Image processing device and image processing method

Abstract

An image processing device 10 comprises a sinogram generation unit 11, a CNN processing unit 12, a convolution integration unit 13, a forward projection computation unit 14, a CNN training unit 15, a storage unit 16, and an image reconstruction unit 17. The forward projection computation unit 14 performs a forward projection calculation on an output image 23 to generate a computed sinogram 24. In a first processing step, the CNN training unit 15 trains a CNN based on an error between a reconstructed image generated by the image reconstruction unit 17 and the output image 23. In a subsequent second processing step, the CNN training unit 15 trains the CNN based on an error between a measured sinogram 21 and the computed sinogram 24.Thus, an image processing device is realized that is able to reduce the number of CNN training epochs by using a DIP technique and to obtain a tomographic image in a short time in which the noise is reduced when the tomographic image of an object is generated by training a CNN based on an evaluation result of an error between a calculated sinogram and a measured sinogram.

Description

Technical field

[0001] The present invention relates to a device and a method for generating a tomographic image of a subject based on coincidence information collected using a radiation tomography device. State of the art

[0002] Examples of a radiation tomography device capable of capturing a tomographic image of a subject (a living body) include a positron emission tomography (PET) device and a single-photon emission computed tomography (SPECT) device.

[0003] The PET apparatus comprises a detection unit with a large number of small radiation detectors arranged around a measurement chamber in which the subject is placed. The PET apparatus detects a photon pair with an energy of 511 keV, generated by electron-positron annihilation in the subject into which a positron-emitting isotope (an RI source) is injected, using a coincidence method with the detection unit, and collects coincidence information of the detection.

[0004] Furthermore, a tomographic image representing the spatial distribution of the photon pair generation frequency in the measurement space (i.e., a spatial distribution of RI sources) can be reconstructed based on the many pieces of coincidence information collected according to the description above. The PET device mentioned above plays an important role in nuclear medicine and related fields and can be used, for example, to investigate biological or higher brain functions.

[0005] Several methods are known for reconstructing the subject's tomographic image based on the collected numerous pieces of coincidence information. In one image processing method used to reconstruct the tomographic image, described in Non-Patent Document 1, the tomographic image is reconstructed using a deep image prior technique (DIP) with a convolutional neural network, which is a type of deep neural network. Hereinafter, the convolutional neural network will be referred to as a "CNN" and the deep image prior technique as a "DIP technique."

[0006] The DIP technique exploits the property of CNNs that meaningful structures in an image are learned faster than random noise (i.e., random noise is less likely to be learned). By using the DIP technique, a tomographic image with reduced noise can be acquired.

[0007] Specifically, the image processing method described in non-patent document 1 is as follows. A sinogram (hereinafter referred to as the "measured sinogram") is generated based on the wealth of coincidence information collected for the subject. Furthermore, a sinogram (hereinafter referred to as the "computed sinogram") is generated by performing a forward projection calculation (Radon transform) on an image output by the CNN when an input (e.g., an MRI image) is received by the CNN.

[0008] Furthermore, an error between the calculated sinogram and the measured sinogram is evaluated, and the CNN is trained based on the error evaluation result. Using the DIP technique, the calculated sinogram gradually approaches the measured sinogram, and the image output by the CNN approaches the tomographic image of the subject, when the image output from the CNN, the generation of the calculated sinogram via forward projection computation, the error evaluation, and the training of the CNN are repeatedly performed.

[0009] The image processing method described above includes a process for performing forward projection from the CNN output image to the computed sinogram, but it does not include a process for performing back projection from the measured sinogram to the tomographic image. This would make it possible to obtain a tomographic image with further noise reduction. Furthermore, in the image processing method described in non-patent document 1, both the measured sinogram and the computed sinogram are divided into a multitude of blocks, and the error between the computed sinogram and the measured sinogram is evaluated for each block to train the CNN, thereby reducing the memory capacity required at the time of processing.

[0010] The sinogram is expressed as a histogram representing a frequency (a generation frequency of coincidence events) at which the coincidence information is captured in a space (a sinogram space) represented by four variables: r, θ, z, and δ. The variable r represents a position in the radial direction of a coincidence detection line (a line connecting two detectors that perform coincidence detection of the photon pair). The variable θ represents an azimuth angle of the coincidence detection line. The variable z represents a position along the midpoint of the coincidence detection line. Furthermore, the variable δ represents a polar angle of the coincidence detection line. List of citations Non-patent literature Non-patent document 1: F. Hashimoto, Y. Onishi, K. Ote, H. Tashima and T. Yamaya, “Fully 3D Implementation of the End-to-end Deep Image Prior-based PET Image Reconstruction Using Block Iterative Algorithm,” arXiv preprint, arXiv:2212.11844, doi: 10.48550 / arXiv.2212.11844, 2022 Non-patent document 2: J. Nuyts et al., “A concave prior penalizing relative differences for maximum-a-posteriori reconstruction in emission tomography,” IEEE TNS, Volume 49, Issue 1, pp. 56-60, 2002 Non-patent document 3: Hiroyuki Kudo, “Image reconstruction methods in low-dose CT: Fundamentals of statistical image reconstruction, iterative image reconstruction and compressed scanning”, Medical Imaging Technology, Volume 32, No. 4, pp. 239-248, 2014 Non-patent document 4: Antonin Chambolle, “An algorithm for minimizing total variation and applications”, Journal of Mathematical Imaging and Vision 20, pp. 89-97, 2004 Summary of the invention: Technical problem

[0011] The image processing method described in non-patent document 1 using the DIP technique exhibits excellent noise reduction performance, but requires a large number of CNN training epochs and takes time to obtain the tomographic image of the object in which the noise is reduced.

[0012] The objective of the present invention is to provide an image processing device and an image processing method with which the number of CNN training epochs can be reduced by using a DIP technique and a tomographic image with reduced noise development can be obtained in a short time when the tomographic image of an object is generated by training a CNN based on an evaluation result of an error between a calculated sinogram and a measured sinogram. Solution to the problem

[0013] One embodiment of the present invention is an image processing device. The image processing device is an image processing device for generating a tomographic image of a subject based on coincidence information collected by a radiation tomography device comprising a plurality of detectors arranged around a measurement chamber in which the subject, into which a radiation source is injected, is placed.and comprises (1) an image reconstruction unit for generating a reconstructed image using an analytical image reconstruction method or a statistical image reconstruction method based on the coincidence information collected by the radiation tomography device; (2) a sinogram generation unit for generating a sinogram based on the coincidence information collected by the radiation tomography device; (3) a CNN processing unit for inputting an input image into a convolutional neural network and for the convolutional neural network to generate an output image; (4) a forward projection computation unit for performing a forward projection computation on the output image to generate a sinogram; and (5) a CNN training unit for training the convolutional neural network,and (6) in a first processing step (a) the CNN training unit uses a scoring function containing an error scoring term that represents a scoring value in terms of an error between the reconstructed image and the original image, and trains the convolutional neural network based on a value of the scoring function and (b) the respective processes of the CNN processing unit and the CNN training unit are repeated in a multitude of cases and (7) in a second processing step following the first processing step, (a) the CNN training unit uses a scoring function containing an error scoring term that represents a scoring value in terms of an error between the sinogram produced by the sinogram generation unit and the sinogram produced by the forward projection computation,and trains the convolutional neural network based on a value of the evaluation function, and (b) the output image, after the respective processes of the CNN processing unit, the forward projection computation unit and the CNN training unit have been performed in a large number of repetitions, is defined as a tomographic image of the subject.

[0014] One embodiment of the present invention is a radiation tomography system. The radiation tomography system comprises a radiation tomography device with a plurality of detectors arranged around a measurement chamber in which a subject, into which an RI source has been injected, is placed, and for collecting coincidence information; and the image processing device of the above-mentioned configuration for generating the tomographic image of the subject based on the coincidence information collected by the radiation tomography device.

[0015] One embodiment of the present invention is an image processing method. The image processing method is an image processing method for generating a tomographic image of a subject based on coincidence information collected by a radiation tomography device comprising a plurality of detectors arranged around a measurement chamber in which the subject, into which a radiation source is injected, is placed.and comprises (1) an image reconstruction step to generate a reconstructed image using an analytical image reconstruction method or a statistical image reconstruction method based on the coincidence information collected by the radiation tomography device; (2) a sinogram generation step to generate a sinogram based on the coincidence information collected by the radiation tomography device; (3) a CNN processing step to input an input image into a convolutional neural network and to generate an output image by the convolutional neural network; (4) a forward projection computation step to perform a forward projection computation on the output image to generate a sinogram; and (5) a CNN training step to train the convolutional neural network; and (6) in a first processing step,(a) in the CNN training step, a scoring function is used which contains an error scoring term that represents a scoring value with respect to an error between the reconstructed image and the original image, and the convolutional neural network is trained based on a value of the scoring function, and (b) the respective processes of the CNN processing step and the CNN training step are carried out in a multitude of repetitions, and (7) in a second processing step following the first processing step, (a) in the CNN training step, a scoring function is used which contains an error scoring term that represents a scoring value with respect to an error between the sinogram produced in the sinogram generation step and the sinogram produced in the forward projection computation step, and the convolutional neural network is trained based on a value of the scoring function,and (b) the initial image after the respective processes of the CNN processing step, the forward projection computation and the CNN training step, which are repeated a multitude of times, is defined as the tomographic image of the subject. Advantageous effects of the invention

[0016] According to the embodiments of the present invention, it is possible to reduce the number of CNN training epochs by using a DIP technique and to obtain a tomographic image in a short time in which the noise is reduced when the tomographic image of a subject is generated by training a CNN based on an evaluation result of an error between a calculated sinogram and a measured sinogram. Brief description of the drawings Fig. Figure 1 is a diagram illustrating a configuration of a radiation tomography system 1. Fig. Figure 2 is a diagram illustrating a configuration example of a CNN. Fig. Figure 3 is a flowchart illustrating an image processing procedure. Fig. Figure 4 is a flowchart illustrating a first processing step of the image processing procedure. Fig. Figure 5 is a flowchart illustrating a second processing step of the image processing procedure. Fig. Figure 6 contains diagrams showing examples of a calculated sinogram 24 in the case where no block subdivision is performed, and calculated sinograms 241 to 24. 16 in the case where block subdivision is performed, in comparison to each other, and shows (a) a diagram schematically showing the calculated sinogram 24 in the case where block subdivision is not performed, and (b) a diagram schematically showing the calculated sinograms 241 to 24. 16in the case where the block splitting is carried out. Fig. Figure 7 is a diagram for describing neighboring pixels in a source image. Fig. Figure 8 is a diagram showing an original image (a correct image) of a tomographic image of a brain. Fig. Figure 9 is a diagram showing the tomographic image of the brain obtained using an initial image processing technique. Fig. Figure 10 is a diagram showing the tomographic image of the brain obtained using a second image processing technique. Description of embodiments

[0017] The following are detailed descriptions of embodiments of an image processing device and an image processing method with reference to the accompanying drawings. In the description of the drawings, identical elements are designated by the same reference numerals, and redundant descriptions are omitted. The present invention is not limited to these examples, and the claims, their equivalents, and any modifications within the scope are to be understood as falling within the scope of the present invention.

[0018] Fig. Figure 1 is a diagram representing a configuration of a radiation tomography system 1. The radiation tomography system 1 comprises a radiation tomography device 2 and an image processing device 10. The image processing device 10 comprises a sinogram generation unit 11, a CNN processing unit 12, a convolution integration unit 13, a forward projection computation unit 14, a CNN training unit 15, a storage unit 16, and an image reconstruction unit 17.

[0019] Furthermore, each of the input images, reconstructed images, output images, and tomographic images used in the image processing device 10 can be a two-dimensional image or a three-dimensional image, and in the following description, it is assumed that each of the aforementioned images is a three-dimensional image. Additionally, each of the measured and calculated sinograms may or may not be subdivided into a plurality of blocks. The following description mainly describes the case in which each of the aforementioned sinograms is subdivided into a plurality of blocks.

[0020] The radiation tomography device 2 is a device for collecting coincidence information to reconstruct a tomographic image of a subject. Examples of radiation tomography device 2 are a PET device and a SPECT device. In the following description, radiation tomography device 2 is assumed to be the PET device.

[0021] The radiation tomography device 2 comprises a detection unit with a large number of small radiation detectors arranged around a measurement chamber in which the object is placed. The radiation tomography device 2 detects a photon pair with an energy of 511 keV, generated by electron-positron annihilation in the subject into which an RI source is injected, using a coincidence method with the detection unit and collects the coincidence information. Furthermore, the radiation tomography device 2 outputs the collected coincidence information to the image processing device 10.

[0022] The image processing device 10 comprises a graphics processing unit (GPU) for performing processing using a convolutional neural network (CNN), an input unit (e.g., a keyboard or a mouse) for input from an operator, a display unit (e.g., a liquid crystal display) for displaying an image, and a storage unit for storing a program and data for performing various types of processing. For example, a computer comprising a CPU, RAM, ROM, hard disk drive, and the like is used as the image processing device 10.

[0023] The storage unit 16 inputs and stores the coincidence information collected by the radiation tomography device 2. The image reconstruction unit 17 generates a reconstructed image based on the coincidence information stored by the storage unit 16 using a general image reconstruction method. This general image reconstruction method is not a reconstruction method performed using a CNN, but rather an analytical or statistical image reconstruction method.

[0024] Examples of analytical image reconstruction methods include filtered backprojection (FBP) and similar techniques. Examples of statistical image reconstruction methods include techniques such as maximum likelihood expectation maximization (ML-EM), ordered subset expectation maximization (OSEM), row action maximum likelihood algorithm (RAMLA), and dynamic RAMLA (DRAMA), and similar techniques.

[0025] The sinogram generation unit 11 generates a measured sinogram 21 based on the coincidence information stored by the storage unit 16. In this case, the sinogram generation unit 11 generates measured sinograms 211 to 212. K , which are subdivided into a multitude of (K) blocks. The measured sinogram 21 kis a measured sinogram of the k-th block from the K blocks. K is an integer of 2 or more, and k is an integer of 1 or more and K or less. The entire measured sinogram 21 is obtained by combining the subdivided measured sinograms 211 to 21. K receive.

[0026] The CNN processing unit 12 inputs a three-dimensional input image 20 into the CNN and generates a three-dimensional output image 22 through the CNN. The three-dimensional input image 20 can be an image representing morphological information of the object, it can be an MRI image, a CT image, or a static PET image of the object, or it can be an image of random noise.

[0027] The CNN processing unit 12 inputs a three-dimensional input image 20 into the CNN and generates a three-dimensional output image 22 through the CNN. The three-dimensional input image 20 can be an image representing morphological information of the object, it can be an MRI image, a CT image, or a static PET image of the object, or it can be an image of random noise.

[0028] The convolution integration unit 13 performs convolution integration of a point spread function on the three-dimensional output image 22 generated by the CNN processing unit 12, producing a new three-dimensional output image 23. The point spread function (PSF) is a function representing the response (impulse response) of the radiation tomography device to a point source and is generally represented by a Gaussian function, specifically an asymmetric Gaussian function where the degree of blurring varies depending on the position within a visual field modeled from the point source measurement data, etc. By providing the convolution integration unit 13, a tomographic image with improved image quality can be obtained, and furthermore, the CNN training can be stabilized.

[0029] The forward projection processing unit 14 performs a forward projection calculation on the three-dimensional source image 23 to generate a calculated sinogram 24. In this case, the forward projection processing unit 14 generates calculated sinograms 241 to 244. K , which are subdivided into K-blocks. The calculated sinogram 24 k is a calculated sinogram of the k-th block from the K-blocks. The entire calculated sinogram 24 is obtained by combining the subdivided calculated sinograms 241 to 24. K receive.

[0030] The block subdivision of the calculated sinogram 24 is performed in the same way as the block subdivision of the measured sinogram 21. The calculated sinogram 24 k of the k-th block and the measured sinogram 21 kThe k-th block comprises the sinograms of a common region within an entire sinogram space. The block partition configuration is arbitrarily chosen, and the partition can be performed for any one, two, or more variables from the four variables representing the sinogram space. The size of each of the k blocks can be different or the same.

[0031] The CNN training unit 15 trains the CNN using the DIP technique with a scoring function that includes an error scoring term. This term differs between a first processing step and a second processing step, which will be described later. In the first processing step, the CNN training unit 15 uses the scoring function, which contains the error scoring term representing a score for an error between the reconstructed image produced by the image reconstruction unit 17 and the original image produced by the convolution integration unit 13, and trains the CNN based on the error scoring result.

[0032] In the second processing step, the CNN training unit 15 uses the error function, which contains the error evaluation term that assigns an evaluation value in terms of an error between the measured sinogram 21. k and the calculated sinogram 24k The CNN is trained based on the error evaluation result for each of the K blocks. CNN training unit 15 continuously performs the second processing step training of the CNN after the first processing step training.

[0033] In the first processing step, the respective processes of the CNN processing unit 12, the convolutional integration unit 13, and the CNN training unit 15 are repeatedly performed in a multitude of cases. A processing sequence of the first processing step is shown in the Fig. The image processing device 10 to be represented is indicated by a dashed arrow.

[0034] In the second processing step, which follows the first, the respective processes of the CNN processing unit 12, the convolution integration unit 13, the forward projection calculation unit 14, and the CNN training unit 15 are carried out in a multitude of repetitions. A processing sequence of the second processing step is indicated by a solid arrow in the Fig. 1 image processing device 10 is shown.

[0035] Furthermore, the three-dimensional output image 22 generated by the CNN processing unit 12 is defined as the three-dimensional tomographic image of the object. The three-dimensional output image 23 generated by the convolution integration unit 13 can also be defined as the three-dimensional tomographic image of the object. The measured sinogram 21 is a sinogram reflecting a response function of the radiation tomography device, and therefore it is preferred that the three-dimensional output image 22 be defined as the three-dimensional tomographic image of the subject before the convolution integration of the point spread function is performed by the convolution integration unit 13.

[0036] Furthermore, the convolutional integration unit 13 can be provided as the last layer of the CNN or separately from the CNN. If the convolutional integration unit 13 is provided as the last layer of the CNN, its weighting coefficient is kept constant during CNN training.

[0037] Furthermore, the convolution integration unit 13 may also be unavailable. In the case where the convolution integration unit 13 is unavailable, the CNN training unit 15, in the first processing step, uses the evaluation function, which contains the error evaluation term. This term represents a value in terms of an error between the reconstructed image generated by the image reconstruction unit 17 and the three-dimensional output image 22 generated by the CNN processing unit 12, and trains the CNN based on the error evaluation result. If the convolution integration unit 13 is unavailable, the forward projection calculation unit 14, in the second processing step, performs the forward projection calculation on the three-dimensional output image 22 generated by the CNN processing unit 12 to generate the calculated sinogram 24.

[0038] Fig. Figure 2 is a diagram illustrating a configuration example of the CNN. The CNN shown in this diagram has a three-dimensional U-Net structure comprising an encoder and a decoder. The diagram shows the size of each layer of the CNN, assuming that the number of pixels in the three-dimensional input image (20) fed into the CNN is N × N × 64.

[0039] Fig. 3 to Fig. Five flowcharts illustrate an image processing procedure. According to the representation in Fig. The image processing method comprises a first processing step S1 and a subsequent second processing step S2. The CNN training unit 15 trains the CNN in the first processing step and continues to train the CNN continuously in the second processing step.

[0040] According to the representation in Fig. 4 The first processing step S1 includes an image reconstruction step S17, which is performed by the image reconstruction unit 17, a CNN processing step S12, which is performed by the CNN processing unit 12, a convolution integration step S13, which is performed by the convolution integration unit 13, and a CNN training step S15, which is performed by the CNN training unit 15.

[0041] In image reconstruction step S17, the reconstructed image is generated using the general image reconstruction method. In CNN processing step S12, the three-dimensional input image 20 is fed into the CNN, and the three-dimensional output image 22 is generated by the CNN. In convolution integration step S13, the convolution integration of the point spread function is performed on the three-dimensional output image 22 generated in CNN processing step S12 to generate the new three-dimensional output image 23.

[0042] In the CNN training step S15, the evaluation function, including the error evaluation term, is used. This term represents the evaluation value in terms of the error between the reconstructed image generated in the image reconstruction step S17 and the three-dimensional original image 23 generated in the convolution integration step S13. The CNN is trained based on the error evaluation result. In the first processing step S1, the respective processes of the CNN processing step S12, the convolution integration step S13, and the CNN training step S15 are repeatedly performed in a multitude of cases.

[0043] According to the representation in Fig. 5, the second processing step S2 comprises a sinogram generation procedure S21, which is performed by the sinogram generation unit 11, a CNN processing step S22, which is performed by the CNN processing unit 12, a convolution integration step S23, which is performed by the convolution integration unit 13, a forward projection calculation S24, which is performed by the forward projection calculation unit 14, and a CNN training step S25, which is performed by the CNN training unit 15.

[0044] In the sinogram generation step S21, the measured sinograms 211 to 21 are K, which are subdivided into K-blocks, based on the coincidence information collected by the radiation tomography device 2. In CNN processing step S22, the three-dimensional input image 20 is fed into the CNN, and the three-dimensional output image 22 is generated by the CNN. In convolution integration step S23, the convolution integration of the point spread function is performed on the three-dimensional output image 22 generated in CNN processing step S22 to generate the new three-dimensional output image 23.

[0045] In the forward projection calculation step S24, the forward projection calculation for the three-dimensional source image 23 is performed in order to generate the calculated sinograms 241 to 24. K to generate blocks that are subdivided into K blocks. In the CNN training step S25, the error between the measured sinogram 21 is determined. k and the calculated sinogram 24 kEach of the K-blocks is evaluated, and the CNN is trained based on the error evaluation result for each of the K-blocks.

[0046] In the second processing step S2, the three-dimensional output image 22, generated in the CNN processing step S22, is defined as the three-dimensional tomographic image of the subject after the respective processes of the CNN processing step S22, the convolution integration step S23, the forward projection computation S24, and the CNN training step S25 have been repeatedly performed in a multitude of cases. The three-dimensional output image 23 generated in the convolution integration step S23 can also be defined as the three-dimensional tomographic image of the subject. Alternatively, the convolution integration step S23 can also be omitted.

[0047] Next, in the second processing step, the processing content of each step of the image processing procedure will be described for the case where the sinogram is not divided into a multitude of blocks. In the image processing procedure for the case where the sinogram is not divided into blocks, the processing is performed on the entire sinogram of each of the measured sinograms and the calculated sinogram.

[0048] In the following, the processing by the CNN is denoted by f, the three-dimensional input image 20 fed into the CNN is denoted by z, and a weighting coefficient parameter representing a training state of the CNN is denoted by θ. As the CNN is trained, θ changes. In the case where the three-dimensional input image z is fed into the CNN with the weighting coefficient θ, the three-dimensional output image 22 from the CNN is set to x. The three-dimensional output image x is represented by the following formula (1). In the CNN processing step, the processing represented by the following formula is performed to generate the three-dimensional output image x. [Formula 1] x=f(θ|z)

[0049] In the convolution integration step, the convolution integration of the point spread function is performed on the three-dimensional output image x, which was generated in the CNN processing step, and the new three-dimensional output image x is generated. Furthermore, in Fig. 1 the three-dimensional initial image x after performing the convolution integration is denoted by PSF(f(θ|z)).

[0050] In the forward projection calculation step, the forward projection calculation is performed on the three-dimensional source image x to generate the calculated sinogram 24. The calculated sinogram 24 is set to y, and a projection matrix for performing the forward projection calculation (Radon transform) from the three-dimensional source image x to the calculated sinogram y is set to P. The projection matrix is ​​also referred to as the system matrix or recognition probability. The processing performed in the forward projection calculation step is represented by the following formula (2). [Formula 2] y=Px

[0051] In the CNN training step, during the second processing step, the measured sinogram 21 is set to y0, the error between the measured sinogram y0 and the calculated sinogram y (formula (2) above) is evaluated, and the CNN is trained based on the error evaluation result. The processing performed in the CNN training step is represented by the following formula (3). An optimized optimization problem represented by the following formula is a problem of optimizing the CNN parameter θ such that a value of an evaluation function E(y;y0) becomes small under the condition that the three-dimensional output image x generated by the CNN is the tomographic image of the subject. [Formula 3] minE=(y;y0)stx=f(θ|z)

[0052] The optimized optimization problem represented by formula (3) above can be rewritten as a non-optimized optimization problem represented by formula (4) below. The evaluation function E is set to an arbitrary function, and, for example, an L1 norm, an L2 norm, a negative log probability in a Poisson distribution, or the like can be used. In the case where the L2 norm is used as the evaluation function, formula (4) below can be rewritten as formula (5) below. [Formula 4] min E(Pf(θ|z)−y0) [Formula 5] θ*=argminθ‖(Pf(θ|z)−y0)‖x*=f(θ*|z)

[0053] Considering the arrangement of the number of detectors provided in the radiation tomography apparatus, there may be a region where collecting coincidence information in the sinogram space is impossible. Given this fact, instead of the optimization problem represented by formula (5) above, an optimization problem represented by formula (6) can be used. In formula (6), m is a binary mask function and has a value of 1 in a region where collecting coincidence information in the sinogram space is possible, and a value of 0 in a region where collecting coincidence information is impossible. Formula (6) is a formula for selectively evaluating the error in the region of the sinogram where collecting coincidence information is possible by forming a Hadamard product of an error (y - y0) and the binary mask function m. [Formula 6] θ*=argminθ‖(Pf(θ|z)−y0)⊙m‖

[0054] In the second processing step, the calculated sinogram y approximates the measured sinogram y0 by repeatedly executing the respective processes of the CNN processing step, the convolution integration step, the forward projection calculation and the CNN training step in the multitude and by solving the above-mentioned optimization problem for the CNN parameter θ, and the three-dimensional output image x generated by the CNN approximates the tomographic image of the subject.

[0055] Next, in the second processing step, the processing content of each step of the image processing procedure is described in detail in the case where the sinogram is divided into blocks. In the case where the sinogram is divided into blocks, the forward projection calculation for the three-dimensional source image x is performed in the forward projection calculation step to generate the calculated sinograms 241 to 24. K to create blocks that are divided into K blocks.

[0056] The calculated sinogram 24 k of the k-th block is on y k set, and the projection matrix for performing the forward projection calculation (Radon transformation) from the three-dimensional initial image x to the calculated sinogram y k will be on P k The processing performed in the forward projection calculation step is represented by the following formula (7). [Formula 7] yk=Pkx(k=1,2,3,⋯,K)

[0057] In the CNN training step, in the second processing step, the measured sinogram 21 k of the k-th block on y 0k set, the error between the measured sinogram y 0k and the calculated sinogram y k Each of the K-blocks is evaluated. The CNN is trained based on the error evaluation result for each of the K-blocks.

[0058] The processing performed in the CNN training step is represented by an unconstrained optimization problem of the following formula (8). In the case where the L2 norm is used as the evaluation function, the following formula (8) can be rewritten as the following formula (9). If the error is selectively evaluated in the region of the sinogram space where coincidence information gathering is possible, it is represented by an unconstrained optimization problem of the following formula (10). kis a binary mask function in the k-th block. [Formula 8] min E(Pkf(θ|z)−y0k) [Formula 9] θ*=argminθ‖(Pkf(θ|z)−y0k)‖(k=1,2,3,⋯,K)x*=f(θ*|z) [Formula 10] θ*=argminθ‖(Pkf(θ|z)−y0k)⊙mk‖(k=1,2,3,⋯,K)

[0059] In the second processing step, the calculated sinogram approaches y. k for each of the K-blocks the measured sinogram y 0k and the three-dimensional output image x, generated by the CNN, approximates the tomographic image of the object by repeatedly performing the respective processes of the CNN processing step, the convolution integration step, the forward projection computation, and the CNN training step in a multitude of cases and solving the above optimization problem for the CNN parameter θ.

[0060] Next, the comparison between the case where the sinogram is not divided into blocks and the case where the sinogram is divided into blocks will be described with regard to the storage capacity required to store the data in the GPU's RAM.

[0061] Generally, the GPU is used in processing with the CNN. The GPU is a processing device specialized for image processing and comprises the processing unit and RAM integrated on a semiconductor chip. Various types of data used by the GPU's processing unit must be stored in the GPU's RAM.

[0062] The data to be stored in the GPU's RAM includes, for example, the CNN input image, the CNN output image, the weighting coefficient representing the CNN's training state, a feature map, the measured sinogram, the calculated sinogram, a parameter required for performing the forward projection calculation, and the like, requiring enormous memory capacity. However, the GPU's RAM capacity is limited, so while the image processing method described above can perform two-dimensional forward projection calculations, three-dimensional forward projection calculations can be difficult.

[0063] In this case, the number of pixels of the three-dimensional output image generated by the CNN is set to 128 × 128 × 64, and the number of pixels of the sinogram space is set to 128 × 128 × 64 × 19. In the image processing procedure, where the sinogram is divided into blocks, K = 16 is set, and the forward projection calculation is performed on the three-dimensional output image to generate the calculated sinograms 241 to 24. 16 to create blocks that are evenly divided into 16 blocks.

[0064] Fig. Figure 6 contains diagrams showing examples of the calculated sinogram 24 in the case where the block division is not carried out, and of the calculated sinograms 241 to 24. 16 in the case where block splitting is carried out, show in comparison to each other. (a) in Fig. Figure 6 is a diagram that schematically shows the calculated sinogram 24 in the case where the block subdivision is not carried out. (b) in Fig. Figure 6 is a diagram that schematically shows the calculated sinograms 241 to 24. 16 in which case the block subdivision is carried out.

[0065] The number of pixels in the calculated sinogram is 24. k Each of the blocks, in the case where block splitting is performed, measures 128 × 8 × 64 × 19, which is 1 / 16 of the number of pixels of the calculated sinogram 24 in the case where block splitting is not performed. Furthermore, the number of elements of the projection matrix P is k to perform the forward projection calculation from the three-dimensional source image to the calculated sinogram 24 kof the k-th block in the case where the block splitting is performed, 1 / 16 of the number of elements of the projection matrix P to perform the forward projection calculation from the three-dimensional source image to the calculated sinogram 24 in the case where the block splitting is not performed.

[0066] In cases where block subdivision is performed, the memory capacity required to store the data used for forward projection calculations can be reduced compared to cases where block subdivision is not performed. This allows the aforementioned data to be stored in the GPU's RAM. Therefore, when block subdivision is implemented, it becomes easier to perform the three-dimensional forward projection calculation from the CNN's output image to the calculated sinogram. The three-dimensional tomographic image of the subject can then be easily generated by training the CNN based on the error analysis results between the calculated and measured sinograms.

[0067] Next, the evaluation function used by CNN training unit 15 in the CNN training step will be described in more detail. In the description above, the evaluation function used in the first processing step is defined as the evaluation function that contains only the error evaluation term, which represents the evaluation value in terms of the error between the reconstructed image produced by image reconstruction unit 17 and the original image produced by convolution integration unit 13. Furthermore, the evaluation function used in the second processing step is defined as the evaluation function that contains only the error evaluation term, which represents the evaluation value in terms of the error between the measured sinogram y0 and the computed sinogram y (= Pf(θ|z)).

[0068] Furthermore, it is preferred to use a scoring function in both the first and second processing steps that includes a regularization term in addition to the error scoring term. The regularization term is a term used to suppress overtraining of the CNN and represents a score based on the difference in pixel values ​​between neighboring pixels in the original image.

[0069] This means that the evaluation function for the case where the sinogram is not divided into blocks is changed from formula (5) above to formula (11). Furthermore, the evaluation function for the case where the sinogram is divided into blocks is changed from formula (9) above to formula (12). [Formula 11] θ*=argminθ‖(Pf(θ|z)−y0)‖+β⋅R(f(θ|z))x*=f(θ*|z) [Formula 12] θ*=argminθ‖(Pkf(θ|z)−y0k)‖+β⋅R(f(θ|z))(k=1,2,3,⋯,K)x*=f(θ*|z)

[0070] In the formulas above, the first term on the right-hand side is the error-weighting term, and the second term on the right-hand side is the regularization term. The regularization term penalizes the difference in pixel values ​​between adjacent pixels in the original image. β is a hyperparameter used to adjust the magnitude of the regularization effect. The smaller the value of β, the less effective the regularization. The larger the value of β, the greater the effect of the regularization (i.e., the greater the effect of suppressing CNN overtraining).

[0071] The regularization term can represent the evaluation value relating to the difference of pixel values ​​between neighboring pixels in the output image 22 (f(θ|z)) output by the CNN processing unit 12, or the evaluation value relating to the difference of pixel values ​​between neighboring pixels in the output image 23 (PSF(f(θ|z)) output by the convolution integration unit 13.

[0072] In the case where the two-dimensional image is used, the pixels adjacent to a given pixel include pixels that border the given pixel in two mutually orthogonal directions, and preferably also include pixels that border the given pixel in diagonal directions. In the case of the two-dimensional image, the number of pixels adjacent to the given pixel is 8, excluding pixels located at the edge or corner of the image.

[0073] When a three-dimensional image is used, the pixels adjacent to a given pixel include pixels that border that pixel in three mutually orthogonal directions, and preferably also pixels that border that pixel in diagonal directions. In the case of the three-dimensional image, the number of pixels adjacent to the given pixel is 26, excluding pixels located at the edge or corner of the image.

[0074] Fig. Figure 7 is a diagram describing the neighboring pixels in the original image. In this diagram, the original image is represented as a two-dimensional image, and 3 × 3 pixels are shown in the image. In this diagram, when a pixel value of the pixel in the middle is set to λ j is set and a pixel value of each of the eight pixels adjacent to the middle pixel is set to λ k(k = 1 to 8) is set, the difference of the pixel values ​​between the neighboring pixels with respect to the central pixel according to the above description is given by |λ j - λ k | shown. The regularization term represents the evaluation value, which refers to the difference in pixel values ​​for all combinations of neighboring pixels in the original image.

[0075] The regularization term is a term used to represent the evaluation value, which relates to the difference in pixel values ​​between neighboring pixels in the original image, and can be represented using various formulas. For example, the regularization term is represented by the following formula (13). In the following formula (13), N represents j γ represents the set of pixels k adjacent to pixel j. γ represents the magnitude of the change in the value of the regularization term with respect to the change in the pixel value λ. jThe following formula (13) contains a term of the difference of the pixel values ​​of the neighboring pixels in the numerator and a term of the sum of the pixel values ​​of the neighboring pixels in the denominator and thus represents the evaluation value, which refers to a relative difference of the pixel values ​​between the neighboring pixels in the original image. [Formula 13] R(f(θ|z))=∑j∑k∈Nj(λj−λk)2(λj+λk)−γ|λj−λk|

[0076] Furthermore, formula (13) is similar to a formula described in non-patent document 2. However, in non-patent document 2, the formula similar to formula (13) is used in the processing for the reconstruction of the subject's tomographic image based on the coincidence information collected using the PET device, and the formula is not used in the noise reduction processing of the tomographic image using the DIP technique.

[0077] Examples of regularization terms that can be used include Gibbs prior (non-patent document 3), total variation (non-patent document 4), or similar terms. Furthermore, the aforementioned documents describe a technique for performing reconstruction processing of the subject's tomographic image but do not describe a technique for performing noise reduction processing of the tomographic image using the DIP technique.

[0078] Next, the result obtained by generating simulation data using a Monte Carlo simulation of a head PET device using a digital brain phantom image and by reconstructing the tomographic image using each of the first and second image processing methods using the simulation data is described.

[0079] In the first image processing procedure of a comparison example, the tomographic image was reconstructed by performing only the second processing step, omitting the first. In the second image processing procedure, the tomographic image was reconstructed by performing the second processing step after the first. In both the first and second image processing procedures, the second processing step divided both the measured sinogram and the calculated sinogram into a multitude of blocks, and the CNN was trained using the scoring function, which included only the error scoring term without the regularization term.

[0080] A three-dimensional image was used as the phantom image. This image was generated using a brain image obtained from BrainWeb (https: / / brainweb.bic.mni.mcgill.ca / brainweb / ) and by embedding a simulated tumor in an area of ​​the white matter of the brain image. The number of pixels in the phantom image was set to 128 × 128 × 64. The number of pixels in the sinogram was set to 128 × 128 × 64 × 19, and the sinogram was divided equally into two blocks. The error weighting term of the evaluation function was set to a mean squared error (MSE). A three-dimensional random noise image was used as the input image for the CNN.

[0081] Fig. Figure 8 is a diagram showing the original (correct) image of the tomographic image of the brain. Fig. Figure 9 is a diagram showing the tomographic image of the brain obtained using the first image processing method. Fig. Figure 10 is a diagram showing the tomographic image of the brain obtained using the second image processing method.

[0082] The PSNR of the tomographic image obtained with the first image processing method ( Fig. 9) is 20.70 dB, and the PSNR of the tomographic image obtained with the second image processing method ( Fig. 10) is 21.60 dB. The peak signal-to-noise ratio (PSNR) is a value that indicates image quality in decibels (dB), with a higher value meaning better image quality.

[0083] According to the description above, the second image processing method results in a higher PSNR of the tomographic image and excellent noise reduction performance compared to the first image processing method. Furthermore, the second image processing method allows for a more detailed reconstruction of the gray matter structure.

[0084] In the first image processing method, the number of repetitions in the second processing step is 11, and the required time is approximately 165 minutes. In the second image processing method, the required time for the first processing step is several minutes, the number of repetitions in the second processing step is 3, and the required time for the second processing step is approximately 45 minutes. According to the description above, the required time for the second image processing method is about one-third compared to the first image processing method.

[0085] According to the above description, in comparison to the first image processing method, in which only the second processing step is performed, the second image processing method, in which the first processing step and the subsequent second processing step are performed, confirms that the tomographic image, in which the noise is further reduced, can be obtained in a shorter time.

[0086] In the present embodiment, it is possible to shorten the time required for the entire processing, including the time required for the second processing step, to a certain extent by performing the training of the CNN in a short time in the first processing step and subsequently performing further training of the CNN in the second processing step, and in addition to obtain the tomographic image in which the noise is reduced.

[0087] The image processing device and the image processing method are not limited to the embodiments and configuration examples shown in the description above, and various modifications are possible.

[0088] The image processing device of a first aspect according to the above embodiment is an image processing device for generating a tomographic image of an object based on coincidence information collected by a radiation tomography device comprising a plurality of detectors arranged around a measurement space in which the object, into which an RI source is injected, is placed.and comprises (1) an image reconstruction unit for generating a reconstructed image using an analytical image reconstruction method or a statistical image reconstruction method based on the coincidence information collected by the radiation tomography device; (2) a sinogram generation unit for generating a sinogram based on the coincidence information collected by the radiation tomography device; (3) a CNN processing unit for inputting an input image into a convolutional neural network and for the convolutional neural network to generate an output image; (4) a forward projection computation unit for performing a forward projection computation on the output image to generate a sinogram; and (5) a CNN training unit for training the convolutional neural network,and (6) in a first processing step, (a) the CNN training unit uses a scoring function containing an error scoring term that represents a scoring value in terms of an error between the reconstructed image and the output image, and trains the convolutional neural network based on a value of the scoring function, and (b) the respective processes of the CNN processing unit and the CNN training unit are repeated a multitude of times, and (7) in a second processing step following the first processing step, (a) the CNN training unit uses a scoring function containing an error scoring term that represents a scoring value in terms of an error between the sinogram produced by the sinogram generation unit and the sinogram produced by the forward projection computation,and trains the convolutional neural network based on a value of the evaluation function, and (b) the output image, after the respective processes of the CNN processing unit, the forward projection computation unit and the CNN training unit have been performed repeatedly a multitude of times, is defined as a tomographic image of the subject.

[0089] In the image processing device of a second aspect, in the configuration of the first aspect, the sinogram generation unit can generate the sinogram, which is divided into a multitude of blocks, based on the coincidence information collected by the radiation tomography device; the forward projection calculation unit can perform the forward projection calculation for the output image to generate the sinogram divided into a multitude of blocks; and in the second processing step, the CNN training unit can train the convolutional neural network based on the value of the evaluation function for each of the number of blocks.

[0090] In the image processing device of a third aspect, in the configuration of the first or second aspect, each of the tomographic images, the reconstructed image, the input image and the output image can be a three-dimensional image.

[0091] In the image processing device of a fourth aspect, the device in the configuration of any of the first to third aspects may further include a convolution integration unit for performing a convolution integration of a point spread function on the output image, and the forward projection calculation unit may perform the forward projection calculation on the output image after a process of the convolution integration unit has been carried out.

[0092] In the image processing device of a fifth aspect, in the configuration of any of the first to fourth aspects, the CNN training unit can evaluate the error in the second processing step using the error evaluation term in a region in a sinogram space where the collection of coincidence information by the radiation tomography device is possible.

[0093] In the image processing device of a sixth aspect, the CNN training unit, in the configuration of any of the first to fifth aspects, can use the evaluation function, which further includes a regularization term representing an evaluation value that relates to a difference of pixel values ​​between neighboring pixels in the output image, and train the convolutional neural network based on the value of the evaluation function.

[0094] In the image processing device of a seventh aspect, the CNN processing unit, in the configuration of any of the first to sixth aspects, can input an image representing morphological information of the subject as an input image into the convolutional neural network.

[0095] In the image processing device of an eighth aspect, the CNN processing unit, in the configuration of any of the first to sixth aspects, can input an MRI image of the subject as an input image into the convolutional neural network.

[0096] In the image processing device of a ninth aspect, in the configuration of any of the first to sixth aspects, the CNN processing unit can input a CT image of the subject as an input image into the convolutional neural network.

[0097] In the image processing device of a tenth aspect, the CNN processing unit, in the configuration of any of the first to sixth aspects, can input a static PET image of the subject as an input image into the convolutional neural network.

[0098] In the image processing device of an eleventh aspect, the CNN processing unit can input a random noise image as an input image into the convolutional neural network in the configuration of any of the first to sixth aspects.

[0099] The radiation tomography system according to the above embodiment comprises a radiation tomography device with a plurality of detectors arranged around a measurement chamber in which a subject, into which an RI source is injected, is placed and which collect coincidence information; and the image processing device which has the above configuration and which generates the tomographic image of the subject based on the coincidence information collected by the radiation tomography device.

[0100] The image processing method of a first aspect according to the above embodiment is an image processing method for generating a tomographic image of a subject based on coincidence information collected by a radiation tomography device comprising a plurality of detectors arranged around a measurement space in which the subject, into which an RI source is injected, is placed.and comprises (1) an image reconstruction step to generate a reconstructed image using an analytical image reconstruction method or a statistical image reconstruction method based on the coincidence information collected by the radiation tomography device; (2) a sinogram generation step to generate a sinogram based on the coincidence information collected by the radiation tomography device; (3) a CNN processing step to input an input image into a convolutional neural network and to generate an output image by the convolutional neural network; (4) a forward projection computation step to perform a forward projection computation on the output image to generate a sinogram; and (5) a CNN training step to train the convolutional neural network; and (6) in a first processing step,(a) in the CNN training step, a scoring function is used which contains an error scoring term that represents a scoring value with respect to an error between the reconstructed image and the original image, and the convolutional neural network is trained based on a value of the scoring function, and (b) the respective processes of the CNN processing step and the CNN training step are carried out in a multitude of repetitions, and (7) in a second processing step following the first processing step, (a) in the CNN training step, a scoring function is used which contains an error scoring term that represents a scoring value with respect to an error between the sinogram produced in the sinogram generation step and the sinogram produced in the forward projection computation step, and the convolutional neural network is trained based on a value of the scoring function,and (b) the initial image after repeated execution of the respective processes of the CNN processing step, the forward projection calculation and the CNN training step is defined as the tomographic image of the subject.

[0101] In the image processing procedure of a second aspect, in the configuration of the first aspect, in the sinogram generation step, the sinogram divided into a multitude of blocks can be generated based on the coincidence information collected by the radiation tomography device; in the forward projection calculation step, the forward projection calculation can be performed on the output image to generate the sinogram divided into a multitude of blocks; and in the second processing step, the CNN training step, the convolutional neural network can be trained based on the value of the evaluation function for each of the multitude of blocks.

[0102] In the image processing method of a third aspect, in the configuration of the first or second aspect, each of the tomographic images, the reconstructed image, the input image and the output image can be a three-dimensional image.

[0103] In the image processing procedure of a fourth aspect, the procedure in the configuration of any of the first to third aspects may further include a convolution integration step in which a convolution integration of a point spread function is performed on the output image, and in the forward projection calculation step, the forward projection calculation may be performed on the output image after a process of the convolution integration step has been carried out.

[0104] In the image processing procedure of a fifth aspect, in the configuration of any of the first to fourth aspects in the second processing step in the CNN training step, the error can be evaluated using the error evaluation term in an area in a sinogram space where the collection of coincidence information by the radiation tomography device is possible.

[0105] In the image processing procedure of a sixth aspect, in the configuration of any of the first to fifth aspects in the CNN training step, the evaluation function, which further includes a regularization term representing an evaluation value in terms of a difference of pixel values ​​between neighboring pixels in the output image, can be used, and the convolutional neural network can be trained based on the value of the evaluation function.

[0106] In the image processing procedure of a seventh aspect, in the configuration of any of the first to sixth aspects in the CNN processing step, an image representing morphological information of the subject can be entered as an input image into the convolutional neural network.

[0107] In the image processing procedure of an eighth aspect, in the configuration of any of the first to sixth aspects in the CNN processing step, an MRI image of the subject can be entered as an input image into the convolutional neural network.

[0108] In the image processing procedure of a ninth aspect, in the configuration of any of the first to sixth aspects in the CNN processing step, a CT image of the subject can be entered as an input image into the convolutional neural network.

[0109] In the image processing procedure of a tenth aspect, in the configuration of any of the first to sixth aspects in the CNN processing step, a static PET image of the subject can be entered as an input image into the convolutional neural network.

[0110] In the image processing method of an eleventh aspect, in the configuration of any of the first to sixth aspects in the CNN processing step, a random noise image can be entered as an input image into the convolutional neural network. Industrial applicability

[0111] The present invention can be used as an image processing device and image processing method that are able to reduce the number of CNN training epochs by using a DIP technique and to obtain a tomographic image in a short time in which the noise is reduced when the tomographic image of a subject is generated by training a CNN based on an evaluation result of an error between a calculated sinogram and a measured sinogram. List of reference symbols

[0112] 1 - Radiation tomography system, 2 - Radiation tomography device, 10 - Image processing device, 11 - Sinogram generation unit, 12 - CNN processing unit, 13 - Convolution integration unit, 14 - Forward projection calculation unit, 15 - CNN training unit, 16 - Storage unit, 17 - Image reconstruction unit. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited non-patent literature

[0000] F. Hashimoto, Y. Onishi, K. Ote, H. Tashima, and T. Yamaya, "Fully 3D Implementation of the End-to-end Deep Image Prior-based PET Image Reconstruction Using Block Iterative Algorithm," arXiv preprint, arXiv:2212.11844, doi: 10.48550 / arXiv.2212.11844, 2022

[0010] J. Nuyts et al., “A concave prior penalizing relative differences for maximum-a-posteriori reconstruction in emission tomography,” IEEE TNS, Volume 49, Issue 1, pp. 56-60, 2002

[0010] Hiroyuki Kudo, “Image reconstruction methods in low-dose CT: Fundamentals of statistical image reconstruction, iterative image reconstruction and compressed scanning”, Medical Imaging Technology, Volume 32, No. 4, pp. 239-248, 2014

[0010] Antonin Chambolle, “An algorithm for minimizing total variation and applications”, Journal of Mathematical Imaging and Vision 20, pp. 89-97, 2004

[0010] https: / / brainweb.bic.mni.mcgill.ca / brainweb /

[0080]

Claims

[1] Image processing device for generating a tomographic image of a subject based on coincidence information collected by a radiation tomography device comprising a plurality of detectors arranged around a measurement space in which the subject, into which an RI source is injected, is placed, wherein the image processing device comprises: a sinogram generation unit for generating a sinogram based on the coincidence information collected by the radiation tomography device; a CNN processing unit for inputting an input image into a convolutional neural network and for generating an output image through the convolutional neural network; a forward projection calculation unit for performing a forward projection calculation on the source image in order to generate a sinogram; and a CNN training unit for training the convolutional neural network, where in a first processing step The CNN training unit uses a scoring function that contains an error scoring term, which represents a score value in terms of an error between the reconstructed image and the original image, and trains the convolutional neural network based on a value of the scoring function, and The respective processes of the CNN processing unit and the CNN training unit are repeated in a large number of cases, and in a second processing step, which follows the first processing step, The CNN training unit uses a scoring function that includes an error scoring term, which represents a score value in terms of an error between the sinogram generated by the sinogram generation unit and the sinogram generated by the forward projection calculation unit, and trains the convolutional neural network based on a value of the scoring function, and The initial image, after the respective processes of the CNN processing unit, the forward projection calculation unit, and the CNN training unit, is repeated in a multitude of cases and defined as a tomographic image of the object. [2] Image processing device according to claim 1, wherein The sinogram generation unit generates the sinogram, which is divided into a multitude of blocks, based on the coincidence information collected by the radiation tomography device. The forward projection calculation unit performs the forward projection calculation for the source image to generate the sinogram, which is divided into a multitude of blocks, and In the second processing step, the CNN training unit trains the convolutional neural network based on the value of the evaluation function for each of the multiple blocks. [3] Image processing device according to claim 1 or 2, wherein each of the tomographic images, the reconstructed image, the input image and the output image is a three-dimensional image. [4] Image processing device according to one of claims 1 to 3, further comprising a convolution integration unit for performing a convolution integration of a point spread function on the output image, wherein the forward projection calculation unit performs the forward projection calculation for the output image after a process of the convolution integration unit has been carried out. [5] Image processing device according to one of claims 1 to 4, wherein in the second processing step the CNN training unit evaluates the error using the error evaluation expression in a region in a sinogram space in which the collection of coincidence information by the radiation tomography device is possible. [6] Image processing device according to any one of claims 1 to 5, wherein the CNN training unit uses the evaluation function which further comprises a regularization term which represents an evaluation value which is related to a difference of pixel values ​​between adjacent pixels in the output image, and trains the convolutional neural network based on the value of the evaluation function. [7] Image processing device according to one of claims 1 to 6, wherein the CNN processing unit inputs an image representing morphological information of the subject as an input image into the convolutional neural network. [8] Image processing device according to one of claims 1 to 6, wherein the CNN processing unit inputs an MRI image of the subject as an input image into the convolutional neural network. [9] Image processing device according to any one of claims 1 to 6, wherein the CNN processing unit inputs a CT image of the subject as an input image into the convolutional neural network. [10] Image processing device according to any one of claims 1 to 6, wherein the CNN processing unit inputs a static PET image of the subject as an input image into the convolutional neural network. [11] Image processing device according to one of claims 1 to 6, wherein the CNN processing unit inputs a random noise image as an input image into the convolutional neural network. [12] Radiation tomography apparatus, comprising: a radiation tomography device with a multitude of detectors arranged around a measurement chamber in which a subject, into which an RI source is injected, is placed, and to collect coincidence information; and the image processing device according to one of claims 1 to 11 for generating the tomographic image of the subject based on the coincidence information collected by the radiation tomography device. [13] Image processing method for generating a tomographic image of a subject based on coincidence information collected by a radiation tomography device comprising a plurality of detectors arranged around a measurement space in which the subject, into which an RI source is injected, is placed, wherein the image processing method comprises: an image reconstruction step to generate a reconstructed image using an analytical image reconstruction method or a statistical image reconstruction method based on the coincidence information collected by the radiation tomography device; a sinogram generation step to generate a sinogram based on the coincidence information collected by the radiation tomography device; a CNN processing step for inputting an input image into a convolutional neural network and for generating an output image through the convolutional neural network; a forward projection calculation step to perform a forward projection calculation on the source image in order to generate a sinogram; and a CNN training step to train the convolutional neural network, where in a first processing step, In the CNN training step, a scoring function is used that contains an error scoring term, which represents a score value in terms of an error between the reconstructed image and the original image, and the convolutional neural network is trained based on a value of the scoring function, and The respective processes of the CNN processing step and the CNN training step are repeated in a large number of cases, and in a second processing step, which follows the first processing step, In the CNN training step, a scoring function is used that contains an error scoring term, which represents a scoring value in terms of an error between the sinogram generated in the sinogram generation step and the sinogram generated in the forward projection computation step, and the convolutional neural network is trained based on a value of the scoring function, and The initial image after the respective processes of the CNN processing step, the forward projection calculation and the CNN training step, which are performed repeatedly in a large number of cases, is defined as a tomographic image of the subject. [14] Image processing method according to claim 13, wherein In the sinogram generation step, the sinogram, divided into a multitude of blocks, is generated based on the coincidence information collected by the radiation tomography device. In the forward projection calculation step, the forward projection calculation is performed on the source image to generate the sinogram divided into a multitude of blocks, and In the second processing step, the CNN training step, the convolutional neural network is trained based on the value of the evaluation function for each of the number of blocks. [15] Image processing method according to claim 13 or 14, wherein each of the tomographic images, the reconstructed image, the input image and the output image is a three-dimensional image. [16] Image processing method according to one of claims 13 to 15, further comprising a convolution integration step in which a point spread function is convolved onto the output image, wherein in the forward projection calculation step the forward projection calculation onto the output image is performed after a process of the convolution integration step has been carried out. [17] Image processing method according to one of claims 13 to 16, wherein in the second processing step, the CNN training step, the error is evaluated using the error evaluation term in an area in a sinogram in which the collection of coincidence information by the radiation tomography device is possible. [18] Image processing method according to one of claims 13 to 17, wherein in the CNN training step the evaluation function, which further comprises a regularization term that represents an evaluation value with respect to a difference of pixel values ​​between neighboring pixels in the source image, is used and the convolutional neural network is trained based on the value of the evaluation function. [19] Image processing method according to one of claims 13 to 18, wherein in the CNN processing step an image representing morphological information of the subject is input as an input image into the convolutional neural network. [20] Image processing method according to one of claims 13 to 18, wherein in the CNN processing step an MRI image of the subject is input as an input image into the convolutional neural network. [21] Image processing method according to one of claims 13 to 18, wherein in the CNN processing step a CT image of the subject is entered as an input image into the convolutional neural network. [22] Image processing method according to one of claims 13 to 18, wherein in the CNN processing step a static PET image of the subject is input as an input image into the convolutional neural network. [23] Image processing method according to one of claims 13 to 18, wherein in the CNN processing step a random noise image is input as an input image into the convolutional neural network.

Images