution pixel size as the pseudo-reconstructed image 16, the proportion (composition ratio) of each tissue contained in a pixel is calculated for the high-resolution tissue label image 14, thereby creating the intermediate image 7a as a tissue composition ratio image.

[0092] Then, in step 116, a large number of pseudo-reconstructed images 16 and intermediate images 7a are used as learning data to create a machine learning model 8 that has been trained. In this embodiment, the machine learning model 8 uses pseudo-reconstructed images 16 generated based on at least one of Monte Carlo simulation calculation and analytical simulation calculation for training.

[0093] (Effects of this implementation method)

[0094] In this embodiment, the following effect can be obtained.

[0095] In this embodiment, as described above, an input image 6 is generated by image processing of measurement data 5 obtained based on the detection of radiation emitted from the subject 100. An intermediate image 7, including an image related to the tissue region, is generated based on the input image 6. An absorption coefficient image 9 is generated based on the intermediate image 7 and the known absorption coefficient of the tissue region. Thus, the absorption coefficient image 9 can be generated based on the intermediate image 7, including an image related to the tissue region. As a result, even when the subject 100 is not subjected to CT or MR imaging and the absorption coefficient image 9 is generated based on the measurement data 5, the absorption coefficient of the absorption coefficient image 9 can be guaranteed to be within an appropriate range (a generally acceptable value).

[0096] Furthermore, in this embodiment, as described above, the intermediate image 7 includes a tissue composition ratio image 71, which represents the proportion of tissue contained in each pixel, as an image related to the tissue region. Therefore, when the intermediate image 7 includes the tissue composition ratio image 71, an absorption coefficient image 9 with an absorption coefficient within an appropriate range can be easily generated based on the proportion of tissue contained in each pixel of the tissue composition ratio image 71.

[0097] Furthermore, in this embodiment, as described above, the step of generating the absorption coefficient image 9 includes assigning absorption coefficients to the tissues in the tissue composition ratio image 71 based on the known absorption coefficients of each tissue region. Therefore, based on the tissue composition ratio image 71 with absorption coefficients assigned according to known absorption coefficients, it is possible to easily generate an absorption coefficient image 9 with absorption coefficients within an appropriate range.

[0098] Furthermore, in this embodiment, as described above, the step of generating the input image 6 includes generating the input image 6 without performing at least one of the absorption correction processing and the scattering correction processing. Therefore, compared to generating the input image 6 by performing at least one of the absorption correction processing and the scattering correction processing, the processing for generating the input image 6 can be simplified to a degree corresponding to the absence of at least one of the absorption correction processing and the scattering correction processing.

[0099] Furthermore, in this embodiment, as described above, the step of generating the input image 6 includes a step of performing a back-projection process on the measurement data 5. Therefore, by performing a back-projection process on the measurement data 5, the input image 6 can be easily generated.

[0100] Furthermore, in this embodiment, as described above, the input image 6 includes at least one of the following: an image obtained from the image-processed measurement data 5 without applying image quality conversion processing, an image obtained from the measurement data 5 with image quality conversion processing applied, and an image obtained from the measurement data 5 with region recognition processing applied. Therefore, an intermediate image 7 can be generated based on at least one of the image obtained from the image-processed measurement data 5 with image quality conversion processing applied and the image obtained from the measurement data 5 with region recognition processing applied.

[0101] Furthermore, in this embodiment, as described above, the step of generating the intermediate image 7 includes applying a pre-learned machine learning model 8 to the input image 6. Therefore, the intermediate image 7 can be easily generated simply by applying the pre-learned machine learning model 8 to the input image 6.

[0102] Furthermore, in this embodiment, as described above, the input image 6a serving as the learning data for the machine learning model 8 includes at least one of the following: a normalized image whose pixel value range has been normalized, an image obtained by multiplying the normalized image by a coefficient greater than 0 and less than 1, and an image obtained by multiplying a specific region of the normalized image or the image before normalization by a positive coefficient. Therefore, the machine learning model 8 can learn from the input image 6a with multiple pixel values. As a result, a machine learning model 8 with strong resistance to pixel value biases can be created.

[0103] Furthermore, in this embodiment, as described above, the machine learning model 8 includes at least one of the following: a machine learning model 8 that takes a 3D image as input, a machine learning model 8 that takes an axial section image as input, a machine learning model 8 that takes a coronal section image as input, a machine learning model 8 that takes a sagittal section image as input, a machine learning model 8 that takes a small patch image cut from a 3D image as input, a machine learning model 8 that takes a small patch image cut from an axial section image as input, a machine learning model 8 that takes a small patch image cut from a coronal section image as input, and a machine learning model 8 that takes a small patch image cut from a sagittal section image as input. Thus, an intermediate image 7 can be generated using a 3D image, an axial section image, a coronal section image, a sagittal section image, a small patch image cut from a 3D image, a small patch image cut from an axial section image, a small patch image cut from a coronal section image, or a small patch image cut from a sagittal section image as input.

[0104] Furthermore, in this embodiment, as described above, the machine learning model 8 includes a deep neural network. Therefore, learning can be performed effectively using the machine learning model 8 that includes a deep neural network.

[0105] Furthermore, in this embodiment, as described above, the deep neural network includes convolutional processing. Therefore, learning can be performed more effectively using a deep neural network that includes convolutional processing.

[0106] Furthermore, in this embodiment, as described above, the machine learning model 8 uses a pseudo-reconstructed image 16 generated based on at least one of Monte Carlo simulation calculations and analytical simulation calculations for learning. Therefore, the machine learning model 8 can be created using a pseudo-reconstructed image 16 generated based on at least one of Monte Carlo simulation calculations and analytical simulation calculations. As a result, unlike the case where the machine learning model 8 is created using actual reconstructed images (clinical images), it is not necessary to collect a large number of clinical images. Therefore, the machine learning model 8 can be easily created.

[0107] Furthermore, in this embodiment, as described above, the step of generating the absorption coefficient image 9 includes the following steps: when the intermediate image 7 includes a tissue composition ratio image 71, a linear combination process is performed on the tissue composition ratio images 71 of each tissue with known absorption coefficients as coefficients. Therefore, when the intermediate image 7 includes a tissue composition ratio image 71, by performing a linear combination process on the tissue composition ratio images 71 of each tissue with known absorption coefficients as coefficients, an absorption coefficient image 9 with absorption coefficients within an appropriate range can be easily generated.

[0108] Furthermore, in this embodiment, as described above, the measurement data 5 is measurement data of a human head, and the elements constituting the intermediate image 7 related to the tissue region include at least one of the following: background, cavities, soft tissue, and bone. Therefore, when the measurement data 5 is measurement data of a human head, an absorption coefficient image 9 with an absorption coefficient within an appropriate range can be easily generated from the intermediate image 7, which includes an image related to the tissue region.

[0109] Furthermore, in this embodiment, as described above, the measurement data 5 is measurement data of a human breast, and the elements constituting the intermediate image 7 related to the tissue region include at least one of the background and soft tissue. Therefore, when the measurement data 5 is measurement data of a human breast, an absorption coefficient image 9 with an absorption coefficient within an appropriate range can be easily generated from the intermediate image 7, which includes an image related to the tissue region.

[0110] Furthermore, in this embodiment, as described above, the processing circuit 41 is configured to perform at least one of absorption correction processing and scattering correction processing based on the absorption coefficient image 9. Therefore, appropriate absorption correction processing or appropriate scattering correction processing can be performed based on the absorption coefficient image 9, which has an absorption coefficient within an appropriate range.

[0111] Furthermore, in this embodiment, as described above, the method for creating the learned model includes the following steps: preparing a tissue label image 11 representing the tissue to which each pixel belongs; creating a pseudo-radioactivity distribution image 12 and a pseudo-absorption coefficient image 13 based on the tissue label image 11; creating pseudo-measurement data 15 by performing simulation calculations based on the pseudo-radioactivity distribution image 12 and the pseudo-absorption coefficient image 13; generating a pseudo-reconstructed image 16 by image processing the pseudo-measurement data 15; and creating a learned model (machine learning model 8) using the pseudo-reconstructed image 16 as learning data. Thus, the learned model (machine learning model 8) can be created using the pseudo-reconstructed image 16 obtained through simulation calculations as learning data. As a result, unlike the case where the learned model (machine learning model 8) is created using actual reconstructed images (clinical images) as learning data, it is not necessary to collect a large number of clinical images. Therefore, the learned model (machine learning model 8) can be created without the work involved in collecting a large number of clinical images, which is not easy from the perspective of personal information protection.

[0112] (First variation)

[0113] Next, refer to Figure 10 The first variation of the above-described embodiment will be described below. In this first variation, an example in which the intermediate image includes a tissue label image will be described. Furthermore, for structures identical to those in the above-described embodiment, the same reference numerals will be used in the illustrations, and their descriptions will be omitted.

[0114] like Figure 10 As shown, in the first variation of the above embodiment, the intermediate image 7 includes a tissue label image 72 representing the tissue to which each pixel belongs, as an image related to the tissue region. Therefore, when the intermediate image 7 includes the tissue label image 72, an absorption coefficient image 9 with an absorption coefficient within an appropriate range can be easily generated based on the tissue to which each pixel of the tissue label image 72 belongs. In the tissue label image 72, if multiple tissues coexist in a single pixel, the pixel is labeled as belonging to the tissue occupying the largest proportion.

[0115] exist Figure 10 In the example shown, tissue-labeled image 72 is an image of a human head, including four labels: background, cavities, soft tissue, and bone. The background label region is configured such that the pixel values ​​assigned to the background label are set to the specified pixel values. Similarly, the cavity label region is configured such that the pixel values ​​assigned to the cavity label are set to the specified pixel values. The soft tissue label region is configured such that the pixel values ​​assigned to the soft tissue label are set to the specified pixel values. The bone label region is configured such that the pixel values ​​assigned to the bone label are set to the specified pixel values. Furthermore, each of the four labels is assigned a distinct pixel value (an integer value).

[0116] In a first variation of the above embodiment, in step 104 of the above embodiment, an absorption coefficient image 9 is generated based on the tissue label image 72 of the intermediate image 7 and the known absorption coefficients of the tissue regions. Specifically, absorption coefficients are assigned to the tissue in the tissue label image 72 based on the known absorption coefficients. Therefore, an absorption coefficient image 9 with absorption coefficients within an appropriate range can be easily generated based on the tissue label image 72 with absorption coefficients assigned according to the known absorption coefficients. More specifically, an assignment process is performed on the known absorption coefficients corresponding to the label values ​​of the tissue label image 72. Therefore, when the intermediate image 7 includes the tissue label image 72, by performing an assignment process on the known absorption coefficients corresponding to the label values ​​of the tissue label image 72, an absorption coefficient image 9 with absorption coefficients within an appropriate range can be easily generated.

[0117] In the first variation of the above embodiment, the known absorption coefficient corresponding to the label value of the tissue label image 72 is assigned using the following formula (3).

[0118] [Number 3]

[0119]

[0120] in,

[0121] j: pixel number

[0122] l j : Label value (organization number) of pixel j

[0123] μ j : Absorption coefficient of pixel j

[0124] μ * l : Absorption coefficient of the label value l (the known absorption coefficient).

[0125] For example, in the case where tissue-labeled image 72 is an image of a human head and includes four labels: background, cavities, soft tissue, and bone, the commonly known absorption coefficient μ of the background is used. * 0. Absorption coefficient μ of voids * 1. Absorption coefficient μ of soft tissue * 2 and the bone resorption coefficient μ * 3. Using the above formula (3), the known absorption coefficients corresponding to the label values ​​of the tissue label image 72 are assigned.

[0126] Furthermore, the machine learning model 8, which outputs the tissue label image 72, does not directly output the tissue label image 72. Instead, it outputs a reliability score for each pixel as an intermediate output. Reliability is a probabilistic indicator used to determine which label a pixel belongs to. Moreover, the machine learning model 8, which outputs the tissue label image 72, is set to output the tissue label image 72 based on the label with the highest obtained reliability. Here, the reliability score can be set to a value between 0 and 1, and the sum of the reliability scores of all labels can be set to 1. Therefore, by replacing the tissue composition ratio in the above embodiment with the reliability score and performing linear combination processing similarly to the tissue composition ratio image 71 in the above embodiment, the absorption coefficient image 9 can also be generated. That is, the absorption coefficient image 9 can also be generated by performing linear combination processing of the reliability image, which is an intermediate output of the tissue label image 72 and has known absorption coefficients set as coefficients. Therefore, when the intermediate image 7 includes the tissue label image 72, by performing linear combination processing of the reliability image as the intermediate output of the tissue label image 72 with the known absorption coefficient as the coefficient, it is possible to easily and accurately generate the absorption coefficient image 9 with the absorption coefficient being within an appropriate range.

[0127] In addition, in the first variation of the above embodiment, in step 115 of the above embodiment, assuming the same low-resolution pixel size as the pseudo-reconstructed image 16, for the high-resolution tissue label image 14, the proportion (composition ratio) of each tissue contained in a pixel is calculated, and the pixel is labeled as belonging to the tissue occupying the largest proportion, thereby creating an intermediate image 7a as a tissue label image.

[0128] (Second variation)

[0129] Next, refer to Figure 11 and Figure 12 A second variation of the above-described embodiment will now be described. In this second variation, an example combining multiple machine learning models will be explained. Furthermore, structures identical to those in the above-described embodiment will be illustrated using the same reference numerals, and their descriptions will be omitted.

[0130] like Figure 11As shown, in the second variation of the above embodiment, the machine learning model 8 includes three machine learning models: a machine learning model for axial sections, a machine learning model for coronal sections, and a machine learning model for sagittal sections. The machine learning model for axial sections uses the input image 6a of the axial section image as input and the intermediate image 7a corresponding to the axial section image as the training image for learning. The machine learning model for coronal sections uses the input image 6a of the coronal section image as input and the intermediate image 7a corresponding to the coronal section image as the training image for learning. The machine learning model for sagittal sections uses the input image 6a of the sagittal section image as input and the intermediate image 7a corresponding to the sagittal section image as the training image for learning.

[0131] Furthermore, in the second variation of the above-described embodiment, such as Figure 12 As shown, in step 102 of the above embodiment, three input images 6 are generated: axial section image, coronal section image, and sagittal section image.

[0132] Furthermore, in the second variation of the above embodiment, in step 103 of the above embodiment, three intermediate images 7 are generated: a tissue composition ratio image corresponding to the axial section image, a tissue composition ratio image corresponding to the coronal section image, and a tissue composition ratio image corresponding to the sagittal section image.

[0133] Furthermore, in the second variation of the above embodiment, in step 104 of the above embodiment, an absorption coefficient image 9 is generated based on three intermediate images 7: the tissue composition ratio image corresponding to the axial section image, the tissue composition ratio image corresponding to the coronal section image, and the tissue composition ratio image corresponding to the sagittal section image. Specifically, a section transformation process is performed such that any two of the tissue composition ratio images corresponding to the axial section image, the coronal section image, and the sagittal section image become the image corresponding to the remaining section. Furthermore, an average image (average tissue composition ratio image) of the three tissue composition ratio images with consistent sections is generated. At this time, a simple averaging process can be performed, or a weighted averaging process with high precision multiplying the sections by weights can be performed. Moreover, a linear combination process of the average tissue composition ratio image is performed in the same manner as in the above embodiment, thereby generating the absorption coefficient image 9.

[0134] (First and second variations of the second variation)

[0135] Next, refer to Figure 13 and Figure 14The first and second modifications of the second variations of the above-described embodiments will be explained below. In the first and second modifications of the second variations of the above-described embodiments, examples in which the intermediate image includes a tissue label image will be described. Furthermore, structures identical to those in the above-described embodiments will be illustrated using the same reference numerals, and their descriptions will be omitted.

[0136] like Figure 13 As shown, in the first variation of the second variation of the above embodiment, in step 103 of the above embodiment, three intermediate images 7 are generated: a tissue label image corresponding to the axial section image, a tissue label image corresponding to the coronal section image, and a tissue label image corresponding to the sagittal section image.

[0137] Furthermore, in the first variation of the second variation of the above-described embodiment, in step 104 of the above-described embodiment, an absorption coefficient image 9 is generated based on three intermediate images 7: a tissue label image corresponding to the axial section image, a tissue label image corresponding to the coronal section image, and a tissue label image corresponding to the sagittal section image. Specifically, a section transformation process is performed such that any two of the tissue label images corresponding to the axial section image, the coronal section image, and the sagittal section image become images corresponding to the remaining section. Moreover, a majority decision image (majority decision tissue label image) is generated, in which the label value of each pixel is determined by the majority decision of the three tissue label images with consistent sections. At this time, if the label value cannot be determined by majority decision, the label value of the pre-determined section can be used. Furthermore, similar to the first variation of the above-described embodiment, the absorption coefficient image 9 is generated by performing a majority decision tissue label image allocation process.

[0138] Additionally, absorption coefficient images can be generated using methods other than majority decision. Specifically, such as... Figure 14 As shown, in the second variation of the second variation of the above embodiment, unlike the first variation of the second variation of the above embodiment, three absorption coefficient images are generated in a manner corresponding to each of the three tissue label images that correspond to the cross-section. Furthermore, an absorption coefficient image that is the average of the three absorption coefficient images is generated as the final absorption coefficient image 9.

[0139] (Third variation)

[0140] Next, refer to Figure 15A third variation of the above-described embodiment will now be described. In this third variation, an example of applying multiple inputs to a machine learning model (multichanneling of the input image) will be explained. Furthermore, structures identical to those in the above-described embodiment will be illustrated using the same reference numerals, and their descriptions will be omitted.

[0141] like Figure 15 As shown, in the third variation of the above embodiment, the two input images 6—the reconstructed image generated by the reconstruction process and the backprojected image generated by the simple backprojection process—become the inputs to the machine learning model 8. In the third variation of the above embodiment, the machine learning model 8 outputs an tissue composition ratio image 71 of the intermediate image 7 based on the two input images 6.

[0142] Furthermore, the combination of input images 6 is not limited to a combination of reconstructed images generated through reconstruction processing and backprojected images generated through simple backprojection processing. For example, the combination of input images 6 can also be a combination of multiple reconstructed images with different reconstruction algorithms. Additionally, for example, the combination of input images 6 can also be a combination of multiple reconstructed images with different iterations of successive approximation image reconstruction processing. Furthermore, for example, the combination of input images 6 can also be a combination of multiple reconstructed images with different resolutions. In this case, input images 6 include images with two or more resolutions. Therefore, compared to the case of including only one resolution image, it is possible to generate an intermediate image 7 based on input images 6 with multiple resolutions. Additionally, the combination of input images 6 can also be a combination of multiple reconstructed images that have undergone different image processing. Furthermore, the combination of input images 6 can also be a combination of these images.

[0143] (The first to fourth variations of the third variation)

[0144] Next, refer to Figures 16-19 The first to fourth modifications of the third modification of the above-described embodiment will be explained. In the first to third modifications of the third modification of the above-described embodiment, examples in which the reconstructed image is combined with information other than the reconstructed image and the back-projected image will be described. Furthermore, for structures that are the same as those in the above-described embodiment, the same reference numerals are used in the illustrations, and their descriptions are omitted.

[0145] like Figure 16As shown, in the first variation of the third variation of the above embodiment, the input image (reconstructed image) 6 and the tissue label image 17 representing the tissue to which each pixel belongs are used as inputs to the machine learning model 8. In the first variation of the third variation of the above embodiment, the machine learning model 8 outputs a tissue composition ratio image 71 of the intermediate image 7 based on the input image 6 and the tissue label image 17 as auxiliary information.

[0146] In addition, such as Figure 17 As shown, in the second variation of the third variation of the above embodiment, both the input image (reconstructed image) 6 and the subject region indication image 18, representing the region of the subject 100, serve as inputs to the machine learning model 8. In the second variation of the third variation of the above embodiment, the machine learning model 8 outputs a tissue composition ratio image 71 of the intermediate image 7 based on the input image 6 and the subject region indication image 18 as auxiliary information. Alternatively, a single tissue region indication image representing a region of a single tissue may be used instead of the subject region indication image 18.

[0147] In addition, such as Figure 18 As shown, in the third variation of the above embodiment, the input image (reconstructed image) 6 and the mixed tissue number image 19, which represents the number of tissue types contained in a pixel, are used as inputs to the machine learning model 8. In the third variation of the above embodiment, the machine learning model 8 outputs a tissue composition ratio image 71 of the intermediate image 7 based on the input image 6 and the mixed tissue number image 19 as auxiliary information.

[0148] In addition, such as Figure 19 As shown, in the fourth variation of the third variation of the above embodiment, both the input image (reconstructed image) 6 and non-image information, namely, information 20 related to the spatial position of the input image 6, become the inputs to the machine learning model 8. That is, in the fourth variation of the third variation of the above embodiment, the machine learning model 8 takes the input image 6 as input and the information 20 related to the spatial position of the input image 6 as input. Thus, by taking not only the input image 6 as input but also the information 20 related to the spatial position of the input image 6 as input, the intermediate image 7 can be generated effectively. The information 20 related to the spatial position of the input image 6 can be, for example, the distance from the center of gravity of the subject and the relative distance. In the fourth variation of the third variation of the above embodiment, the machine learning model 8 outputs the tissue composition ratio image 71 of the intermediate image 7 based on the input image 6 and the information 20 related to the spatial position of the input image 6 as auxiliary information.

[0149] (Fourth variation)

[0150] Next, refer to Figure 20 A fourth variation of the above-described embodiment will now be described. In this fourth variation, an example of generating multiple outputs from a machine learning model will be explained. Furthermore, structures identical to those in the above-described embodiment will be illustrated using the same reference numerals, and their descriptions will be omitted.

[0151] like Figure 20 As shown, in the fourth variation of the above embodiment, the output of the machine learning model 8 includes both the intermediate image 7 (the tissue composition ratio image 71) and the reconstructed image 21 obtained by applying at least one of absorption correction processing and scattering correction processing. In the fourth variation of the above embodiment, the machine learning model 8 outputs the reconstructed image 21 obtained by applying at least one of absorption correction processing and scattering correction processing, in addition to the intermediate image 7. Therefore, by applying the pre-learned machine learning model 8 to the input image 6, it is possible to generate not only the intermediate image 7 but also the reconstructed image 21 obtained by applying at least one of absorption correction processing and scattering correction processing. In the fourth variation of the above embodiment, for example, the machine learning model 8 outputs the intermediate image 7 and the reconstructed image 21 obtained by applying absorption correction processing.

[0152] In addition, in the fourth variation of the above-described embodiment, the machine learning model 8 includes a multi-output (multi-task) deep convolutional neural network that outputs two images, the intermediate image 7 and the reconstructed image 21.

[0153] (Fifth variation)

[0154] Next, refer to Figure 21 and Figure 22 The fifth variation of the above-described embodiment will now be described. In this fifth variation, an example of combining multiple machine learning models corresponding to each tissue in the tissue composition ratio image will be described. Furthermore, structures identical to those in the above-described embodiment will be illustrated using the same reference numerals, and their descriptions will be omitted.

[0155] like Figure 21 As shown, in the fifth variation of the above embodiment, the machine learning model 8 includes multiple machine learning models corresponding to the tissue composition of the intermediate image 7 compared to each tissue in image 71. Figure 21In the example shown, tissue composition ratio image 71 is an image of a human head. In this case, machine learning model 8 includes four machine learning models: a machine learning model for the background, a machine learning model for holes, a machine learning model for soft tissue, and a machine learning model for bone. The machine learning model for the background takes input image 6 as input and outputs tissue composition ratio image 71 corresponding to the background. The machine learning model for holes takes input image 6 as input and outputs tissue composition ratio image 71 corresponding to holes. The machine learning model for soft tissue takes input image 6 as input and outputs tissue composition ratio image 71 corresponding to soft tissue. The machine learning model for bone takes input image 6 as input and outputs tissue composition ratio image 71 corresponding to bone.

[0156] Furthermore, in the fifth variation of the above-described embodiment, such as Figure 22 As shown, the tissue composition ratio image 71 of each tissue with known absorption coefficients is linearly combined using the following equation (4).

[0157] [Number 4]

[0158]

[0159] in,

[0160] n: Organization number

[0161] j: pixel number

[0162] μ j : Absorption coefficient of pixel j

[0163] μ * n The absorption coefficient of tissue n (a known absorption coefficient).

[0164] r nj The composition ratio of pixel j to n (0≤r) nj ≤1).

[0165] In the fifth variation of the above embodiment, the tissue composition ratio image 71 of each tissue is generated by mutually independent machine learning models. Therefore, unlike the above embodiment, the condition that the sum of the composition ratios is 1 is not automatically satisfied. Therefore, in the fifth variation of the above embodiment, as shown in the above equation (4), normalization processing (dividing by the denominator term of the above equation (4)) is performed during linear combination processing.

[0166] (Sixth variation)

[0167] Next, refer to Figure 23The sixth variation of the above-described embodiment will now be explained. In this sixth variation, an example of a machine learning model processing images of different cross-sections will be described. Furthermore, structures identical to those in the above-described embodiment will be illustrated using the same reference numerals, and their descriptions will be omitted.

[0168] like Figure 23 As shown, in the sixth variation of the above embodiment, the machine learning model 8 includes a deep neural network for axial sections, a deep neural network for coronal sections, and a deep neural network for sagittal sections. The deep neural network for axial sections takes the input image 6, a three-dimensional axial section image, as input and outputs a tissue composition ratio image corresponding to the axial section image. The deep neural network for coronal sections takes the input image 6, a three-dimensional coronal section image, as input and outputs a tissue composition ratio image corresponding to the coronal section image. The deep neural network for sagittal sections takes the input image 6, a three-dimensional sagittal section image, as input and outputs a tissue composition ratio image corresponding to the sagittal section image.

[0169] Furthermore, in the sixth variation of the above embodiment, the machine learning model 8 is configured to perform cross-section transformation processing by taking any two of the tissue composition ratio images corresponding to the axial cross-section image, the coronal cross-section image, and the sagittal cross-section image as images corresponding to the remaining cross-section. Additionally, in the sixth variation of the above embodiment, the machine learning model 8 includes a deep neural network that takes three tissue composition ratio images corresponding to the cross-section as input and outputs a three-dimensional tissue composition ratio image corresponding to the three tissue composition ratio images corresponding to the cross-section. Thus, in the sixth variation of the above embodiment, an intermediate image 7 is generated in the form of a three-dimensional tissue composition ratio image.

[0170] [Variation Example]

[0171] Furthermore, the embodiments disclosed herein should be considered illustrative rather than restrictive in all respects. The scope of the invention is not indicated by the description of the above embodiments, but by the claims, and includes all modifications (variations) within the meaning and scope equivalent to the claims.

[0172] For example, in this invention, structures that can be applied to each other in the above embodiments and the first to sixth variations can be appropriately combined.

[0173] Furthermore, while the above embodiments illustrate an example of a PET device as a nuclear medicine diagnostic device, the present invention is not limited thereto. For example, a nuclear medicine diagnostic device may also be a SPECT (Single Photon Emission Computed Tomography) device other than a PET device.

[0174] Furthermore, the above embodiment illustrates an example of normalizing the pixel value range to [0, 1] for a pseudo-reconstructed image, but the present invention is not limited thereto. In the present invention, the normalization range can also be any range other than [0, 1], such as [-1, 1].

[0175] Furthermore, while the above embodiments illustrate an example of a machine learning model using pseudo-images prepared based on simulation calculations, the present invention is not limited to this. In the present invention, the machine learning model can also learn using both pseudo-images and actual images (images of the actual subjects). This allows the machine learning model to learn from a variety of data. As a result, a machine learning model with strong resistance to biases from individual subjects can be created. Additionally, the machine learning model can use a machine learning model learned from pseudo-images as a base model and then perform additional learning using actual images. Therefore, even when using both pseudo-images and actual images to learn, the machine learning model can learn efficiently.

[0176] Furthermore, the above embodiments illustrate an example of generating an input image without performing at least one of absorption correction and scattering correction processing, but the present invention is not limited thereto. In the present invention, it is also possible to generate an input image that has undergone both absorption correction and scattering correction processing.

[0177] Furthermore, the above embodiments illustrate an example of generating an absorption coefficient image by performing linear combination processing of tissue composition ratio images of each tissue with known absorption coefficients as coefficients; however, the present invention is not limited to this. In the present invention, when generating an absorption coefficient image based on a tissue composition ratio image, the absorption coefficient image can also be generated by performing an allocation processing of known absorption coefficients corresponding to the tissue with the largest tissue composition ratio in each pixel.

[0178] Furthermore, in the second variation of the above embodiment, an example of three machine learning models—a machine learning model for axial sections, a machine learning model for coronal sections, and a machine learning model for sagittal sections—was described, but the present invention is not limited thereto. In the present invention, the machine learning model may also include any two of the machine learning models for axial sections, coronal sections, and sagittal sections.

[0179] Furthermore, in the fourth variation of the above embodiment, an example was described where both intermediate images and reconstructed images were used as the output of the machine learning model; however, the present invention is not limited to this. In the present invention, three or more images may be used as the output of the machine learning model. Additionally, intermediate images and images other than reconstructed images may be used as the output of the machine learning model. For example, intermediate images and mixed tissue number images representing the number of tissue types contained in pixels may be used as the output of the machine learning model.

[0180] Furthermore, in the above embodiments, for ease of explanation, a "flow-driven" flowchart was used to describe each process of the processing circuit, but the present invention is not limited thereto. In the present invention, it can also be described using an "event-driven" approach, where each of the above processes is executed on an event-by-event basis. In this case, it can be described using a completely event-driven approach, or it can be described using a combination of event-driven and flow-driven approaches.

[0181] [Way]

[0182] Those skilled in the art will understand that the exemplary embodiments described above are specific examples of the following approaches.

[0183] (Project 1)

[0184] An absorption coefficient image generation method is provided for a nuclear medicine diagnostic device used to generate absorption coefficient images within a subject. The absorption coefficient image generation method includes the following steps:

[0185] An input image is generated by image processing of measurement data obtained from the detection of radiation emitted from the subject;

[0186] Based on the input image, an intermediate image is generated, including images related to the tissue region; and

[0187] An absorption coefficient image is generated based on the known absorption coefficients of the intermediate image and the tissue region.

[0188] (Project 2)

[0189] According to the absorption coefficient image generation method described in Project 1, where,

[0190] The intermediate image includes at least one of a tissue composition ratio image representing the proportion of tissue contained in each pixel and a tissue label image representing the tissue to which each pixel belongs, as an image related to the tissue region.

[0191] (Project 3)

[0192] According to the absorption coefficient image generation method described in Project 2, where,

[0193] The step of generating the absorption coefficient image includes the following steps: assigning absorption coefficients to tissues in the tissue composition ratio image or tissues in the tissue label image based on known absorption coefficients.

[0194] (Project 4)

[0195] According to the absorption coefficient image generation method described in Project 1, where,

[0196] The step of generating the input image includes the following steps: generating the input image without performing at least one of absorption correction processing and scattering correction processing.

[0197] (Project 5)

[0198] According to the absorption coefficient image generation method described in Project 1, where,

[0199] The step of generating the input image includes the following steps: processing the measurement data, including back projection processing.

[0200] (Project 6)

[0201] According to the absorption coefficient image generation method described in Project 1, where,

[0202] The input image includes at least one of the following: an image obtained by applying image quality conversion processing to the image-processed measurement data without applying image quality conversion processing; an image obtained by applying image quality conversion processing to the image-processed measurement data; and an image obtained by applying region recognition processing to the image-processed measurement data.

[0203] (Project 7)

[0204] According to the absorption coefficient image generation method described in Project 1, where,

[0205] The input image includes images with two or more resolutions.

[0206] (Project 8)

[0207] According to the absorption coefficient image generation method described in Project 1, where,

[0208] The steps for generating the intermediate image include the following: applying a pre-learned machine learning model to the input image.

[0209] (Project 9)

[0210] According to the absorption coefficient image generation method described in Project 8, among which,

[0211] The input image used as learning data for the machine learning model includes at least one of the following images: a normalized image whose pixel value range has been normalized, an image obtained by multiplying the normalized image by a coefficient greater than 0 and less than 1, and an image obtained by multiplying a specific region of the normalized image or the image before normalization by a positive coefficient.

[0212] (Project 10)

[0213] According to the absorption coefficient image generation method described in Project 8, among which,

[0214] In addition to outputting the intermediate image, the machine learning model also outputs a reconstructed image obtained by applying at least one of absorption correction and scattering correction processing.

[0215] (Project 11)

[0216] According to the absorption coefficient image generation method described in Project 8, among which,

[0217] The machine learning model includes at least one of the following machine learning models:

[0218] Machine learning models that take 3D images as input;

[0219] A machine learning model that takes axial cross-sectional images as input;

[0220] Machine learning models that take coronal cross-sectional images as input;

[0221] A machine learning model that takes sagittal cross-sectional images as input;

[0222] A machine learning model that takes small image patches cut out from a 3D image as input;

[0223] A machine learning model that takes small images cut out from axial cross-sectional images as input;

[0224] A machine learning model that takes small image patches cut from coronal cross-sectional images as input; and

[0225] A machine learning model that takes small image patches cut out from a sagittal cross-sectional image as input.

[0226] (Project 12)

[0227] According to the absorption coefficient image generation method described in Project 8, among which,

[0228] In addition to the input image, the machine learning model also takes information related to the spatial location of the input image as input.

[0229] (Project 13)

[0230] According to the absorption coefficient image generation method described in Project 8, among which,

[0231] The machine learning model includes a deep neural network.

[0232] (Project 14)

[0233] According to the absorption coefficient image generation method described in Project 13, where,

[0234] The deep neural network includes convolution processing.

[0235] (Project 15)

[0236] According to the absorption coefficient image generation method described in Project 8, among which,

[0237] The machine learning model learns using pseudo-images generated by at least one of Monte Carlo simulation and analytical simulation.

[0238] (Project 16)

[0239] According to the absorption coefficient image generation method described in Project 15, where,

[0240] The machine learning model uses both the pseudo-images and images of the actual subjects to learn.

[0241] (Project 17)

[0242] According to the absorption coefficient image generation method described in Project 16, among which,

[0243] The machine learning model uses the machine learning model learned from the pseudo-images as a base model to perform additional learning using images of the actual subjects.

[0244] (Project 18)

[0245] According to the absorption coefficient image generation method described in Project 3, where,

[0246] The step of generating the absorption coefficient image includes the following steps: when the intermediate image includes the tissue composition ratio image, performing linear combination processing on the tissue composition ratio images of each tissue with known absorption coefficients set as coefficients.

[0247] The step of generating the absorption coefficient image includes the following steps: when the intermediate image includes the tissue label image, performing an assignment process of known absorption coefficients corresponding to the label values ​​of the tissue label image.

[0248] (Project 19)

[0249] According to the absorption coefficient image generation method described in Project 3, where,

[0250] The step of generating the absorption coefficient image includes the following steps: when the intermediate image includes the tissue label image, performing linear combination processing on a reliability image that is an intermediate output of the tissue label image, with the known absorption coefficients set as coefficients.

[0251] (Project 20)

[0252] According to the absorption coefficient image generation method described in Project 1, where,

[0253] The measurement data mentioned are measurements of a human head.

[0254] The elements that constitute an image related to a tissue region include at least one of the following: background, cavities, soft tissue, and bone.

[0255] (Project 21)

[0256] According to the absorption coefficient image generation method described in Project 1, where,

[0257] The measurement data mentioned are measurements of human breast tissue.

[0258] The elements that constitute an image related to an organized region include at least one of the background and soft tissue.

[0259] (Project 22)

[0260] A nuclear medicine diagnostic device, comprising:

[0261] The detection department detects radiation generated by radioactive agents within the body of the subject; and

[0262] The processing unit generates an image of the radioactivity distribution within the subject based on the detection of radiation by the detection unit.

[0263] The processing unit is configured as follows:

[0264] An input image is generated by image processing of measurement data obtained from the detection of radiation emitted from the subject;

[0265] Based on the input image, an intermediate image is generated that includes images related to the tissue region;

[0266] An absorption coefficient image is generated based on the known absorption coefficients of the intermediate image and the tissue region to generate the radioactivity distribution image.

[0267] (Project 23)

[0268] According to the nuclear medicine diagnostic device described in Project 22, among which,

[0269] The processing unit is configured to perform at least one of absorption correction processing and scattering correction processing based on the absorption coefficient image.

[0270] (Project 24)

[0271] A method for creating a learned model, which is used in nuclear medicine diagnostic devices, includes the following steps:

[0272] Prepare tissue label images representing the tissue to which each pixel belongs;

[0273] Based on the tissue label image, a pseudo-radioactivity distribution image and a pseudo-absorption coefficient image are created;

[0274] Pseudo-measurement data are generated by performing simulation calculations based on the pseudo-radioactivity distribution image and the pseudo-absorption coefficient image.

[0275] A pseudo-image is generated by image processing of the pseudo-measurement data; and

[0276] The pseudo-images are used as learning data to create a fully learned model.

[0277] Explanation of reference numerals in the attached figures

[0278] 1: PET device (nuclear medicine diagnostic device); 2: Detector ring (detection unit); 5: Measurement data; 6, 6a: Input image; 7, 7a: Intermediate image; 8: Machine learning model; 9: Absorption coefficient image; 10: Radioactivity distribution image; 11: Tissue tag image; 12: Pseudo-radioactivity distribution image; 13: Pseudo-absorption coefficient image; 15: Pseudo-measurement data; 16: Pseudo-reconstructed image (pseudo-image); 20: Information related to spatial location; 21: Reconstructed image; 41: Processing circuit (processing unit); 71: Tissue composition ratio image; 72: Tissue tag image; 100: Subject.

Claims

1. A method for generating an absorption coefficient image, which is a method for generating an absorption coefficient image of a subject in a nuclear medicine diagnostic device, the method comprising the following steps: An input image is generated by image processing of measurement data obtained from the detection of radiation emitted from the subject; Based on the input image, generate multiple tissue composition ratio images representing the proportion of multiple tissues contained in each pixel, or multiple reliability images representing the probability that each pixel belongs to which of the multiple tissues. as well as An absorption coefficient image is generated by weighted summation of the images after multiplying each image in a plurality of tissue composition ratio images or a plurality of reliability images by a known absorption coefficient of the tissue region. In the step of generating multiple tissue composition ratio images or multiple reliability images, the input image is input into a machine learning model, and the output of the machine learning model is used to generate multiple tissue composition ratio images or multiple reliability images.

2. The absorption coefficient image generation method according to claim 1, wherein, The step of generating the absorption coefficient image includes the following steps: assigning absorption coefficients to tissues in multiple tissue composition ratio images or multiple reliability images based on known absorption coefficients.

3. The absorption coefficient image generation method according to claim 1, wherein, The step of generating the input image includes the following steps: generating the input image without performing at least one of absorption correction processing and scattering correction processing.

4. The absorption coefficient image generation method according to claim 1, wherein, The step of generating the input image includes the following steps: processing the measurement data, including back projection processing.

5. The absorption coefficient image generation method according to claim 1, wherein, The input image includes at least one of the following: an image obtained by applying image quality conversion processing to the image-processed measurement data without applying image quality conversion processing; an image obtained by applying image quality conversion processing to the image-processed measurement data; and an image obtained by applying region recognition processing to the image-processed measurement data.

6. The absorption coefficient image generation method according to claim 1, wherein, The input image includes images with two or more resolutions.

7. The absorption coefficient image generation method according to claim 1, wherein, The input image used as learning data for the machine learning model includes at least one of the following images: a normalized image whose pixel value range has been normalized, an image obtained by multiplying the normalized image by a coefficient greater than 0 and less than 1, and an image obtained by multiplying a specific region of the normalized image or the image before normalization by a positive coefficient.

8. The absorption coefficient image generation method according to claim 1, wherein, In addition to outputting the tissue composition ratio image or the reliability image, the machine learning model also outputs a reconstructed image obtained by applying at least one of absorption correction processing and scattering correction processing.

9. The absorption coefficient image generation method according to claim 1, wherein, The machine learning model includes at least one of the following machine learning models: Machine learning models that take 3D images as input; A machine learning model that takes axial cross-sectional images as input; Machine learning models that take coronal cross-sectional images as input; A machine learning model that takes sagittal cross-sectional images as input; A machine learning model that takes small image patches cut out from a 3D image as input; A machine learning model that takes small images cut out from axial cross-sectional images as input; A machine learning model that takes small image patches cut from coronal cross-sectional images as input; and A machine learning model that takes small image patches cut out from a sagittal cross-sectional image as input.

10. The absorption coefficient image generation method according to claim 1, wherein, In addition to the input image, the machine learning model also takes information related to the spatial location of the input image as input.

11. The absorption coefficient image generation method according to claim 1, wherein, The machine learning model includes a deep neural network.

12. The absorption coefficient image generation method according to claim 11, wherein, The deep neural network includes convolution processing.

13. The absorption coefficient image generation method according to claim 1, wherein, The machine learning model learns using pseudo-images generated by at least one of Monte Carlo simulation and analytical simulation.

14. The absorption coefficient image generation method according to claim 13, wherein, The machine learning model uses both the pseudo-images and images of the actual subjects to learn.

15. The absorption coefficient image generation method according to claim 14, wherein, The machine learning model uses the machine learning model learned from the pseudo-images as a base model to perform additional learning using images of the actual subjects.

16. The absorption coefficient image generation method according to claim 2, wherein, The step of generating the absorption coefficient image includes the following steps: when generating multiple tissue composition ratio images, performing linear combination processing of the tissue composition ratio images of each tissue with known absorption coefficients set as coefficients.

17. The absorption coefficient image generation method according to claim 2, wherein, The step of generating the absorption coefficient image includes the following steps: when generating multiple reliability images, performing linear combination processing of the reliability images, which are intermediate outputs of tissue label images and have known absorption coefficients set as coefficients.

18. The absorption coefficient image generation method according to claim 1, wherein, The measurement data mentioned are measurements of a human head. The elements constituting the tissue composition ratio image or the reliability image include at least one of the following: background, cavities, soft tissue, and bone.

19. The absorption coefficient image generation method according to claim 1, wherein, The measurement data mentioned are measurements of human breast tissue. The elements constituting the tissue composition ratio image or the reliability image include at least one of the background and soft tissue.

20. A nuclear medicine diagnostic device, comprising: The detection department detects radiation generated by radioactive agents within the body of the subject; and The processing unit generates an image of the radioactivity distribution within the subject based on the detection of radiation by the detection unit. in, The processing unit is configured as follows: An input image is generated by image processing of measurement data obtained from the detection of radiation emitted from the subject; Based on the input image, generate multiple tissue composition ratio images representing the proportion of multiple tissues contained in each pixel, or multiple reliability images representing the probability that each pixel belongs to which of the multiple tissues. An absorption coefficient image for generating the radioactivity distribution image is generated by a weighted summation of the images obtained by multiplying each image in a plurality of tissue composition ratio images or a plurality of reliability images by a known absorption coefficient of the tissue region. The processing unit is further configured to: input the input image into a machine learning model, and generate multiple tissue composition ratio images or multiple reliability images through the output of the machine learning model.

21. The nuclear medicine diagnostic device according to claim 20, wherein, The processing unit is configured to perform at least one of absorption correction processing and scattering correction processing based on the absorption coefficient image.

22. A method for fabricating a learned model, used in a nuclear medicine diagnostic device, the method comprising the following steps: Prepare tissue label images representing the tissue to which each pixel belongs; Based on the tissue label image, a pseudo-radioactivity distribution image and a pseudo-absorption coefficient image are created; Pseudo-measurement data are generated by performing simulation calculations based on the pseudo-radioactivity distribution image and the pseudo-absorption coefficient image. A pseudo-image is generated by image processing of the pseudo-measurement data; as well as The pseudo-images are used as learning data to create a fully learned model.

Images