Image processing device and image processing method

Abstract

To provide an image processing device which can easily reduce noise in an object image by performing teacherless learning after pre-learning with teacher to a CNN.SOLUTION: An image processing device 1 comprises an input image creation unit 10, a first calculation unit 20 and a second calculation unit 30, and creates a noise reduction image by reducing noise in an object image 52. The first calculation unit 20 includes a first CNN processing part 21 and a first CNN learning part 22 and performs processing of pre-learning with teacher. A first input image 40 is obtained by changing a pixel value of a partial region on the basis of a teacher image 42. The second calculation unit 30 includes a second CNN processing part 31 and a second CNN learning part 32 and performs processing of teacherless learning.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an image processing device and an image processing method for reducing noise in a target image to create a noise-reduced image. [Background technology]

[0002] Images may contain noise. An example of a noisy image is a tomographic image of a subject reconstructed based on information acquired by a radiation tomography device. Radiation tomography devices include PET (Positron Emission Tomography) and SPECT (Single Photon Emission Computed Tomography) devices.

[0003] A PET system has a detector section with many small radiation detectors arranged around the measurement space in which the subject is placed. The detector section detects photon pairs with an energy of 511 keV, which are generated when electrons and positrons annihilate in a subject administered with a positron-emitting isotope (RI source), using coincidence counting to collect this coincidence information. Based on the collected coincidence information, a tomographic image can be reconstructed that shows the spatial distribution of the frequency of photon pair occurrence in the measurement space (i.e., the spatial distribution of the RI source). PET systems play an important role in fields such as nuclear medicine, and can be used to study, for example, biological functions and higher-level brain functions.

[0004] Reconstructed tomographic images of a subject contain a large amount of statistical noise, making it necessary to reduce the noise in these images. In addition to PET (tomographic) images, there are also cases where it is necessary to reduce noise from noisy target images to create noise-reduced images. Various noise reduction techniques are known. Among these, a technique that reduces noise using Deep Image Prior technology, which uses a convolutional neural network (CN), a type of deep neural network, has attracted attention. Hereinafter, Convolutional Neural Network (CNN) is referred to as "CNN," and Deep Image Prior technology is referred to as "DIP technology." DIP technology utilizes the property of CNNs that meaningful structures in an image are learned faster than random noise (i.e., random noise is difficult to learn), enabling it to create noise-reduced images through unsupervised learning.

[0005] Non-Patent Document 1 describes a technique for reducing noise in a target image (PET image) using the DIP technique. The technique described in this document divides the noise reduction process into two steps. In the first step, for each of a plurality of pairs of a first input image (MRI image) and a teacher image (PET image), the first input image is input to a CNN to generate a first output image, and the CNN is trained based on an evaluation result of the error between the first output image and the teacher image (supervised pre-training). In the subsequent second step, a second input image (MRI image) is input to the CNN pre-trained in the first step to generate a second output image, and the CNN is further trained based on an evaluation result of the error between the second output image and the target image (PET image) (unsupervised training). The second output image is then a noise-reduced image in which the noise in the target image (PET image) has been reduced.

[0006] In short, the noise reduction technology described in Non-Patent Document 1 creates a noise-reduced image in which the noise of the target image is reduced by training the CNN to convert a first input image into a teacher image in a first step (supervised pre-training), and further training the CNN in a second step (unsupervised training). By doing this, the noise reduction technology described in Non-Patent Document 1 is said to be able to improve noise reduction performance compared to the case of DIP technology alone (unsupervised learning only). [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] J. Cui et al. Populational and individual information based PET image denoising using conditional unsupervised learning. Phys. Med. Biol. 66:155001, 2021. Summary of the Invention [Problem to be solved by the invention]

[0008] The noise reduction technology described in Non-Patent Document 1 requires the preparation of not only a large number of teacher images but also a large number of first input images in order to perform supervised pre-learning in the first step. However, it is not easy to prepare a large number of first input images separately from the preparation of a large number of teacher images.

[0009] The present invention has been made to solve the above problems, and aims to provide an image processing device and an image processing method that can easily reduce noise in a target image by performing supervised pre-learning on a CNN followed by unsupervised learning. [Means for solving the problem]

[0010] The image processing device of the present invention is an image processing device that reduces noise in a target image to create a noise-reduced image. A first aspect of the image processing device of the present invention includes: (1) a first CNN processing unit that inputs a first input image, in which pixel values ​​of a partial region are changed based on a teacher image, into a convolutional neural network to create a first output image using the convolutional neural network; (2) a first CNN training unit that evaluates an error between the first output image and the teacher image and trains the convolutional neural network based on the error evaluation result; (3) a second CNN processing unit that inputs a second input image into the convolutional neural network to create a second output image using the convolutional neural network; and (4) a second CNN training unit that evaluates an error between the second output image and a target image and trains the convolutional neural network based on the error evaluation result. Then, after repeatedly performing processing by the first CNN processing unit and the first CNN training unit multiple times for each of a plurality of pairs of the teacher image and the first input image, the second CNN processing unit and the second CNN training unit each perform processing multiple times to create a noise-reduced second output image.

[0011] The image processing device of the present invention may have the following configuration. In the second aspect, in addition to the first aspect, it is preferable that the image processing device further comprises an input image creation unit that creates a first input image in which pixel values ​​of a partial area are changed based on the teacher image. In a third aspect, in addition to the first or second aspect, it is preferable that the convolutional neural network has a U-net structure including an encoder and a decoder, the first CNN training unit trains both the encoder and the decoder of the convolutional neural network, and the second CNN training unit selectively trains the decoder of the convolutional neural network. In a fourth aspect, in addition to any one of the first to third aspects, it is preferable that the first CNN processing unit inputs, for each teacher image, a plurality of first input images each having a different manner of pixel value change to the neural network, and creates a first output image for each first input image by the convolutional neural network. In a fifth aspect, in addition to any one of the first to fourth aspects, it is preferable that the first CNN processing unit inputs a first input image, in which pixel values ​​of a part of the noise-reduction target area are changed based on the teacher image, to the convolutional neural network. In a sixth aspect, in addition to any of the first to fifth aspects, it is preferable that the target image and the teaching image are tomographic images of the subject reconstructed based on information acquired by a radiation tomography apparatus.

[0012] The image processing method of the present invention is an image processing method for reducing noise in a target image to create a noise-reduced image. A first aspect of the image processing method of the present invention includes: (1) a first CNN processing step of inputting a first input image, in which pixel values ​​of a partial region are changed based on a teacher image, into a convolutional neural network to generate a first output image using the convolutional neural network; (2) a first CNN learning step of evaluating an error between the first output image and the teacher image and training the convolutional neural network based on the error evaluation result; (3) a second CNN processing step of inputting a second input image into the convolutional neural network to generate a second output image using the convolutional neural network; and (4) a second CNN learning step of evaluating an error between the second output image and a target image and training the convolutional neural network based on the error evaluation result. Then, the first CNN processing step and the first CNN learning step are each repeated multiple times for each of multiple pairs of the teacher image and the first input image, and then the second CNN processing step and the second CNN learning step are each repeated multiple times to generate a second output image as a noise-reduced image.

[0013] The image processing method of the present invention may be configured as follows. In the second aspect, in addition to the first aspect, it is preferable that the image processing method further comprises an input image creating step of creating a first input image in which pixel values ​​of a partial area are changed based on the teacher image. In a third aspect, in addition to the first or second aspect, the convolutional neural network has a U-net structure including an encoder and a decoder, and it is preferable that in the first CNN learning step, both the encoder and the decoder of the convolutional neural network are trained, and in the second CNN learning step, the decoder of the convolutional neural network is selectively trained. In a fourth aspect, in addition to any one of the first to third aspects, in the first CNN processing step, for each teacher image, it is preferable to input a plurality of first input images having mutually different pixel value change modes into a neural network, and to create a first output image for each first input image by a convolutional neural network. In a fifth aspect, in addition to any of the first to fourth aspects, in the first CNN processing step, it is preferable that a first input image in which pixel values ​​of a part of the noise reduction target area are changed based on a teacher image is input to the convolutional neural network. In a sixth aspect, in addition to any of the first to fifth aspects, it is preferable that the target image and the teaching image are tomographic images of the subject reconstructed based on information acquired by a radiation tomography apparatus. [Effects of the Invention]

[0014] According to the present invention, an input image to a convolutional neural network (CNN) used in supervised pre-learning can be easily created based on a supervised image, and noise in a target image can be easily reduced by performing unsupervised learning on the CNN after supervised pre-learning. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is a diagram showing the configuration of an image processing device 1. As shown in FIG. [Figure 2] FIG. 2 is a flowchart of the image processing method. [Figure 3] FIG. 3 is a diagram illustrating an example of the configuration of a CNN. [Figure 4]FIG. 4 shows an input image (MRI image) to the CNN. [Figure 5] FIG. 5 is a diagram showing a phantom image (correct image). [Figure 6] FIG. 6 is a diagram showing a tomographic image (target image) simulating a head having a tumor and administered with 18F-FDG as a drug. [Figure 7] FIG. 7 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 1. As shown in FIG. [Figure 8] FIG. 8 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 2. As shown in FIG. [Figure 9] FIG. 9 is a diagram showing a tomographic image after noise reduction processing by the image processing method of the embodiment. [Figure 10] FIG. 10 is a table summarizing the PSNR, SSIM, and CNR values ​​of the tomographic images after noise reduction processing using the image processing methods of the example and comparative examples 1 and 2. [Figure 11] FIG. 11 is a diagram showing an input image (MRI image) to the CNN. [Figure 12] FIG. 12 shows a tomographic image (target image) of the head to which 18F-AV45 was administered as a drug. [Figure 13] FIG. 13 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 1. As shown in FIG. [Figure 14] FIG. 14 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 2. As shown in FIG. [Figure 15] FIG. 15 is a diagram showing a tomographic image after noise reduction processing by the image processing method of the embodiment. [Figure 16] FIG. 16 is a diagram showing an input image (MRI image) to the CNN. [Figure 17] FIG. 17 shows a tomographic image (target image) of the head to which 11C-PIB was administered as a drug. [Figure 18] FIG. 18 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 1. As shown in FIG. [Figure 19]FIG. 19 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 2. As shown in FIG. [Figure 20] FIG. 20 is a diagram showing a tomographic image after noise reduction processing by the image processing method of the embodiment. [Figure 21] FIG. 21 is a diagram showing an input image (MRI image) to the CNN. [Figure 22] FIG. 22 is a diagram showing a tomographic image (target image) of the head to which 18F-FDG was administered as a drug. [Figure 23] FIG. 23 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 1. As shown in FIG. [Figure 24] FIG. 24 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 2. As shown in FIG. [Figure 25] FIG. 25 is a diagram showing a tomographic image after noise reduction processing by the image processing method of the embodiment. [Figure 26] FIG. 26 is a table summarizing the CNR values ​​of the tomographic images after noise reduction processing using the image processing methods of the example and comparative examples 1 and 2. DETAILED DESCRIPTION OF THE INVENTION

[0016] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same elements are designated by the same reference numerals, and duplicate explanations will be omitted. The present invention is not limited to these examples, but is defined by the claims, and is intended to include all modifications within the meaning and scope equivalent to the claims.

[0017] 1 is a diagram showing the configuration of an image processing device 1. The image processing device 1 includes an input image creation unit 10, a first calculation unit 20, and a second calculation unit 30, and reduces noise in a target image 52 to create a noise-reduced image.

[0018] In this figure, a PET image of the head is shown as an example of each of the first input image 40, first output image 41, teacher image 42, second output image 51, and target image 52, and an MRI image of the head is shown as an example of the second input image 50. The following description will be based on this premise. However, this embodiment is not limited to this. This embodiment is suitable for reducing noise in a target image that is a tomographic image (e.g., a PET image) of a subject reconstructed based on information acquired by a radiation tomography apparatus. These images may be two-dimensional images or three-dimensional images.

[0019] The image processing device 1 includes a GPU (Graphics Processing Unit) that performs processing using CNN, an input unit (e.g., a keyboard or mouse) that accepts input from an operator, a display unit (e.g., a liquid crystal display) that displays images, etc., and a storage unit that stores programs and data for executing various processes. A computer having a CPU, RAM, ROM, a hard disk drive, etc. is used as the image processing device 1.

[0020] The input image creation unit 10 creates a first input image 40 by changing pixel values ​​in a partial region based on the teacher image 42.

[0021] The first calculation unit 20 includes a first CNN processing unit 21 and a first CNN learning unit 22, and performs supervised pre-learning processing. The first CNN processing unit 21 inputs a first input image 40 to the CNN, and creates a first output image 41 using the CNN. The first CNN learning unit 22 evaluates the error between the first output image 41 and a teacher image 42, and trains the CNN based on the error evaluation result. The first calculation unit 20 repeatedly performs the processing of the first CNN processing unit 21 and the first CNN learning unit 22 multiple times for each of multiple pairs of teacher image 42 and first input image 40.

[0022] The second calculation unit 30 includes a second CNN processing unit 31 and a second CNN training unit 32, and performs unsupervised learning processing. The second CNN processing unit 31 inputs a second input image 50 to the CNN, and creates a second output image 51 using the CNN. The second CNN training unit 32 evaluates the error between the second output image 51 and the target image 52, and trains the CNN based on the error evaluation result. After the iterative processing of the first calculation unit 20 is completed, the second calculation unit 30 sets the CNN training state at the time of completion as the initial value and repeats the processing of the second CNN processing unit 31 and the second CNN training unit 32 multiple times. Then, the second output image 51 at the time of completion of the iterative processing of the second calculation unit 30 is designated as a noise-reduced image.

[0023] 2 is a flowchart of the image processing method, which reduces noise in target image 52 and creates a noise-reduced image by sequentially performing input image creation step S10, first calculation step S20, and second calculation step S30.

[0024] The input image creation step S10 is a process performed by the input image creation unit 10. In the input image creation step S10, a first input image 40 is created by changing pixel values ​​in a partial area based on the teacher image 42.

[0025] The first calculation step S20 is a supervised pre-learning process performed by the first calculation unit 20, and includes a first CNN processing step S21 and a first CNN learning step S22. The first CNN processing step S21 is a process performed by the first CNN processing unit 21. The first CNN learning step S22 is a process performed by the first CNN learning unit 22. In the first CNN processing step S21, a first input image 40 is input to the CNN, and a first output image 41 is created by the CNN. In the first CNN learning step S22, the error between the first output image 41 and a teacher image 42 is evaluated, and the CNN is trained based on the error evaluation result. In the first calculation step S20, the processes of the first CNN processing step S21 and the first CNN learning step S22 are each repeated multiple times for each of multiple pairs of teacher image 42 and first input image 40.

[0026] The second calculation step S30 is an unsupervised learning process performed by the second calculation unit 30 and includes a second CNN processing step S31 and a second CNN training step S32. The second CNN processing step S31 is a process performed by the second CNN processing unit 31. The second CNN training step S32 is a process performed by the second CNN training unit 32. In the second CNN processing step S31, a second input image 50 is input to the CNN, and a second output image 51 is generated by the CNN. In the second CNN training step S32, the error between the second output image 51 and the target image 52 is evaluated, and the CNN is trained based on the error evaluation result. In the second calculation step S30, after the iterative process in the first calculation step S20 is completed, the CNN training state at the time of completion is set as the initial value, and the processes in the second CNN processing step S31 and the second CNN training step S32 are each repeated multiple times. Then, the second output image 51 at the time of completion of the iterative process in the second calculation step S30 is designated as the noise-reduced image.

[0027] FIG. 3 is a diagram showing an example of the configuration of a CNN. As shown in this diagram, the CNN preferably has a U-net structure including an encoder and a decoder. This diagram shows the size of each layer of the CNN, with the number of pixels of the input image input to the CNN being N×N×64. In the first CNN training step S22, the first CNN training unit 22 trains both the encoder and decoder of the CNN. In the second CNN training step S32, the second CNN training unit 32 may train both the encoder and decoder of the CNN, but it is preferable to selectively train the decoder of the CNN encoder and decoder to shorten the time required for training.

[0028] Any function may be used to evaluate the error between two images in the first CNN training unit 22 in the first CNN training step S22 and in the second CNN training unit 32 in the second CNN training step S32. The error evaluation function may be, for example, the L1 norm, the L2 norm, or the negative logarithmic likelihood in a Poisson distribution.

[0029] In the image processing device 1 or the image processing method, each image is as follows. The target image 52 is an image for which noise is to be reduced, and in this case is a PET image of the head. Such a PET image generally consists of an image region of the head, which is the region to be subjected to noise reduction (noise reduction target region), and a background region surrounding the head region.

[0030] The teacher image 42 is a PET image of the head, similar to the target image 52. A plurality of teacher images 42 may be prepared, or a single teacher image 42 may be prepared. The target image 52 itself may be used as the teacher image 42.

[0031] The first input image 40 is obtained by changing pixel values ​​of a partial region based on the teacher image 42, preferably by changing pixel values ​​of a partial region of the noise reduction target region based on the teacher image 42. Multiple first input images 40 may be prepared for each teacher image 42. The pixel values ​​of the first input image 40 relative to the teacher image 42 can be changed in any manner. The partial regions whose pixel values ​​are changed may have any shape, size, and number. The change of pixel values ​​in the partial regions may be, for example, a nonlinear conversion of pixel values, an exchange of pixel values ​​between multiple pixels, replacement with a constant pixel value, replacement with a random pixel value, etc. It is preferable that the manner of pixel value change (e.g., the shape, size, and number of partial regions whose pixel values ​​are changed, and the method of pixel value change) differ among the multiple first input images 40. In the input image creation step S10, the input image creation unit 10 creates such first input images 40 based on the teacher image 42.

[0032] The second input image 50 may be an image representing morphological information of the subject, such as an MRI image as shown in Fig. 1, a CT image, or a static PET image. The second input image 50 may also be a random noise image.

[0033] Next, target images were created using a Monte Carlo simulation of a head PET scanner using digital brain phantom images, and the results of noise reduction processing using these target images and the image processing methods used in Example and Comparative Examples 1 and 2 are described with reference to Figures 4 to 10. The phantom images were obtained from BrainWeb (https: / / brainweb.bic.mni.mcgill.ca / brainweb / ).

[0034] The image processing method of the example is based on the above embodiment. The image processing method of the comparative example 1 is based on the conventional DIP technology. The image processing method of the comparative example 2 is based on the technology described in Non-Patent Document 1.

[0035] In the supervised pre-learning in each of the image processing methods of the Example and Comparative Example 2, 20 teacher images created based on phantom images were used. In the image processing method of the Example, 32 first input images created by varying pixel value change modes for each teacher image were used. In the image processing method of Comparative Example 2, MRI images were used as input images to the CNN.

[0036] In the unsupervised learning in the image processing methods of the Example and Comparative Examples 1 and 2, the target image has a tumor and a drug. 18 We used tomographic images of a simulated head injected with F-FDG, and MRI images as input images to the CNN.

[0037] 4 to 9 are diagrams showing tomographic images used or created in unsupervised learning in the image processing methods of Example and Comparative Examples 1 and 2. FIG. 4 is a diagram showing an input image (MRI image) to CNN. FIG. 5 is a diagram showing a phantom image (correct image). FIG. 6 is a diagram showing a tomographic image of a subject having a tumor and treated with a drug. 184 to 9, (a) is a diagram showing a tomographic image (target image) simulating a head into which F-FDG has been administered. FIG. 7 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 1. FIG. 8 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 2. FIG. 9 is a diagram showing a tomographic image after noise reduction processing by the image processing method of the Example. In each of FIGS. 4 to 9, (a) is an image of a transverse section, (b) is an image of a coronal section, and (c) is an image of a sagittal section.

[0038] FIG. 10 is a table summarizing the PSNR, SSIM, and CNR values ​​of the tomographic images after noise reduction processing using the image processing methods of the Example and Comparative Examples 1 and 2. PSNR (Peak Signal to Noise Ratio) represents image quality in decibels (dB). SSIM (structural similarity index) is an index that quantifies changes in the brightness, contrast, and structure of an image. CNR (Contrast to Noise Ratio) is the ratio of image contrast to noise. For all of PSNR, SSIM, and CNR, the larger the value, the better the image quality. As shown in FIGS. 4 to 10, the values ​​of PSNR, SSIM, and CNR were all larger in the Example compared to Comparative Examples 1 and 2, confirming that noise in the target image could be reduced more effectively.

[0039] Next, the results of noise reduction processing performed using clinical data by the image processing methods of Example and Comparative Examples 1 and 2 will be described with reference to FIGS.

[0040] In the supervised pre-learning in the image processing methods of Example and Comparative Example 2, 24 tomographic images of clinical data were used as training images. 18 This is a PET image of the head after administration of F-AV45. In the image processing method of the example, 32 first input images were used, each created by varying the pixel value change mode for each training image. In the image processing method of comparative example 2, an MRI image was used as the input image to the CNN.

[0041] In the unsupervised learning in the image processing methods of the Example and Comparative Examples 1 and 2, the target image was a drug. 18 F-AV45, 11 C-PIB and 18 We used tomographic images of the head after administration of either F-FDG or MRI images as input images to the CNN.

[0042] Figures 11 to 15 each show 18 11 shows the tomographic images used or created in unsupervised learning in the image processing methods of Example and Comparative Examples 1 and 2, respectively, when a tomographic image of a head to which F-AV45 was administered was used as the target image. FIG. 12 shows the input image (MRI image) to the CNN. 18 11 to 15, (a) is an image of a cross-sectional view of a head after administration of F-AV45 (target image). FIG. 13 is a view showing a cross-sectional view after noise reduction processing by the image processing method of Comparative Example 1. FIG. 14 is a view showing a cross-sectional view after noise reduction processing by the image processing method of Comparative Example 2. FIG. 15 is a view showing a cross-sectional view after noise reduction processing by the image processing method of the Example. In each of FIGS. 11 to 15, (a) is an image of a transverse section, (b) is an image of a coronal section, and (c) is an image of a sagittal section.

[0043] Figures 16 to 20 each show 11 16 shows the tomographic images used or created in unsupervised learning in the image processing methods of Example and Comparative Examples 1 and 2, respectively, when a tomographic image of a head to which C-PIB was administered was used as the target image. FIG. 17 shows the input image (MRI image) to the CNN. 1116A to 16C are diagrams showing tomographic images (target images) of the head after administration of C-PIB. FIG. 18 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 1. FIG. 19 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 2. FIG. 20 is a diagram showing a tomographic image after noise reduction processing by the image processing method of the example. In each of FIGS. 16A to 16C, (a) is an image of a transverse section, and (b) is an image of a sagittal section.

[0044] Figures 21 to 25 each show 18 21 shows the tomographic images used or created in unsupervised learning in the image processing methods of Example and Comparative Examples 1 and 2, respectively, when a tomographic image of a head into which F-FDG was administered was used as the target image. FIG. 22 shows the input image (MRI image) to the CNN. 18 21 to 25 are diagrams showing tomographic images (target images) of the head into which F-FDG was administered. FIG. 23 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 1. FIG. 24 is a diagram showing a tomographic image after noise reduction processing by the image processing method of Comparative Example 2. FIG. 25 is a diagram showing a tomographic image after noise reduction processing by the image processing method of the example. In each of FIGS. 21 to 25, (a) is an image of a transverse section, and (b) is an image of a sagittal section.

[0045] 26 is a table summarizing the CNR values ​​of the tomographic images after noise reduction processing by the image processing methods of Example and Comparative Examples 1 and 2. 18 F-AV45, 11 C-PIB and 18 The CNR values ​​are shown for each F-FDG administered to the head.

[0046] 11 to 26 reveal the following: Compared to Comparative Examples 1 and 2, the CNR values ​​in the Examples were larger, and it was confirmed that noise in the target image could be reduced more effectively. 18When a tomographic image obtained when F-FDG was administered to the head was used as the target image, the degree of improvement in CNR was small in Comparative Example 2 compared to Comparative Example 1, whereas the degree of improvement in CNR was greater in the Examples compared to Comparative Examples 1 and 2. Thus, when the types of drugs administered to the head differ between the teacher image and the target image, it was confirmed that the degree of noise removal in Comparative Example 2 was similar to that of Comparative Example 1, whereas the degree of noise removal in the Examples was improved compared to Comparative Examples 1 and 2.

[0047] As described above, according to this embodiment, input images to be used in supervised pre-training of CNN can be easily generated based on teacher images, and therefore noise in the target image can be easily reduced by performing unsupervised training after supervised pre-training of CNN. Furthermore, even if the types of drugs administered to the subject differ between the teacher image and the target image, noise in the target image can be effectively reduced. [Explanation of symbols]

[0048] 1...image processing device, 10...input image creation unit, 20...first calculation unit, 21...first CNN processing unit, 22...first CNN learning unit, 30...second calculation unit, 31...second CNN processing unit, 32...second CNN learning unit.

Claims

1. An image processing device that reduces noise in a target image to create a noise-reduced image, a first CNN processing unit that inputs a first input image, in which pixel values ​​of a partial region are changed based on a teacher image, into a convolutional neural network and creates a first output image by the convolutional neural network; a first CNN learning unit that evaluates an error between the first output image and the teacher image and trains the convolutional neural network based on the error evaluation result; a second CNN processing unit that inputs a second input image to the convolutional neural network and creates a second output image by the convolutional neural network; a second CNN learning unit that evaluates an error between the second output image and the target image and trains the convolutional neural network based on the error evaluation result; Equipped with After repeatedly performing the processing of the first CNN processing unit and the first CNN learning unit a plurality of times for each of a plurality of pairs of the teacher image and the first input image, the processing of the second CNN processing unit and the second CNN learning unit a plurality of times is repeatedly performed to set the second output image as the noise-reduced image. Image processing device.

2. The image processing device further includes an input image creating unit that creates a first input image by changing pixel values ​​of a partial region based on the teacher image. The image processing device according to claim 1 .

3. The convolutional neural network has a U-net structure including an encoder and a decoder, The first CNN training unit trains both an encoder and a decoder of the convolutional neural network; The second CNN learning unit selectively learns a decoder of the convolutional neural network. The image processing device according to claim 1 .

4. The first CNN processing unit inputs a plurality of first input images, each having a different manner of pixel value change, to the neural network for each teacher image, and creates a first output image for each first input image using the convolutional neural network. The image processing device according to claim 1 .

5. The first CNN processing unit inputs a first input image, in which pixel values ​​of a part of a noise reduction target area are changed based on a teacher image, to the convolutional neural network. The image processing device according to claim 1 .

6. the target image and the teacher image are tomographic images of a subject reconstructed based on information acquired by a radiation tomography apparatus; The image processing device according to claim 1 .

7. 1. An image processing method for reducing noise in a target image to create a noise-reduced image, comprising: a first CNN processing step of inputting a first input image, in which pixel values ​​of a partial region have been changed based on a teacher image, into a convolutional neural network, and generating a first output image by the convolutional neural network; a first CNN learning step of evaluating an error between the first output image and the teacher image and learning the convolutional neural network based on the error evaluation result; a second CNN processing step of inputting a second input image to the convolutional neural network and generating a second output image by the convolutional neural network; a second CNN learning step of evaluating an error between the second output image and the target image and learning the convolutional neural network based on the error evaluation result; Equipped with After repeatedly performing the first CNN processing step and the first CNN learning step a plurality of times for each of a plurality of pairs of the teacher image and the first input image, the second CNN processing step and the second CNN learning step are repeatedly performed a plurality of times to set the second output image as the noise-reduced image. Image processing methods.

8. The method further includes an input image creating step of creating a first input image in which pixel values ​​of a partial region are changed based on the teacher image. The image processing method according to claim 7.

9. The convolutional neural network has a U-net structure including an encoder and a decoder, In the first CNN training step, both the encoder and the decoder of the convolutional neural network are trained; In the second CNN training step, a decoder of the convolutional neural network is selectively trained. The image processing method according to claim 7.

10. In the first CNN processing step, for each teacher image, a plurality of first input images having different pixel value change modes are input to the neural network, and a first output image is created for each first input image by the convolutional neural network. The image processing method according to claim 7.

11. In the first CNN processing step, a first input image in which pixel values ​​of a part of a noise reduction target area are changed based on a teacher image is input to the convolutional neural network. The image processing method according to claim 7.

12. the target image and the teacher image are tomographic images of a subject reconstructed based on information acquired by a radiation tomography apparatus; The image processing method according to claim 7.

Images