Image processing apparatus and image processing method

Abstract

To provide a device and a method capable of creating a tomographic image in which noise is reduced on the basis of list data collected by a radiation tomography apparatus.SOLUTION: An image processing method repeatedly performs plural times of a reconstruction step S1, a CNN processing step S2 and an update step S3 from a certain initial state to create a tomographic image of a subject. In the reconstruction step S1, a first image is created by performing processing by a list mode iterative approximation reconstruction method. In the CNN processing step S2, a second image is created with a CNN by inputting input information to the CNN by a DIP technique to have the CNN learned. In the updating step S3, a third image is updated on the basis of the first image and the second image.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to an apparatus and method for creating tomographic images based on list data collected by a radiographic tomography apparatus. [Background technology]

[0002] Examples of radiographic tomography devices that can acquire cross-sectional images of a subject (living organism) include PET (Positron Emission Tomography) devices and SPECT (Single Photon Emission Computed Tomography) devices.

[0003] A PET scanner has a detection unit equipped with numerous small radiation detectors arranged around the measurement space where the subject is placed. The PET scanner detects photon pairs with an energy of 511 keV generated by electron-positron annihilation within the subject into which a positron-emitting isotope (RI source) is introduced, using the coincidence counting method, and collects this coincidence counting information. Based on this collected information, a tomographic image representing the spatial distribution of the frequency of photon pair generation in the measurement space (i.e., the spatial distribution of the RI source) can be reconstructed. At this time, a list data of the coincidence counting information collected by the PET scanner arranged in chronological order is divided into multiple frames in the order of collection, and by performing image reconstruction processing using the data group contained in each frame of the list data, a dynamic PET image consisting of tomographic images from multiple frames can be obtained. This PET scanner plays an important role in fields such as nuclear medicine, and can be used to study, for example, biological functions and higher brain functions.

[0004] The tomographic image thus reconstructed contains a lot of noise, so noise removal processing using an image filter is necessary. Examples of image filters used for noise removal include the Gaussian filter and the guided filter. Conventionally, the Gaussian filter has been used. In contrast, the guided filter was developed in recent years and has the feature of being able to better preserve the intensity boundaries in the image compared to the Gaussian filter.

[0005] Also, a technique for removing noise from tomographic images using the Deep Image Prior technique that uses a convolutional neural network, which is a type of deep neural network, has been proposed (Non-Patent Document 1). Hereinafter, a deep neural network will be referred to as "DNN", a convolutional neural network will be referred to as "CNN", and the Deep Image Prior technique will be referred to as "DIP technique". The DIP technique can reduce the noise of the target image by utilizing the property of the CNN that meaningful structures in the target image are learned faster than random noise (that is, random noise is difficult to learn).

[0006] These noise removal techniques process the tomographic image to reduce noise after creating the tomographic image by the histogram mode reconstruction method using list data. Alternatively, the noise removal technique may be incorporated as regularization into the histogram mode reconstruction method. In the histogram mode reconstruction method, based on the list data, a histogram representing the number of coincidence events detected by each detector pair is created, and the tomographic image is reconstructed based on this histogram. As the format of the histogram, for example, a four-dimensional array (three-dimensional sinogram) of radius × body axis × azimuth angle × tilt angle is used.

[0007] By the way, with the recent evolution of radiation tomography technology, when new information such as the detection time difference information (Time of Flight, TOF) of a pair of radiation detectors and the depth of photon interaction (Depth of Interaction, DOI) in the detector is added to the list data, the array format of the histogram has become as high as five or six dimensions. As a result, in the histogram mode reconstruction method that reconstructs a tomographic image from list data through a histogram, the load on the arithmetic unit including a CPU that performs these series of processes has become excessive.

[0008] As a technology that can address the problems of such a histogram mode reconstruction method, a list mode successive approximation reconstruction method has been proposed (Non-Patent Document 2). In the list mode successive approximation reconstruction method, a tomographic image is reconstructed by directly repeating successive approximations from the list data (without going through a histogram).

Prior Art Documents

Non-Patent Documents

[0009]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0010] Many denoising techniques have been researched and developed for histogram mode reconstruction. However, list mode iterative reconstruction has the problem that it is difficult to process list data directly with a CNN, and no list mode iterative reconstruction technique that can effectively remove noise by incorporating a CNN is known.

[0011] The present invention was made to solve the above problems, and aims to provide an image processing apparatus and an image processing method that can create a tomographic image with reduced noise based on list data collected by a radiographic tomography apparatus. [Means for solving the problem]

[0012] The present invention relates to an image processing device that creates a tomographic image based on list data collected by a radiographic tomography device.

[0013] image The processing unit comprises: (1) a reconstruction unit that creates a new first image by repeatedly updating the first image based on list data using a list-mode successive reconstruction method and bringing the resulting image closer to the difference between the second and third images; (2) a CNN processing unit that inputs input information to a convolutional neural network to create a second image using the convolutional neural network and trains the convolutional neural network so that the created second image approaches the sum of the first and third images; and (3) an update unit that updates the third image based on the first and second images. It's fine. .this Image processing The device starts with the training state of the convolutional neural network and the initial states of the first, second, and third images, and repeatedly performs the following: the reconstruction unit creates the first image, the CNN processing unit creates the second image and trains the convolutional neural network, and the update unit updates the third image. The device then converts either the first or second image obtained through these iterative processes into a tomographic image. You may .

[0014] This invention The picture The image processing device comprises: (1) a reconstruction unit that creates a first image by updating a third image based on list data using a list-mode successive reconstruction method; (2) a CNN processing unit that inputs input information to a convolutional neural network to create a second image using the convolutional neural network, and trains the convolutional neural network so that the created second image approaches the third image; and (3) an update unit that updates the third image based on the first and second images. Image processing The device starts from the initial state of the convolutional neural network's training state and the initial state of the third image, and repeatedly performs the following: the reconstruction unit creates the first image, the CNN processing unit creates the second image and trains the convolutional neural network, and the update unit updates the third image. One of the first, second, or third images obtained through these iterative processes is used as the tomographic image.

[0015] The CNN processing unit of the image processing device of the present invention may be input to the convolutional neural network with an image representing the morphological information of the subject as input information, or with an MRI or CT image of the subject as input information, or with a random noise image as input information.

[0016] The radiographic tomography system of the present invention comprises a radiographic tomography apparatus that collects list data for reconstructing a tomographic image of a subject, and the image processing apparatus of the present invention described above that creates a tomographic image based on the list data collected by the radiographic tomography apparatus.

[0017] The present invention relates to an image processing method for creating a tomographic image based on list data collected by a radiographic tomography device.

[0018] imageThe processing method comprises: (1) a reconstruction step in which a new first image is created by repeatedly updating the first image based on list data using a list-mode successive reconstruction method and bringing the resulting image closer to the difference between the second and third images; (2) a CNN processing step in which input information is fed into a convolutional neural network to create a second image using the convolutional neural network, and the convolutional neural network is trained so that the created second image approaches the sum of the first and third images; and (3) an update step in which the third image is updated based on the first and second images. It's fine. . this The image processing method starts with the training state of the convolutional neural network and the initial states of the first, second, and third images, and repeatedly performs the following steps: creating the first image in the reconstruction step, creating the second image and training the convolutional neural network in the CNN processing step, and updating the third image in the update step. Either the first or second image obtained through these iterative processes is then used as the tomographic image. You may .

[0019] This invention The picture The image processing method comprises: (1) a reconstruction step of creating a first image by updating a third image based on list data using a list-mode successive reconstruction method; (2) a CNN processing step of inputting input information to a convolutional neural network to create a second image using the convolutional neural network, and training the convolutional neural network so that the created second image approaches the third image; and (3) an update step of updating the third image based on the first and second images. this The image processing method starts with the initial state of the convolutional neural network's training state and the initial state of the third image, and repeatedly performs the following steps: creating the first image in the reconstruction step, creating the second image and training the convolutional neural network in the CNN processing step, and updating the third image in the update step. One of the first, second, or third images obtained through these iterative processes is used as the tomographic image.

[0020] In the CNN processing step of the image processing method of the present invention, an image representing the morphological information of a subject may be input to the convolutional neural network as input information, or an MRI or CT image of the subject may be input to the convolutional neural network as input information, or a random noise image may be input to the convolutional neural network as input information. [Effects of the Invention]

[0021] According to the present invention, noise-reduced tomographic images can be created based on list data collected by a radiographic tomography device. [Brief explanation of the drawing]

[0022] [Figure 1] Figure 1 shows the configuration of the radiographic tomography system 1. [Figure 2] Figure 2 is a flowchart of the image processing method. [Figure 3] Figure 3 shows the sequence of the image processing method according to the first embodiment. [Figure 4] Figure 4 shows the phantom image (ground truth image). Figures 4(a) and 4(b) are cross-sectional images. [Figure 5] Figure 5 shows the phantom image (ground truth image). Figure 5(a) is a coronal section image, and Figure 5(b) is a sagittal section image. [Figure 6] Figure 6 shows a tomographic image obtained using the image processing method of Comparative Example 1. Figures 6(a) and 6(b) are cross-sectional images. [Figure 7] Figure 7 shows tomographic images obtained using the image processing method of Comparative Example 1. Figure 7(a) is a coronal section image, and Figure 7(b) is a sagittal section image. [Figure 8] Figure 8 shows a tomographic image obtained using the image processing method of Comparative Example 2. Figures 8(a) and 8(b) are cross-sectional images. [Figure 9]Figure 9 shows tomographic images obtained using the image processing method of Comparative Example 2. Figure 9(a) is a coronal section image, and Figure 9(b) is a sagittal section image. [Figure 10] Figure 10 shows tomographic images obtained by the image processing method of the embodiment. Figures 10(a) and 10(b) are cross-sectional images. [Figure 11] Figure 11 shows tomographic images obtained by the image processing method of the embodiment. Figure 11(a) is a coronal section image, and Figure 11(b) is a sagittal section image. [Figure 12] Figure 12 is a graph showing the PSNR of the tomographic images for Comparative Examples 1 and 2 and the Example. [Figure 13] Figure 13 is a graph showing the CRC of tomographic images for Comparative Examples 1 and 2 and the Example. [Figure 14] Figure 14 shows the sequence of the image processing method according to the second embodiment. [Modes for carrying out the invention]

[0023] Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the attached drawings. In the description of the drawings, the same elements will be denoted by the same reference numerals, and redundant descriptions will be omitted. The present invention is not limited to these examples, but is indicated by the claims, and all modifications within the meaning and scope equivalent to the claims are intended to be included.

[0024] Figure 1 shows the configuration of the radiographic tomography system 1. The radiographic tomography system 1 comprises a radiographic imaging device 2 and an image processing device 10. The image processing device 10 comprises a reconstruction unit 11, a CNN processing unit 12, an update unit 13, and a storage unit 14. A computer having a CPU, RAM, ROM, and a hard disk drive is used as the image processing device 10. The image processing device 10 also includes an input unit (e.g., keyboard and mouse) for receiving input from the operator and a display unit (e.g., liquid crystal display) for displaying images, etc.

[0025] The tomography system 2 is a device that collects list data for reconstructing tomographic images of the subject. Examples of tomography systems 2 include PET scanners and SPECT scanners. In the following explanation, we will assume that the tomography system 2 is a PET scanner.

[0026] The tomography apparatus 2 is equipped with a detection unit having numerous small radiation detectors arranged around the measurement space where the subject is placed. The tomography apparatus 2 detects photon pairs with an energy of 511 keV generated by electron-positron annihilation within the subject into which a positron-emitting isotope (RI source) is introduced, using the coincidence counting method, and stores this coincidence counting information. The tomography apparatus 2 then outputs this large amount of stored coincidence counting information, arranged in chronological order, as list data to the image processing device 10.

[0027] The list data includes identification information and detection time information for a pair of radiation detectors that simultaneously counted photon pairs. The list data may also include detection time difference information (TOF information) for the pair of radiation detectors, photon interaction depth information (DOI information) for the radiation detectors, and energy information for the photons detected by the radiation detectors, etc.

[0028] The memory unit 14 stores list data collected by the tomography device 2. The memory unit 14 also stores programs for executing processing by the reconstruction unit 11, the CNN processing unit 12, and the update unit 13, respectively. The reconstruction unit 11, the CNN processing unit 12, and the update unit 13 use the programs and list data stored in the memory unit 14 to create tomographic images of the subject.

[0029] The reconstruction unit 11 creates a first image by performing processing based on the list-mode iterative reconstruction method (see Non-Patent Document 2). The processing based on the list-mode iterative reconstruction method in the reconstruction unit 11 uses iterative update formulas such as LM-MLEM (Maximum Likelihood Expectation Maximization), LM-OSEM (Ordered Subset EM), and LM-DRAMA (Dynamic Row Action Maximum Likelihood Algorithm).

[0030] The CNN processing unit 12 creates a second image by performing processing based on DIP technology (see Non-Patent Literature 1). In the DIP technology processing in the CNN processing unit 12, input information is input to the CNN, which then creates the second image and trains the CNN. The input information to be input to the CNN may be morphological information of the subject, an MRI or CT image of the subject, or a random noise image.

[0031] The update unit 13 updates the third image based on the first and second images. Details of the processes of the reconstruction unit 11, the CNN processing unit 12, and the update unit 13 will be described later.

[0032] The memory unit 14 also stores the input information that is input to the CNN, and also stores the first image, the second image, and the third image. The image processing device 10 starts from a certain initial state and repeatedly performs the processes of the reconstruction unit 11, the CNN processing unit 12, and the update unit 13 to create a tomographic image of the subject.

[0033] Figure 2 is a flowchart of the image processing method. The image processing method comprises a reconstruction step S1 performed by the reconstruction unit 11, a CNN processing step S2 performed by the CNN processing unit 12, and an update step S3 performed by the update unit 13. Starting from an initial state, the reconstruction step S1, CNN processing step S2, and update step S3 are repeated multiple times (N times) to create a tomographic image of the subject.

[0034] In step S4, the value of parameter n is set to its initial value of 0. In the following step S5, the value of parameter n is increased by 1. After step S5, the reconstruction step S1, CNN processing step S2, and update step S3 are performed. In the following step S6, the value of parameter n is compared with N, and if it is determined that n is less than N, the process returns to step S5. If it is determined in step S6 that n has reached N, the iteration is terminated and a tomographic image of the subject is obtained. Hereinafter, the nth reconstruction step S1 out of the N iterations will be referred to as reconstruction step S1(n), the nth CNN processing step S2 as CNN processing step S2(n), and the nth update step S3 as update step S3(n). n is an integer between 1 and N, inclusive.

[0035] Next, the details of the image processing device 10 and the image processing method will be described. First, the list data U is formulated as shown in equation (1) below. t is a number representing a simultaneous counting event. T is the total number of events. i(t) represents a number that identifies the detector pair that detected the t-th event.

[0036]

number

[0037] For this list data U, we consider the constrained optimization problem expressed in equation (2) below. x is a tomographic image. L(U│x) is the likelihood representing the probability that the list data U is observed from the tomographic image x. z is the input information input to the CNN. θ is a parameter representing the learning state of the CNN, such as the connection weights, and changes as the CNN's learning progresses. θ (z) is the image output from a CNN when input information z is input to a CNN whose learning state is θ. This constrained optimization problem of equation (2) is when the tomographic image x is the CNN output image f θ The constraint is that (z) is (x=f θ Under (z), the problem is to optimize the tomographic image x and CNN parameters θ so that the likelihood L(U│x) is high.

[0038]

number

[0039] There are two methods for solving this optimization problem, as described below. The first method is based on the Alternating Direction Method of Multipliers (ADMM method). The second method is based on Forward Backward Splitting (FBS method). The FBS method also includes the De Pierro method as a special variant. The details of the image processing methods for the first and second methods will be described below.

[0040] In the first embodiment of the image processing method, the constrained optimization problem in equation (2) above is rewritten based on the extended Lagrangian function method and then solved using the ADMM method. In the extended Lagrangian function method, the constraints in equation (2) above are replaced with regularization terms, and the constrained optimization problem in equation (2) above is rewritten as the unconstrained optimization problem in equation (3) below. ρ is a positive constant that adjusts the strength of regularization. μ is called the Lagrangian multiplier or dual variable, and in the following explanation it will be referred to as the "third image".

[0041]

number

[0042] In the first embodiment, ADMM solves the unconstrained optimization problem of equation (3) by repeatedly performing the processes of equations (4) to (6) below.

[0043]

number

[0044]

number

[0045]

Number

[0046] Figure 3 is a diagram showing the sequence of the image processing method of the first aspect. Prior to the iterative process, the learning state θ of the CNN (0) , the first image x (0) , the second image f θ (0) (z), and the third image μ (0) are each initialized. Note that the second image f θ (0) (z) is an image output from the CNN when input information z is input to the CNN in the initial learning state θ (0) .

[0047] In the n-th reconstruction step S1(n), according to the above formula (4), the image obtained by updating the first image x (n-1) once based on the list data U by the list mode successive approximation reconstruction method is made to approach the difference (f θ (n-1) (z) - μ (n-1) ) between (z) and the third image μ θ (n-1) by repeating the process, and a new first image x (n-1) is created. (n)

[0048] In the n-th CNN processing step S2(n), according to the above formula (5), input information z is input to the CNN to create the second image f θ (n) (z) by the CNN, and the CNN is trained so that the created second image f θ (n) (z) approaches the sum (x (n) + μ (n-1) ) of the first image x (n) and the third image μ (n-1) . The learning state of the CNN after this training is denoted as θ (n) .

[0049] ​In the nth update step S3(n), according to equation (6) above, the first image x (n) and the second image f θ (n) The difference with (z) (x (n) -f θ (n) (z)) is the third image μ (n-1) By adding this, the third image μ (n) Update to [date / time].

[0050] After the reconstruction step S1, CNN processing step S2, and update step S3 are repeated N times, the resulting first image x (N) and second image f θ (N) (z) is used as the tomographic image of the subject.

[0051] In the first embodiment, the first image x and the second image f θ Instead of optimizing (z) simultaneously, we optimize the first image x and the second image f θ Since (z) is optimized alternately, the problem is easy to solve. Furthermore, the reconstruction step S1 by the reconstruction unit 11 and the CNN processing step S2 by the CNN processing unit 12 can be performed using conventional methods as described in Non-Patent Documents 1 and 2, making implementation easy.

[0052] Figures 4 to 11 show the results of simulations performed to confirm the effectiveness of the first embodiment of the image processing method. In this simulation, simulation data was created using MC simulation of a head PET device with digital brain phantom images, and this data was used to confirm the effectiveness of the first embodiment of the image processing method.

[0053] Figures 4 and 5 show phantom images (ground truth images). Figures 4(a) and 4(b) are cross-sectional images, Figure 5(a) is a coronal section image, and Figure 5(b) is a sagittal section image.

[0054] Figures 6 and 7 show tomographic images obtained using the image processing method of Comparative Example 1. Figures 6(a) and 6(b) are cross-sectional images, Figure 7(a) is a coronal section image, and Figure 7(b) is a sagittal section image. The tomographic images of Comparative Example 1 were reconstructed using only the successive update formula of the LM-DRAMA list-mode iterative reconstruction method, and no noise reduction processing was performed.

[0055] Figures 8 and 9 show tomographic images obtained using the image processing method of Comparative Example 2. Figures 8(a) and 8(b) are cross-sectional images, Figure 9(a) is a coronal section image, and Figure 9(b) is a sagittal section image. The tomographic images of Comparative Example 2 were obtained by reducing noise in the reconstructed tomographic images of Comparative Example 1 using DIP technology. The number of CNN parameter updates was set to 20. The input information to the CNN was MRI images.

[0056] Figures 10 and 11 show tomographic images obtained by the image processing method of the embodiment. Figures 10(a) and 10(b) are cross-sectional images, Figure 11(a) is a coronal section image, and Figure 11(b) is a sagittal section image. The tomographic images of the embodiment were created by the image processing method of the first embodiment. ρ = 0.05 was set. The input information z to the CNN was an MRI image. The number of iterations in the reconstruction step S1 was set to 2, the number of iterations of CNN training in the CNN processing step S2 was set to 20, and the total number of iterations N was set to 200.

[0057] As can be seen by comparing these tomographic images, the tomographic images obtained by the image processing method of the example have less noise and the structure of the cerebral cortex is well reconstructed.

[0058] Figure 12 is a graph showing the PSNR of the tomographic images for Comparative Examples 1 and 2 and the Example. PSNR is the peak signal-to-noise ratio [unit: dB], which is an indicator of noise. Figure 13 is a graph showing the CRC of the tomographic images for Comparative Examples 1 and 2 and the Example. CRC is the contrast recovery coefficient of the tumor, which is an indicator of quantitative accuracy. The tomographic images of the Example had higher PSNR and CRC compared to the tomographic images of Comparative Examples 1 and 2, and the CRC was close to the ideal value of 1.0. As can be seen from these results, the image processing method of the Example makes it possible to generate highly quantitative tomographic images while suppressing the increase in noise artifacts.

[0059] Next, the second embodiment of the image processing method will be described. In the second embodiment of the image processing method, the constrained optimization problem of equation (2) above is solved by the FBS method based on the framework of the Maximum Apostilio (MAP) estimation method. In the MAP estimation method, the constrained optimization problem of equation (2) above is rewritten as the unconstrained optimization problem of equation (7) below.

[0060]

number

[0061] In the second embodiment, the unconstrained optimization problem of equation (7) is solved by repeatedly performing the processes of equations (8) to (11) below using FBS.

[0062]

number

[0063]

number

[0064]

number

[0065]

number

[0066] Here, γ is a pre-defined parameter image. ω is an image representing the sensitivity of the detector for each pixel. p is the probability that a pair of gamma rays emitted from pixel j is detected by detector pair i. ij Therefore, the sensitivity image ω is expressed by equation (12) below.

[0067]

number

[0068] Figure 14 shows the sequence of the image processing method in the second embodiment. Prior to the iterative processing, the learning state θ of the CNN is... (0) and third image x (0) Initialize each of them.

[0069] In the nth reconstruction step S1(n), a third image x is generated based on the list data U by the list-mode iterative reconstruction method according to equation (8) above. (n-1) By updating the first image x, ML (n) Create.

[0070] In the nth CNN processing step S2(n), the input information z is input to the CNN according to equation (9) above, and the CNN processes the second image f θ (n) Create (z), and this created second image f θ (n) (z) is the third image x (n-1) The CNN is trained to approach this state. The trained state of the CNN after this training is θ. (n) Let's assume that.

[0071] In the nth update step S3(n), according to equations (10) and (11) above, the first image x ML (n) and second image f θ (n) Based on (z), the third image x(n-1) The third image x (n) Update to [date / time].

[0072] After the reconstruction step S1, CNN processing step S2, and update step S3 are repeated N times, the resulting first image x ML (N) , second image f θ (N) (z) and third image x (N) One of the following will be used as the tomographic image of the subject.

[0073] In the second embodiment, the reconstruction step S1(n) and the CNN processing step S2(n) may be performed in any order, or they may be performed in parallel.

[0074] The De Pierro method is equivalent to the FBS method with ρ=1 and γ=1 / ω. That is, in the De Pierro method, equations (13) and (14) below are used instead of equations (10) and (11) above. The processing contents of the reconstruction step S1, CNN processing step S2, and update step S3 are the same as above.

[0075]

number

[0076]

number

[0077] In the second aspect, the first image x ML and second image f θ Instead of simultaneously optimizing (z), the first image x ML and second image f θSince (z) is optimized separately, the problem is easy to solve. Furthermore, the reconstruction step S1 by the reconstruction unit 11 and the CNN processing step S2 by the CNN processing unit 12 can be performed using conventional methods as described in Non-Patent Documents 1 and 2, making implementation easy.

[0078] The tomographic images obtained by the image processing method of the second embodiment, like those of the first embodiment, have low noise and the structure of the cerebral cortex is well reconstructed. In the second embodiment as well, it is possible to generate highly quantitative tomographic images while suppressing the increase in noise artifacts.

[0079] The present invention is not limited to the above embodiments, and various modifications are possible. For example, although the radiographic imaging device 2 is a PET device in the above embodiments, it may be a SPECT device. [Explanation of Symbols]

[0080] 1...Radiation tomography system, 2...Radiation tomography device, 10...Image processing device, 11...Reconstruction unit, 12...CNN processing unit, 13...Update unit, 14...Storage unit.

Claims

1. An image processing device that creates a tomographic image based on list data collected by a radiographic tomography device, A reconstruction unit that creates a first image by updating a third image based on the list data using a list-mode iterative reconstruction method, A CNN processing unit that inputs input information to a convolutional neural network to create a second image using the convolutional neural network, and trains the convolutional neural network so that the created second image approaches the third image, An update unit updates the third image based on the first and second images, Equipped with, Starting from the learning state of the convolutional neural network and the initial state of each of the third images, the reconstruction unit creates the first image, the CNN processing unit creates the second image and the convolutional neural network is trained, and the update unit updates the third image, and any of the first image, second image, and third image obtained through these iterative processes is taken as the tomographic image. Image processing device.

2. The CNN processing unit causes the convolutional neural network to input an image representing the morphological information of the subject as input information. The image processing apparatus according to claim 1.

3. The CNN processing unit inputs the MRI image of the subject as input information to the convolutional neural network. The image processing apparatus according to claim 1.

4. The CNN processing unit inputs the CT image of the subject as input information to the convolutional neural network. The image processing apparatus according to claim 1.

5. The CNN processing unit inputs the random noise image as input information to the convolutional neural network. The image processing apparatus according to claim 1.

6. A tomography device that collects list data for reconstructing tomographic images of a subject, An image processing apparatus according to any one of claims 1 to 5, which creates a tomographic image based on list data collected by the aforementioned radiographic tomography apparatus, A radiographic tomography system equipped with [specific features / equipment].

7. An image processing method for creating a tomographic image based on list data collected by a radiographic tomography device, A reconstruction step in which a first image is created by updating a third image based on the list data using a list-mode iterative reconstruction method, A CNN processing step involves inputting input information into a convolutional neural network to create a second image using the convolutional neural network, and training the convolutional neural network so that the created second image approaches the third image. An update step in which the third image is updated based on the first image and the second image, Equipped with, Starting from the learning state of the convolutional neural network and the initial state of each of the third images, the creation of the first image in the reconstruction step, the creation of the second image in the CNN processing step and the learning of the convolutional neural network, and the updating of the third image in the update step are repeated, and any of the first image, second image, and third image obtained through these repeated processes is taken as the tomographic image. Image processing methods.

8. In the CNN processing step, an image representing the morphological information of the subject is input to the convolutional neural network as input information. The image processing method according to claim 7.

9. In the CNN processing step, the MRI image of the subject is input to the convolutional neural network as input information. The image processing method according to claim 7.

10. In the CNN processing step, the CT image of the subject is input to the convolutional neural network as input information. The image processing method according to claim 7.

11. In the CNN processing step, a random noise image is input to the convolutional neural network as input information. The image processing method according to claim 7.

Images