Attenuation distribution image creating device, image processing device, radiation computed tomography system, attenuation distribution image creating method, and image processing method

Abstract

An image processing device 20 comprises an image reconstruction unit 21, an image conversion unit 22, a CNN processing unit 23, and an attenuation correction unit 24. The image conversion unit 22 generates converted images in which the pixel value of a partial region has been changed on the basis of tomographic images for respective slice positions of a subject. In the CNN processing unit 23, the converted images for the slice positions generated by the image conversion unit 22 are inputted to a CNN 31, and slice position information of the converted images is inputted to an embedding layer 32. Output data from the embedding layer 32 is added to a feature map of the CNN 31. The CNN processing unit 23 outputs output images from the CNN 31 as attenuation distribution images for the subject. Thus, even when a medicine administered to a subject when acquiring tomographic images of an attenuation correction target is different from a medicine administered to the subject when acquiring tomographic images for CNN learning, a device capable of creating more accurate attenuation distribution images from the tomographic images of the attenuation correction target is realized.

Description

Attenuation distribution image creating device, image processing device, radiation tomography system, attenuation distribution image creating method, and image processing method

[0001] The present disclosure relates to an apparatus and method for creating an attenuation distribution image, and an apparatus and method for performing attenuation correction of a tomographic image using the attenuation distribution image.

[0002] Radiation tomography apparatuses capable of acquiring tomographic images of a subject (living body) include PET (Positron Emission Tomography) apparatuses and SPECT (Single Photon Emission Computed Tomography) apparatuses.

[0003] The PET device includes a detection unit having a large number of small radiation detectors arranged around a measurement space in which a subject is placed. The PET device detects photon pairs with an energy of 511 keV generated in association with electron-positron annihilation in a subject into which a positron-emitting isotope (RI source) has been introduced, using a coincidence method with the detection unit, and collects this coincidence information. Then, based on the collected large amount of coincidence information, a tomographic image can be reconstructed that represents the spatial distribution of the occurrence frequency of photon pairs in the measurement space (i.e., the spatial distribution of the RI source).

[0004] In this case, list data in which coincidence counting information collected by the PET device is arranged in chronological order is divided into multiple frames in the order of collection, and image reconstruction processing is performed using the data contained in each frame of the list data, thereby obtaining a dynamic PET image consisting of cross-sectional images of multiple frames. PET devices 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.

[0005] Photons generated within the subject are absorbed and attenuated as they pass through the subject, and also as they pass through the bed on which the subject is placed. Therefore, attenuation correction is necessary to obtain more accurate tomographic images.

[0006] For example, a PET device is used to perform measurements (emission scans) on a subject administered with a drug containing an RI radiation source to obtain tomographic images (emission scan images), and measurements (transmission scans) are performed on the subject without the drug while rotating the RI radiation source around the subject to obtain tomographic images (transmission scan images).Then, attenuation correction of the emission scan images is performed using the transmission scan images.

[0007] For attenuation correction of emission scan images, transmission scan images obtained by performing a transmission scan using a PET device as described above can be used, as well as X-ray CT scan images obtained by performing a scan using an X-ray CT device.

[0008] Hereinafter, the attenuation distribution image at an energy of 511 keV obtained from the transmission scan image and the X-ray CT scan image will be collectively referred to as the attenuation distribution image. The attenuation distribution image represents the attenuation distribution of photons within the subject. Also, below, the emission scan image will be referred to as the tomographic image. The attenuation distribution image can be used to perform attenuation correction of the tomographic image.

[0009] To perform attenuation correction and obtain more accurate tomographic images, a PET device must have a transmission scanning mechanism or an X-ray CT scanning mechanism. This is one of the factors that hinders the simplification and cost reduction of PET devices. A technology intended to solve this problem is described in Non-Patent Document 1.

[0010] The technology described in Non-Patent Document 1 trains a convolutional neural network (CNN) using a large number of sets of training tomographic images and attenuation distribution images as teacher images, creates an attenuation distribution image from a tomographic image to be attenuation corrected using this trained CNN, and performs attenuation correction on the tomographic image using this attenuation distribution image. This technology enables attenuation correction by using the trained CNN even if the PET device does not have a transmission scan mechanism or an X-ray CT scan mechanism.

[0011] F. Hashimoto et al., "Deep learning-based attenuation correction for brain PET with various radiotracers", Annals of Nuclear Medicine, 2021

[0012] The present inventors have found that the technology described in Non-Patent Document 1 has the following problem: When the drug administered to the subject when the tomographic image to be attenuation corrected was acquired is the same as the drug administered to the subject when the tomographic image for CNN training was acquired, the attenuation distribution image created from the tomographic image to be attenuation corrected is relatively accurate.

[0013] However, if the drug administered to the subject when the tomographic image to be attenuation corrected was acquired is different from the drug administered to the subject when the tomographic image for CNN training was acquired, the accuracy of the attenuation distribution image created from the tomographic image to be attenuation corrected will be inferior.

[0014] The present invention aims to provide an apparatus and method that can create a more accurate attenuation distribution image from a tomographic image to be attenuation corrected, even if the drug administered to the subject when the tomographic image to be attenuation corrected was acquired is different from the drug administered to the subject when the tomographic image for CNN learning was acquired.

[0015] An embodiment of the present invention is an attenuation distribution image creation device that creates an attenuation distribution image of a subject based on a tomographic image of the subject, and includes: (1) an image conversion unit that generates a converted image by changing pixel values ​​of a partial region based on the tomographic image of each slice position of the subject, and (2) a CNN processing unit that inputs the converted image of each slice position generated by the image conversion unit to a convolutional neural network, inputs slice position information of the converted image to an embedding layer, adds output data from the embedding layer to a feature map of the convolutional neural network, and outputs the output image from the convolutional neural network as an attenuation distribution image of the subject.

[0016] An embodiment of the present invention is an image processing device that includes the attenuation distribution image creation device having the above-described configuration that creates an attenuation distribution image of a subject based on a tomographic image of the subject, and an attenuation correction unit that performs attenuation correction on the tomographic image of the subject using the attenuation distribution image of the subject.

[0017] An embodiment of the present invention is a radiation tomography system including: a radiation tomography apparatus for acquiring a tomographic image of a subject; an attenuation distribution image creation device configured as described above for creating an attenuation distribution image of the subject based on the tomographic image of the subject; and an attenuation correction unit for performing attenuation correction on the tomographic image of the subject using the attenuation distribution image of the subject.

[0018] An embodiment of the present invention is a method for creating an attenuation distribution image of a subject based on a tomographic image of the subject, the method including: (1) an image conversion step of generating a converted image in which pixel values ​​of a partial region are changed based on the tomographic image of each slice position of the subject; and (2) a CNN processing step of inputting the converted image of each slice position generated in the image conversion step to a convolutional neural network, inputting slice position information of the converted image to an embedding layer, adding output data from the embedding layer to a feature map of the convolutional neural network, and outputting the output image from the convolutional neural network as the attenuation distribution image of the subject.

[0019] An embodiment of the present invention is an image processing method, comprising: the attenuation distribution image creation method having the above-described configuration for creating an attenuation distribution image of a subject based on a tomographic image of the subject; and an attenuation correction step for performing attenuation correction on the tomographic image of the subject using the attenuation distribution image of the subject.

[0020] According to an embodiment of the present invention, even if the drug administered to the subject when the tomographic image to be attenuation corrected was acquired is different from the drug administered to the subject when the tomographic image for CNN learning was acquired, a more accurate attenuation distribution image can be created from the tomographic image to be attenuation corrected.

[0021] FIG. 1 is a diagram showing the configuration of a radiation tomography system 1. FIG. 2 is a diagram showing the configuration of an attenuation distribution image creation device. FIG. 3 is a diagram explaining the configuration of a CNN processing unit 23. FIG. １８ 5 shows images of subjects administered [F]FDG. １１ C] Images of subjects administered with raclopride. １１ Figure 7 shows images of subjects administered [C]PIB. １１ C] Images of subjects administered PBB3. １１ C] A figure showing an image of a subject administered with methionine. １１ [C] Figures showing images of a subject administered DPA713. Figure 10 is a graph showing, for each drug, the PSNR of the attenuation distribution image of all slices obtained based on the tomographic image to be attenuation corrected. Figure 11 is a graph showing, for each drug, the PSNR of the attenuation distribution image of slice 5 obtained based on the tomographic image to be attenuation corrected. Figure 12 is a graph showing, for each drug, the PSNR of the attenuation distribution image of slice 20 obtained based on the tomographic image to be attenuation corrected.

[0022] Hereinafter, with reference to the accompanying drawings, detailed descriptions will be given of embodiments of an attenuation distribution image creation device, an image processing device, a radiation tomography system, an attenuation distribution image creation method, and an image processing method. In the description of the drawings, the same elements are given the same reference numerals, and duplicate descriptions 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.

[0023] 1 is a diagram showing the configuration of a radiation tomography system 1. The radiation tomography system 1 includes a radiation tomography apparatus 10 and an image processing device 20. The image processing device 20 includes an image reconstruction unit 21, an image conversion unit 22, a CNN processing unit 23, and an attenuation correction unit 24. Of these, the image conversion unit 22 and the CNN processing unit 23 constitute an attenuation distribution image creation device that creates an attenuation distribution image based on a tomographic image to be attenuation corrected.

[0024] The radiation tomography apparatus 10 is an apparatus that performs an emission scan to collect list data (emission scan data) for reconstructing a tomographic image (emission scan image) of a subject. Examples of the radiation tomography apparatus 10 include a PET apparatus and a SPECT apparatus. In the following description, the radiation tomography apparatus 10 will be described as a PET apparatus.

[0025] The radiation tomography apparatus 10 includes a detection unit having a large number of small radiation detectors arranged around a measurement space in which a subject placed on a bed is placed. The radiation tomography apparatus 10 detects photon pairs with an energy of 511 keV generated in association with electron-positron annihilation within a subject into which a positron-emitting isotope (RI source) has been introduced, using a coincidence method with the detection unit, and accumulates this coincidence information. The radiation tomography apparatus 10 then outputs this accumulated coincidence information, arranged in chronological order, to the image processing apparatus 20 as list data (emission scan data).

[0026] The list data includes identification information and detection time information of a pair of radiation detectors that have counted the photon pairs coincidentally. The list data may further include energy information of the photons detected by each radiation detector and information about the detection time difference between the pair of radiation detectors.

[0027] The image processing device 20 is a device that reconstructs a tomographic image of the subject after attenuation correction based on the emission scan data of the subject output from the radiation tomography device 10 .

[0028] The image processing device 20 reconstructs an emission scan image (a tomographic image to be attenuation corrected) using an image reconstruction unit 21 based on the emission scan data, creates an attenuation distribution image based on this tomographic image to be attenuation corrected using an attenuation distribution image creation device (image conversion unit 22, CNN processing unit 23), and performs attenuation correction of the tomographic image based on this attenuation distribution image using an attenuation correction unit 24.

[0029] A computer having a CPU, GPU, RAM, ROM, a hard disk drive, etc. is used as the image processing device 20. The image processing device 20 also includes an input unit (e.g., a keyboard and a mouse) that accepts input from an operator, and a display unit (e.g., a liquid crystal display) that displays images, etc.

[0030] The image reconstruction unit 21 reconstructs a tomographic image based on the emission scan data acquired by the radiation tomography apparatus 10 .

[0031] The maximum likelihood expectation maximization (ML-EM) method and an improved iterative image reconstruction technique using a block iteration method are known as techniques for reconstructing tomographic images based on emission scan data in the image reconstruction unit 21. Other known iterative image reconstruction techniques using a block iteration method include the ordered subset ML-EM (OSEM) method, the row-action maximum likelihood algorithm (RAMLA) method, and the dynamic RAMLA (DRAMA) method.

[0032] The image conversion unit 22 generates a converted image in which pixel values ​​of a partial region are changed based on the tomographic image at each slice position of the subject. The converted image is a converted image in which pixel values ​​of a partial region are changed based on the tomographic image to be attenuation corrected, and preferably a converted image in which pixel values ​​of a partial region of the subject region in the tomographic image to be attenuation corrected are changed.

[0033] The pixel values ​​of the converted image for the tomographic image can be changed in any manner. The partial regions whose pixel values ​​are changed may have any shape, any size, and any number. The change in 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.

[0034] The CNN processing unit 23 receives the converted image for each slice position generated by the image conversion unit 22 and slice position information of the converted image. The CNN processing unit 23 inputs the converted image for each slice position to a trained convolutional neural network (CNN) 31, and also inputs the slice position information of the converted image to an embedding layer 32. Output data from the embedding layer 32 is added to the feature map of the CNN 31. The CNN processing unit 23 then outputs the output image from the CNN 31 as an attenuation distribution image of the subject.

[0035] The attenuation correction unit 24 receives the emission scan data and also receives the attenuation distribution image created by the CNN processing unit 23. Then, the attenuation correction unit 24 performs image reconstruction with attenuation correction based on the emission scan data and the attenuation distribution image, and creates a tomographic image after the correction.

[0036] The attenuation distribution image creation method of this embodiment is a method for creating an attenuation distribution image of a subject based on a tomographic image of the subject. The image processing method of this embodiment includes the attenuation distribution image creation method of this embodiment and is a method for performing attenuation correction of a tomographic image using the created attenuation distribution image.

[0037] In the image processing method of this embodiment, an emission scan image (a tomographic image to be attenuation corrected) is reconstructed based on emission scan data in an image reconstruction step, an attenuation distribution image is created based on the tomographic image to be attenuation corrected in an attenuation distribution image creation step (image conversion step, CNN processing step), and attenuation correction of the tomographic image is performed based on the attenuation distribution image in an attenuation correction step.

[0038] In the image reconstruction step, the image reconstruction unit 21 reconstructs a tomographic image based on the emission scan data acquired by the radiation tomography apparatus 10. In the image conversion step, the image conversion unit 22 generates a converted image in which pixel values ​​of a partial region are changed based on the tomographic image at each slice position of the subject.

[0039] In the CNN processing step, the CNN processing unit 23 inputs the converted image at each slice position to the trained CNN 31, and also inputs slice position information of the converted image to the embedding layer 32. The output data from the embedding layer 32 is added to the feature map of the CNN 31. Then, in the CNN processing step, the output image from the CNN 31 is output as an attenuation distribution image of the subject.

[0040] In the attenuation correction step, the attenuation correction unit 24 performs image reconstruction with attenuation correction based on the emission scan data and the attenuation distribution image, and creates a tomographic image after the correction.

[0041] 2 is a diagram showing the configuration of the attenuation distribution image creation device. The attenuation distribution image creation device includes an image conversion unit 22, a CNN processing unit 23, and a CNN learning unit 25. This diagram shows the configuration of a system for training a CNN 31 and an embedding layer 32 used in the CNN processing unit 23.

[0042] When training the CNN 31 and the embedding layer 32, first, a plurality of sets of training tomographic images and attenuation distribution images as teacher images are prepared. The training tomographic images prepared here may be tomographic images of a subject to which a given drug has been administered. Furthermore, the attenuation distribution images prepared here as teacher images may be transmission scan images obtained by performing a transmission scan with a PET device, or may be X-ray CT scan images obtained by performing a scan with an X-ray CT device.

[0043] A converted image at each slice position generated by the image conversion unit 22 based on the learning tomographic image is input to the CNN 31 , and slice position information of the converted image is input to the embedding layer 32 .

[0044] The CNN learning unit 25 evaluates the error (e.g., L1 norm, L2 norm, negative logarithmic likelihood in Poisson distribution, etc.) between the output image from the CNN 31 and the attenuation distribution image as a teacher image, and corrects the parameters of the CNN 31 and the embedding layer 32 based on the error evaluation result. By performing this series of processes using multiple sets of learning tomographic images and attenuation distribution images as teacher images, the CNN 31 and the embedding layer 32 are trained (CNN learning step).

[0045] 3 is a diagram illustrating the configuration of the CNN processing unit 23. The CNN 31 has a U-net structure including an encoder unit and a decoder unit.

[0046] The encoder unit of the CNN 31 includes a convolutional layer, a normalization layer, a nonlinear activation function, and downsampling. The encoder unit receives the converted image output from the image conversion unit 22 and updates the feature map using the convolutional layer, normalization layer, and nonlinear activation function. After performing processing using the convolutional layer, normalization layer, and nonlinear activation function several times, the encoder unit downsamples to reduce the spatial size of the feature map and increase the number of channels in the feature map. For example, the spatial size of the feature map is initially 128 x 128, but is subsequently reduced to 64 x 64, 32 x 32, and 16 x 16.

[0047] The decoder unit of the CNN 31 is provided after the encoder unit. The decoder unit includes a convolutional layer, a normalization layer, a nonlinear activation function, and upsampling. The decoder unit increases the spatial size of the feature map and reduces the number of channels in the feature map through upsampling. For example, the spatial size of the feature map is increased to 16x16, 32x32, 64x64, or 128x128. After performing the upsampling process, the decoder unit updates the feature map using the convolutional layer, normalization layer, and nonlinear activation function.

[0048] The embedding layer 32 receives slice position information of the converted image input to the CNN 31, converts the slice position information into a data string, and outputs the data string. The output data from the embedding layer 32 is added to the feature map of the CNN 31.

[0049] In this figure, a data sequence [0.2, 0.5, ..., 0.3, ...] is output from the embedding layer 32, and the first data, 0.2, is added to the feature map of the first channel after the first downsampling stage of the encoder unit of the CNN 31, and the second data, 0.5, is added to the feature map of the second channel after the first downsampling stage of the encoder unit of the CNN 31.

[0050] The output data from the embedding layer 32 is added to the feature map of each channel after downsampling or upsampling of the CNN 31. The output data from the embedding layer 32 may be added to the feature maps of each channel after all downsampling and all upsampling of the CNN 31.

[0051] Next, an example will be described. In this example, 188 sets of tomographic images and attenuation distribution images without attenuation correction obtained for the head of a subject using a PET device dedicated to the head were used.

[0052] As a drug １８ F]FDG, [ １１ C] Raclopride, １１ C]PIB, [ １１ C] PBB3, [ １１ C] Methionine and [ １１ C]DPA713 was administered to the subjects. １８ For learning, uncorrected tomographic images and attenuation distribution images obtained from subjects administered [F]FDG were used. The attenuation distribution images created by CNN in this example and the comparative example were compared using the technology described in Non-Patent Document 1 as a comparative example.

[0053] 4 to 9 are diagrams showing images obtained in the example and comparative example, respectively. １８Figure 5 shows images of subjects administered [F]FDG. １１ C] Images of subjects administered with raclopride. １１ Figure 7 shows images of subjects administered [C]PIB. １１ C] Images of subjects administered PBB3. １１ C] Images of a subject administered with methionine are shown in FIG. １１ C] Images of subjects administered DPA713 are shown.

[0054] In each of Figures 4 to 9, the top row shows a tomographic image without attenuation correction, the second row shows an attenuation distribution image (correct image), the third row shows an attenuation distribution image created by CNN in the comparative example, and the bottom row shows an attenuation distribution image created by CNN in the example.

[0055] 4 to 9, the left column shows images of slice 5, and the right column shows images of slice 20. The larger the slice number, the higher the position of the subject's head (the closer to the top of the head). In a head-specific PET device, emission scans and transmission scans are performed using a jig that positions the subject's head, so the same slice number indicates approximately the same position of the head even if the subject is different.

[0056] 10 to 12 are graphs showing the PSNR of attenuation distribution images obtained for each drug based on tomographic images to be attenuation corrected in the examples and comparative examples. Fig. 10 shows the PSNR of the attenuation distribution images for all slices, Fig. 11 shows the PSNR of the attenuation distribution image for slice 5, and Fig. 12 shows the PSNR of the attenuation distribution image for slice 20. The PSNR (Peak Signal to Noise Ratio) represents image quality in decibels (dB), with a higher value indicating better image quality.

[0057] As shown in FIGS. 4 to 12, when the tomographic image to be attenuation corrected was acquired, the drug administered to the subject [ １８ In the case of [F]FDG, accurate attenuation distribution images could be created even in the comparative example. １１ C] DPA 713 is [ １８F] shows an accumulation distribution similar to that of FDG, so when the tomographic image to be attenuation corrected is acquired, the drug administered to the subject is [ １１ In both the case of C]DPA713 and the comparative example, accurate attenuation distribution images could be created.

[0058] When the tomographic image to be attenuation corrected was acquired, the drug administered to the subject was [ １８ When the attenuation distribution image was different from [F]FDG, in this example, a more accurate attenuation distribution image could be created from the tomographic image to be attenuation corrected compared to the comparative example.

[0059] As described above, in this embodiment, even if a transmission scanning mechanism or an X-ray CT scanning mechanism is not provided, an attenuation distribution image of a subject can be created based on a tomographic image of the subject that is to be attenuation corrected, and an attenuation-corrected tomographic image can be created, thereby simplifying the device and reducing its cost. Furthermore, in this embodiment, a more accurate attenuation distribution image can be created, and by using this, attenuation correction of a tomographic image can be performed with higher precision.

[0060] The attenuation distribution image creating device, image processing device, radiation tomography system, attenuation distribution image creating method, and image processing method are not limited to the above-described embodiments and configuration examples, and various modifications are possible.

[0061] The attenuation distribution image creation device of the first aspect according to the above embodiment is a device that creates an attenuation distribution image of a subject based on a tomographic image of the subject, and includes: (1) an image conversion unit that generates a converted image in which pixel values ​​of a partial region are changed based on the tomographic image of each slice position of the subject; and (2) a CNN processing unit that inputs the converted image of each slice position generated by the image conversion unit to a convolutional neural network, inputs slice position information of the converted image to an embedding layer, adds output data from the embedding layer to a feature map of the convolutional neural network, and outputs the output image from the convolutional neural network as an attenuation distribution image of the subject.

[0062] The attenuation distribution image creation device of the second aspect may be configured in the configuration of the first aspect by using a plurality of sets of learning tomographic images and attenuation distribution images as teacher images, inputting converted images for each slice position generated by an image conversion unit based on the learning tomographic images to a convolutional neural network, and by inputting slice position information of the converted images to an embedding layer, evaluating an error between an output image from the convolutional neural network and the attenuation distribution image as a teacher image, and training the convolutional neural network and the embedding layer based on the error evaluation result.

[0063] The image processing device according to the above embodiment includes an attenuation distribution image creation device having the above configuration that creates an attenuation distribution image of the subject based on a tomographic image of the subject, and an attenuation correction unit that performs attenuation correction on the tomographic image of the subject using the attenuation distribution image of the subject.

[0064] The radiation tomography system according to the above embodiment includes a radiation tomography device for acquiring a tomographic image of a subject, an attenuation distribution image creation device having the above configuration for creating an attenuation distribution image of the subject based on the tomographic image of the subject, and an attenuation correction unit for performing attenuation correction on the tomographic image of the subject using the attenuation distribution image of the subject.

[0065] The attenuation distribution image creation method of the first aspect according to the above embodiment is a method for creating an attenuation distribution image of a subject based on a tomographic image of the subject, and includes: (1) an image conversion step of generating a converted image in which pixel values ​​of a partial region are changed based on the tomographic image of each slice position of the subject; and (2) a CNN processing step of inputting the converted image of each slice position generated in the image conversion step to a convolutional neural network, inputting slice position information of the converted image to an embedding layer, adding output data from the embedding layer to a feature map of the convolutional neural network, and outputting the output image from the convolutional neural network as an attenuation distribution image of the subject.

[0066] The attenuation distribution image creation method of the second aspect may be configured in the configuration of the first aspect, further comprising a CNN learning step in which a plurality of sets of learning tomographic images and attenuation distribution images as teacher images are used, and converted images for each slice position generated in the image conversion step based on the learning tomographic images are input to a convolutional neural network, and when slice position information of the converted images is input to an embedding layer, an error between an output image from the convolutional neural network and the attenuation distribution image as the teacher image is evaluated, and the convolutional neural network and the embedding layer are trained based on the error evaluation result.

[0067] The image processing method according to the above embodiment includes a step of performing the attenuation distribution image creation method of the above configuration, which creates an attenuation distribution image of the subject based on a tomographic image of the subject, and an attenuation correction step of performing attenuation correction on the tomographic image of the subject using the attenuation distribution image of the subject.

[0068] The present invention can be used as an apparatus and method that can create a more accurate attenuation distribution image from a tomographic image to be attenuation corrected, even if the drug administered to the subject when the tomographic image to be attenuation corrected was acquired is different from the drug administered to the subject when the tomographic image for CNN training was acquired.

[0069] 1...Radiation tomography system, 10...Radiation tomography apparatus, 20...Image processing device, 21...Image reconstruction unit, 22...Image conversion unit, 23...CNN processing unit, 24...Attenuation correction unit, 25...CNN learning unit, 31...Convolutional neural network (CNN), 32...Embedding layer

Claims

1. An apparatus for creating an attenuation distribution image of a subject based on a tomographic image of the subject, comprising: an image conversion unit that generates a converted image by changing pixel values ​​of a partial area based on the tomographic image of each slice position of the subject; and a CNN processing unit that inputs the converted image of each slice position generated by the image conversion unit to a convolutional neural network, inputs slice position information of the converted image to an embedding layer, adds output data from the embedding layer to a feature map of the convolutional neural network, and outputs an output image from the convolutional neural network as the attenuation distribution image of the subject.

2. The attenuation distribution image creation device of claim 1, further comprising a CNN learning unit that uses a plurality of sets of learning tomographic images and attenuation distribution images as teacher images, inputs converted images for each slice position generated by the image conversion unit based on the learning tomographic images to the convolutional neural network, and inputs slice position information of the converted images to the embedding layer, evaluates an error between an output image from the convolutional neural network and the attenuation distribution image as the teacher image, and trains the convolutional neural network and the embedding layer based on the error evaluation result.

3. An image processing device comprising: an attenuation distribution image creation device according to claim 1 or 2, which creates an attenuation distribution image of a subject based on a tomographic image of the subject; and an attenuation correction unit which performs attenuation correction of the tomographic image of the subject using the attenuation distribution image of the subject.

4. A radiation tomography system comprising: a radiation tomography device for acquiring a tomographic image of a subject; an attenuation distribution image creation device according to claim 1 or 2 for creating an attenuation distribution image of the subject based on the tomographic image of the subject; and an attenuation correction unit for performing attenuation correction of the tomographic image of the subject using the attenuation distribution image of the subject.

5. A method for creating an attenuation distribution image of a subject based on a tomographic image of the subject, comprising: an image conversion step of generating a converted image in which pixel values ​​of a partial area are changed based on the tomographic image of each slice position of the subject; and a CNN processing step of inputting the converted image of each slice position generated in the image conversion step to a convolutional neural network, inputting slice position information of the converted image to an embedding layer, adding output data from the embedding layer to a feature map of the convolutional neural network, and outputting an output image from the convolutional neural network as an attenuation distribution image of the subject.

6. The attenuation distribution image creating method according to claim 5, further comprising a CNN learning step of using a plurality of sets of learning tomographic images and attenuation distribution images as teacher images, inputting converted images for each slice position generated in the image conversion step based on the learning tomographic images to the convolutional neural network, and inputting slice position information of the converted images to the embedding layer, evaluating an error between an output image from the convolutional neural network and the attenuation distribution image as the teacher image, and training the convolutional neural network and the embedding layer based on the error evaluation result.

7. An image processing method comprising: an attenuation distribution image creation method according to claim 5 or 6, which creates an attenuation distribution image of a subject based on a tomographic image of the subject; and an attenuation correction step of performing attenuation correction of the tomographic image of the subject using the attenuation distribution image of the subject.

Images