X-ray imaging method and X-ray imaging device

Abstract

To provide an X-ray imaging method capable of reducing noise sufficiently and acquiring a re-configured image in which image quality is improved, in a photon counting type X-ray imaging for detecting an X-ray photon.SOLUTION: There is provided an X-ray imaging method comprising: a step for radiating an X-ray to a subject 200 and acquiring a detection result of an X-ray photon generated by an X-ray passed through the subject 200; a step for, on the basis of the detection result of the X-ray photon, generating a plurality of re-configured images 11 in energy ranges which are different from each other; and a step for using a deep layer convolution neural network 7 to perform deep layer image processing in which the processing is stopped before convergence of the deep layer convolution neural network 7 with the plurality of re-configured images 11 as pieces of inputs of the deep layer convolution neural network 7, thereby generating a plurality of noise reduced re-configured images 12.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to an X-ray imaging method and an X-ray imaging apparatus, and more particularly to an X-ray imaging method and an X-ray imaging apparatus that detect X-ray photons.

[0002] BACKGROUND ART Conventionally, an X-ray imaging apparatus equipped with an energy discrimination detector that detects X-ray photons is known (see, for example, Patent Document 1). [Background technology]

[0003] The photon-counting CT device disclosed in Patent Document 1 above includes an X-ray detection unit (energy discrimination detector) that detects X-ray photons generated by an X-ray generation unit and transmitted through a subject. This photon-counting CT device is configured to collect count data representing the number of detected X-ray photons over multiple energy ranges. This photon-counting CT device is also configured to reconstruct photon-counting CT images (reconstructed images) for the multiple energy ranges. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2015-131028 Summary of the Invention [Problem to be solved by the invention]

[0005] Although not explicitly stated in Patent Document 1, photon-counting CT devices such as those described in Patent Document 1 typically have a low upper limit for the X-ray dose that can be detected by an energy discrimination detector, making it difficult to obtain a sufficient X-ray detection amount and prone to producing noisy photon-counting CT images. Furthermore, because the X-ray dose is distributed across multiple energy ranges, the X-ray detection amount in the photon-counting CT image for each energy range is low, making it prone to producing a lot of noise. This poses the problem of difficulty in obtaining a photon-counting CT image (reconstructed image) with sufficiently reduced noise and improved image quality.

[0006] The present invention has been made to solve the above-mentioned problems, and one object of the present invention is to provide an X-ray imaging method and an X-ray imaging apparatus that can obtain reconstructed images with improved image quality by sufficiently reducing noise in photon-counting X-ray imaging that detects X-ray photons. [Means for solving the problem]

[0007] In order to achieve the above object, an X-ray imaging method according to a first aspect of the present invention includes the steps of: irradiating an object with X-rays and obtaining a detection result of X-ray photons due to the X-rays that have passed through the object; generating a plurality of reconstructed images in different energy ranges based on the detection result of the X-ray photons; and generating a plurality of reconstructed images in different energy ranges using a deep convolutional neural network. Processing results Deep image processing that stops processing before convergence is performed by using multiple reconstructed images. Each of By using this as the input of a deep convolutional neural network, This is an image obtained by reducing noise from multiple input reconstructed images.and generating a plurality of noise-reduced reconstructed images. Note that the deep image processing is, for example, DIP (Deep Image Prior) processing, which is a type of image quality improvement processing using a deep convolutional neural network. In such deep image processing, noise information cannot be restored unless the number of convergences is large, while structural information other than noise information can be restored with a small number of convergences. By utilizing this fact, processing can be stopped before the deep convolutional neural network converges, thereby generating an image with less noise. Furthermore, in order to achieve the above object, an X-ray imaging method according to a second aspect of the present invention includes the steps of: irradiating an object with X-rays and obtaining a detection result of X-ray photons from the X-rays that have passed through the object; generating a plurality of reconstructed images having different energy ranges based on the detection result of the X-ray photons; and generating a plurality of noise-reduced reconstructed images by using a deep convolutional neural network and performing deep image processing, in which the processing is stopped before the processing result of the deep convolutional neural network converges, with the plurality of reconstructed images as input to the deep convolutional neural network. The step of generating a plurality of noise-reduced reconstructed images includes a step of generating a plurality of noise-reduced reconstructed images by performing deep image processing in which a plurality of reconstructed images representing the same cross-section of the same object are input to the deep convolutional neural network, and the step of generating a plurality of noise-reduced reconstructed images includes a step of using the plurality of reconstructed images as input and performing deep image processing while sharing structural information possessed by each of the plurality of reconstructed images in the deep convolutional neural network, thereby generating a plurality of noise-reduced reconstructed images. In addition, in order to achieve the above object, an X-ray imaging method in a third aspect of the present invention includes the steps of irradiating an object with X-rays and obtaining a detection result of X-ray photons from the X-rays that have passed through the object; generating a plurality of reconstructed images with different energy ranges based on the detection result of the X-ray photons; and generating a plurality of noise-reduced reconstructed images by performing deep image processing using a deep convolutional neural network, in which the processing is stopped before the processing result of the deep convolutional neural network converges, using the plurality of reconstructed images as input to the deep convolutional neural network, and further includes the step of performing deep image processing on at least one of a virtual monochromatic image, an effective atomic number image, and a density distribution image obtained based on the plurality of reconstructed images.

[0008] In order to achieve the above object, the first aspect of the present invention is 4 The X-ray imaging device in this aspect includes an X-ray irradiation unit that irradiates an object with X-rays, an energy discrimination detector that detects X-ray photons due to X-rays that have passed through the object, and an image processing unit that generates a plurality of reconstructed images in different energy ranges based on the detection result of the X-ray photons by the energy discrimination detector, and the image processing unit uses a deep convolutional neural network to generate a plurality of reconstructed images in different energy ranges. Processing results Deep image processing that stops processing before convergence is performed by using multiple reconstructed images. Each of By using this as the input of a deep convolutional neural network, This is an image obtained by reducing noise from multiple input reconstructed images. A plurality of noise-reduced reconstructed images are generated. [Effects of the Invention]

[0009] Above No. 1 ~3rd The X-ray imaging method and the above-mentioned 4 In the X-ray imaging device, a deep convolutional neural network is used to Processing results Deep image processing that stops processing before convergence is performed by using multiple reconstructed images. Each of By using this as the input of a deep convolutional neural network, This is an image obtained by reducing noise from multiple input reconstructed images.Multiple noise-reduced reconstructed images are generated. This allows for easy generation of noise-reduced reconstructed images from reconstructed images using deep image processing. Furthermore, by performing deep image processing on multiple reconstructed images, unlike when deep image processing is simply performed on a single reconstructed image, deep image processing can be performed on multiple reconstructed images while interacting with each other, thereby achieving a higher noise reduction effect than when deep image processing is performed on a single reconstructed image. As a result, in photon-counting X-ray imaging that detects X-ray photons, noise is sufficiently reduced, and a noise-reduced reconstructed image with improved image quality can be obtained.

[0010] Furthermore, since a noise-reduced reconstructed image with improved image quality can be obtained with sufficiently reduced noise, when material decomposition is performed based on the noise-reduced reconstructed image, the material decomposition can be performed with high accuracy. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a schematic diagram showing the configuration of an X-ray imaging apparatus according to an embodiment. [Figure 2] 3A to 3C are diagrams for explaining the detection results of X-ray photons and a reconstructed image of the X-ray imaging apparatus according to one embodiment. [Figure 3] FIG. 2 is a diagram for explaining deep image processing of an X-ray imaging apparatus according to an embodiment. [Figure 4] FIG. 1 is a diagram for explaining a deep convolutional neural network of an X-ray imaging apparatus according to an embodiment. [Figure 5] 10 is a flowchart for explaining a control process relating to X-ray imaging by the X-ray imaging apparatus according to an embodiment. [Figure 6] 1A and 1B are diagrams showing an actual reconstructed image before deep image processing and an actual reconstructed image after deep image processing. [Figure 7] FIG. 10 shows correct data from an experiment to distinguish resin types with and without deep image processing. [Figure 8]This figure shows the accuracy rate without deep image processing and the accuracy rate with deep image processing in an experiment to distinguish resin types with and without deep image processing. DETAILED DESCRIPTION OF THE INVENTION

[0012] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will now be described with reference to the accompanying drawings.

[0013] (Configuration of X-ray equipment) The configuration of an X-ray imaging apparatus 100 according to one embodiment of the present invention will be described with reference to FIGS.

[0014] 1, the X-ray imaging device 100 is a photon-counting type X-ray imaging device that generates an image of the inside of the subject 200 by using X-ray photons that have passed through the subject 200. Specifically, the X-ray imaging device 100 is a PCCT (Photon Counting Computed Tomography) device. The X-ray imaging device 100 is configured to be capable of performing material decomposition by utilizing the fact that the energy value characteristics of the X-ray photons that have passed through the subject 200 differ depending on the material.

[0015] The X-ray imaging device 100 includes an X-ray irradiation unit 1, an energy discrimination detector (X-ray detection unit) 2, a counting unit 3, a rotating stage 4, a control unit 5, and a storage unit 6. The control unit 5 is an example of the "image processing unit" in the claims.

[0016] The X-ray irradiator 1 is configured to irradiate the subject 200 with X-rays. The X-ray irradiator 1 includes an X-ray tube, and is configured to generate X-rays when a high voltage is applied thereto, and to irradiate the generated X-rays toward the energy discrimination detector 2.

[0017] The energy discrimination detector 2 is a photon count type X-ray detection unit. The energy discrimination detector 2 is configured to detect X-ray photons generated by X-rays irradiated from the X-ray irradiation unit 1 and transmitted through the subject 200. The energy discrimination detector 2 is also configured to convert the detected X-ray photons into an electrical signal and output the converted electrical signal. The electrical signal (electrical pulse) has a peak value corresponding to the energy of the X-ray photon.

[0018] The energy discrimination detector 2 also includes a plurality of detection elements that convert X-ray photons into electrical signals. The detection elements can be, for example, direct conversion semiconductor elements that directly convert X-ray photons into electrical signals. Examples of such semiconductor elements include cadmium telluride semiconductor elements. Alternatively, the detection elements can be, for example, indirect conversion detection elements that include a scintillator that converts X-ray photons into light and a photodiode that converts light into an electrical signal, and that indirectly convert X-ray photons into an electrical signal.

[0019] The counting unit 3 is configured to count the X-ray photons detected by the energy discrimination detector 2 for each energy range. The energy ranges are not particularly limited, but for example, as shown in FIG. 2, the counting unit 3 counts X-ray photons in multiple (four) energy ranges E1 to E4. At this time, the counting result of X-ray photons according to the X-ray spectrum of the original X-rays is obtained. The counting unit 3 is also configured to discriminate the electric signals (electric pulses) from the energy discrimination detector 2 according to their peak values, and count the X-ray photons in each of the multiple energy ranges, using the number of electric signals as the number of X-ray photons.

[0020] The rotation stage 4 is configured to rotate the subject 200. The rotation stage 4 includes a mounting table for placing the subject 200 thereon, and a drive unit such as a motor that rotates the mounting table. The rotation stage 4 is configured to rotate under the control of a control unit 5. By rotating the subject 200 using the rotation stage 4, the subject 200 can be rotated relative to an imaging system including the X-ray irradiation unit 1 and the energy discrimination detector 2. This makes it possible to perform X-ray CT imaging of the subject 200 while changing the imaging position (view) of the imaging system relative to the subject 200.

[0021] The control unit 5 is configured to control each unit of the X-ray imaging apparatus 100. The control unit 5 includes a processor such as a CPU (Central Processing Unit) and a memory. The control unit 5 also functions as an image processing unit that performs image processing. Specifically, as shown in FIG. 2, the control unit 5 is configured to acquire multiple reconstructed images 11 in different energy ranges based on the detection results of X-ray photons by the energy discrimination detector 2. More specifically, the control unit 5 is configured to acquire multiple reconstructed images 11 by performing reconstruction processing such as FBP (Filtered Back Projection) based on the counting results of X-ray photons for each energy range for each imaging position (view) of X-ray CT imaging, which are detected by the energy discrimination detector 2 and counted by the counter 3. Note that in the figure, black dots in the reconstructed images 11 represent noise. Reconstructed images 11 in an energy range with a small number of X-ray photons (e.g., E4) have relatively more noise, while reconstructed images 11 in an energy range with a large number of X-ray photons (e.g., E2) have relatively less noise.

[0022] The control unit 5 is also configured to perform material decomposition processing to distinguish materials in the reconstructed images 11 by acquiring linear attenuation coefficients based on the reconstructed images 11 in different energy ranges. The control unit 5 is also configured to generate, based on the reconstructed images 11 in different energy ranges, a virtual monochromatic image obtained by weighting and combining the reconstructed images 11, an effective atomic number image representing the effective atomic number for each pixel obtained from the reconstructed images 11, and a density distribution image representing the material density for each pixel obtained from the reconstructed images 11. The memory unit 6 is configured, for example, with a storage device such as a hard disk drive. The memory unit 6 stores a deep convolutional neural network 7 (hereinafter referred to as "DCNN") used for image processing. Details of the DCNN 7 will be described later.

[0023] In this embodiment, as shown in FIGS. 3 and 4 , the control unit 5 is configured to perform deep image processing using a DCNN 7 that does not undergo pre-training using training data, stopping the processing before the DCNN 7 converges, using multiple reconstructed images 11 as input to the DCNN 7, thereby generating multiple noise-reduced reconstructed images 12 from the multiple reconstructed images 11. In this embodiment, the deep image processing is deep image prior (DIP) processing, a type of image quality improvement processing using a deep convolutional neural network. DIP processing is an image quality improvement processing proposed by Dmitry Ulyanov et al. in a paper titled "Deep Image Prior" in 2017. In DIP processing, in a deep convolutional neural network that does not undergo pre-training using training data, noise information cannot be restored unless many convergences are performed, while structural information other than noise information can be restored with fewer convergences. By stopping the processing before the deep convolutional neural network converges, it is possible to generate images with less noise.

[0024] The DCNN7 is configured with a U-Net (see FIG. 4), a type of Fully Convolution Network (FCN). The DCNN7 also includes an input layer, an intermediate layer, and an output layer. The input layer is configured to receive input images (multiple reconstructed images 11). The input layer has a number of channels (e.g., four) corresponding to the number of multiple reconstructed images 11. The intermediate layer has a downsampling unit that acquires feature components from the input images while reducing the size of the input images received by the input layer, and an upsampling unit that restores the image, including the feature components, reduced in size by the downsampling unit to its original size (the size of the input image). The output layer is configured to generate and output an output image. The output layer has a number of channels (e.g., four) corresponding to the number of multiple reconstructed images 11 input to the input layer. That is, the output layer is configured to generate and output multiple output images corresponding to the multiple reconstructed images 11 input to the input layer. Note that the final multiple output images obtained when the processing of the DCNN7 is stopped before convergence become multiple noise-reduced reconstructed images 12.

[0025] Moreover, in this embodiment, a plurality of reconstructed images 11 having different degrees of noise are input to the DCNN 7. For this reason, the control unit 5 is configured to perform deep image processing (hereinafter referred to as "DIP processing") in which the plurality of reconstructed images 11 having different degrees of noise are input to the DCNN 7, thereby generating a plurality of noise-reduced reconstructed images 12. Furthermore, in this embodiment, a plurality of reconstructed images 11 representing the same cross-sectional plane of the same subject 200 are input to the DCNN 7. For this reason, the control unit 5 is configured to perform DIP processing in which the plurality of reconstructed images 11 representing the same cross-sectional plane of the same subject 200 are input to the DCNN 7, thereby generating a plurality of noise-reduced reconstructed images 12.

[0026] Furthermore, the DCNN 7 receives as input a plurality of reconstructed images 11 that are the same image size and that are successive in the energy range from low to high, such as E1 to E4 shown in Fig. 2. That is, unlike when only a single reconstructed image 11 is input, the DCNN 7 is configured to perform processing using not only spatial information corresponding to the vertical and horizontal directions of the image, but also energy information corresponding to the energy direction of the X-rays (i.e., X-ray spectrum information).

[0027] Furthermore, in this embodiment, the control unit 5 is configured to receive a plurality of reconstructed images 11 as input, and perform DIP processing while sharing structural information possessed by each of the plurality of reconstructed images 11 in the DCNN 7, thereby generating a plurality of noise-reduced reconstructed images 12. Specifically, the control unit 5 is configured to receive a plurality of reconstructed images 11 as input, and perform DIP processing to repeatedly update parameters of the DCNN 7 while sharing structural information possessed by each of the plurality of reconstructed images 11 in the DCNN 7, thereby generating a plurality of noise-reduced reconstructed images 12.

[0028] Specifically, the control unit 5 is configured to perform deep image processing that repeatedly updates the parameters of the DCNN 7 based on the following formula (1). Formula (1) defines the parameters of the DCNN 7 that minimize the error function between the input image (plurality of reconstructed images 11) and the output image. θ is a parameter of the DCNN 7, specifically, a weight parameter of each layer (such as a convolutional layer) of the DCNN 7. A random initial value is assigned to θ. x0 represents the input image (plurality of reconstructed images 11). f θ (z) is the code vector z to image x * This represents a network (DCNN7) that maps to f θ (z) and x * represents the output image.

number

[0029] The control unit 5 is configured to repeatedly update the parameters of the DCNN 7 using an optimization method such as gradient descent so that the parameters of the DCNN 7 converge to parameters that minimize the error function of Equation (1). That is, the control unit 5 is configured to repeatedly generate an output image while repeatedly updating the parameters of the DCNN 7. At this time, the parameters of the DCNN 7 are repeatedly updated based on multiple reconstructed images 11 (x0 in Equation (1)). Therefore, the parameters of the DCNN 7 are repeatedly updated while reflecting structural information (structural information of the cross-section of the subject 200) possessed by each of the multiple reconstructed images 11 in the parameters of the DCNN 7 (while sharing the structural information in the DCNN 7). The control unit 5 is configured to stop updating the parameters of the DCNN 7 at a predetermined timing before the parameters of the DCNN 7 converge to parameters that minimize the error function of Equation (1), thereby acquiring an output image suitable as the noise-reduced reconstructed image 12. Note that if processing is continued until the DCNN 7 converges, an image identical to the input image will be generated as the output image.

[0030] Furthermore, the timing to stop updating the parameters of the DCNN 7 (i.e., the timing to stop the processing of the DCNN 7) is not particularly limited as long as it is possible to obtain the noise-reduced reconstructed image 12, but for example, it can be determined by a user while checking the degree of noise reduction of the output image, or it can be determined in advance based on an experiment, etc. Furthermore, for example, the stop timing can be determined based on index information obtained from at least the output image for determining the degree of noise reduction of the output image.

[0031] In this embodiment, the control unit 5 is configured to perform DIP processing on the virtual monochromatic image, effective atomic number image, and density distribution image acquired based on the multiple reconstructed images 11. That is, the control unit 5 is configured to acquire a virtual monochromatic image based on the multiple reconstructed images 11 and perform DIP processing on the acquired virtual monochromatic image to generate a noise-reduced virtual monochromatic image from the virtual monochromatic image. Similarly, the control unit 5 is configured to acquire an effective atomic number image based on the multiple reconstructed images 11 and perform DIP processing on the acquired effective atomic number image to generate a noise-reduced effective atomic number image from the effective atomic number image. The control unit 5 is also configured to acquire a density distribution image based on the multiple reconstructed images 11 and perform DIP processing on the acquired density distribution image to generate a noise-reduced density distribution image from the density distribution image.

[0032] (X-ray imaging control processing) Next, control processing relating to X-ray imaging by the X-ray imaging apparatus 100 of this embodiment will be described based on a flowchart with reference to Fig. 5. Each processing in the flowchart is performed by the control unit 5.

[0033] As shown in Fig. 5, first, X-ray CT imaging is performed in step 101. In step 101, while changing the imaging position of the X-ray CT imaging, X-rays are irradiated onto the subject 200 from the X-ray irradiation unit 1, and X-ray photons due to the X-rays that have passed through the subject 200 are detected by the energy discrimination detector 2. In addition, the X-ray photons detected by the energy discrimination detector 2 are counted for each energy range by the counting unit 3. As a result, the counting results of X-ray photons for each energy range for each imaging position are obtained.

[0034] Then, in step 102, a reconstructed image 11 is generated for each energy range of X-ray photons based on the X-ray photon counting results (X-ray photon detection results). In step 102, a plurality of reconstructed images 11 for different energy ranges are generated. In step 102, for example, a reconstructed image 11 corresponding to energy range E1 is generated based on the X-ray photon counting results for energy range E1 at all imaging positions. Also, for example, a reconstructed image 11 corresponding to energy range E4 is generated based on the X-ray photon counting results for energy range E4 at all imaging positions.

[0035] Then, in step 103, a DCNN 7 that does not perform pre-learning using training data is used, and a DIP process that stops processing before the DCNN 7 converges is performed with a plurality of reconstructed images 11 as input to the DCNN 7, thereby generating a plurality of noise-reduced reconstructed images 12 from the plurality of reconstructed images 11. At this time, a plurality of reconstructed images 11 that have different levels of noise and represent the same cross-section of the same subject 200 are input to the DCNN 7. Also, in step 103, by using a plurality of reconstructed images 11 as input, a DIP process is performed in which the DCNN 7 repeatedly updates the parameters of the DCNN 7 while sharing structural information possessed by each of the plurality of reconstructed images 11, thereby generating a plurality of noise-reduced reconstructed images 12. Thereafter, the control process is terminated.

[0036] (Noise reduction effect of DIP processing) Next, the noise reduction effect of the reconstructed image 11 by the DIP processing will be described with reference to FIG.

[0037] The object 200 used had the structure shown in FIG. 6. Specifically, the object 200 was a cylindrical body made of ASA (Acrylate Styrene Acrylonitrile) resin, with a cylindrical member made of PEI (Polyetherimide) resin and a cylindrical member made of ABS (Acrylonitrile Butadiene Styrene) resin embedded in it, providing a cylindrical hollow portion (air portion). The energy ranges were set to four: 10-15 keV, 15-18 keV, 18-22 keV, and 22 keV or higher. Four reconstructed images 11 corresponding to the four energy ranges of 10-15 keV, 15-18 keV, 18-22 keV, and 22 keV or higher were reconstructed by FBP processing.

[0038] All four reconstructed images 11 contain noise. The reconstructed image 11 on the low-energy side contains relatively little noise, while the reconstructed image 11 on the high-energy side contains relatively more noise. When DIP processing was performed on the four reconstructed images 11, four noise-reduced reconstructed images 12 with low noise and high contrast were obtained. Furthermore, despite differences in the level of noise in the original reconstructed images 11, all four noise-reduced reconstructed images 12 had similarly low noise and high contrast. This is thought to be due to the DIP processing performed while the four reconstructed images 11 interact with each other. Furthermore, the four noise-reduced reconstructed images 12 had reduced noise, and the structure of the object 200 (e.g., ASA, PEI, ABS, and air) was clearly identifiable. This demonstrates that DIP processing on the four reconstructed images 11 can produce multiple noise-reduced reconstructed images 12 with low noise and excellent structural information extraction capabilities.

[0039] (Experiment to distinguish resin types with and without DIP treatment) Next, an experiment to distinguish resin types with and without DIP treatment will be described with reference to Figures 7 and 8. Figure 7 shows the correct answer data from the experiment to distinguish resin types, and Figure 8 shows the results of distinguishing resin types without DIP treatment and with DIP treatment.

[0040] The object 200 used had the structure of the correct data shown in Figure 7. Specifically, the object 200 used was a cylindrical body of resin type P1, with multiple cylindrical members of resin type P2 with different diameters and multiple cylindrical members of resin type P3 with different diameters embedded in it. The effective atomic numbers of resin types P1, P2, and P3 were 6.12, 6.30, and 6.40, respectively. The densities of resin types P1, P2, and P3 were 1.12, 1.18, and 1.21 kg / cm, respectively. 3 It was.

[0041] In addition, when DIP processing was not performed, linear attenuation coefficients were obtained from multiple reconstructed images 11 that had only been subjected to FBP processing (without DIP processing), and the resin type was determined based on the obtained linear attenuation coefficients.The discrimination results were compared with the correct data to obtain the accuracy rate.In addition, when DIP processing was performed, linear attenuation coefficients were obtained from multiple noise-reduced reconstructed images 12 that had been subjected to DIP processing on multiple reconstructed images 11 that had been subjected to FBP processing, and the resin type was determined based on the obtained linear attenuation coefficients.The discrimination results were compared with the correct data to obtain the accuracy rate.

[0042] As shown in Figure 8, without DIP processing (FBP processing only), the accuracy rate for resin type discrimination was 60.2%. On the other hand, with DIP processing (FBP processing + DIP processing), the accuracy rate for resin type discrimination was 95.6%. This is thought to be because the low-noise, high-contrast noise-reduced reconstructed image 12 obtained by DIP processing was able to obtain linear attenuation coefficients more accurately than the reconstructed image 11 without DIP processing. This shows that material decomposition can be performed more accurately with DIP processing than without DIP processing.

[0043] (Effects of this embodiment) In this embodiment, the following effects can be obtained.

[0044] In this embodiment, as described above, the X-ray imaging method includes the steps of irradiating X-rays onto the subject 200 and obtaining a detection result of X-ray photons from the X-rays that have passed through the subject 200; generating a plurality of reconstructed images 11 having different energy ranges based on the detection result of the X-ray photons; and generating a plurality of noise-reduced reconstructed images 12 by using a deep convolutional neural network 7 as input to the deep convolutional neural network 7 to perform deep image processing in which the processing is stopped before the deep convolutional neural network 7 converges, with the plurality of reconstructed images 11 being used as input to the deep convolutional neural network 7.

[0045] Furthermore, in this embodiment, as described above, the X-ray imaging apparatus 100 includes an X-ray irradiation unit 1 that irradiates X-rays onto the subject 200, an energy discrimination detector 2 that detects X-ray photons from the X-rays that have passed through the subject 200, and a control unit 5 that generates multiple reconstructed images 11 with different energy ranges based on the detection results of the X-ray photons by the energy discrimination detector 2. The control unit 5 is configured to use a deep convolutional neural network 7 to perform deep image processing in which the processing is stopped before the deep convolutional neural network 7 converges, and to generate multiple noise-reduced reconstructed images 12 by using the multiple reconstructed images 11 as input to the deep convolutional neural network 7.

[0046] As a result, a plurality of noise-reduced reconstructed images 12 are generated by performing deep image processing using a deep convolutional neural network 7, in which processing is stopped before the deep convolutional neural network 7 converges, with a plurality of reconstructed images 11 as input to the deep convolutional neural network 7. As a result, by using deep image processing, a noise-reduced reconstructed image 12 in which noise has been reduced can be easily generated from the reconstructed image 11. Furthermore, by performing deep image processing on a plurality of reconstructed images 11, unlike when deep image processing is simply performed on a single reconstructed image 11, deep image processing can be performed while the plurality of reconstructed images 11 interact with each other, thereby achieving a higher noise reduction effect than when deep image processing is performed on a single reconstructed image 11. As a result, in photon-counting X-ray imaging that detects X-ray photons, noise is sufficiently reduced, and a noise-reduced reconstructed image 12 with improved image quality can be obtained.

[0047] Furthermore, since a noise-reduced reconstructed image 12 with sufficiently reduced noise and improved image quality can be obtained, when material decomposition is performed based on the noise-reduced reconstructed image 12, the material decomposition can be performed with high accuracy.

[0048] Furthermore, in the above embodiment, the following additional effects can be obtained by configuring as follows.

[0049] That is, in this embodiment, as described above, the step of generating the plurality of noise-reduced reconstructed images 12 includes a step of generating the plurality of noise-reduced reconstructed images 12 by performing deep image processing in which the plurality of reconstructed images 11 with different degrees of noise are input to the deep convolutional neural network 7. This allows deep image processing to be performed so as to utilize the low-noise reconstructed images 11 with a large amount of effective information while reducing the noise in the high-noise reconstructed images 11 with a small amount of effective information, thereby effectively reducing the noise in the plurality of reconstructed images 11.

[0050] Furthermore, in this embodiment, as described above, the step of generating the plurality of noise-reduced reconstructed images 12 includes a step of performing deep image processing to generate the plurality of noise-reduced reconstructed images 12 by inputting the plurality of reconstructed images 11 representing the same tomographic plane of the same subject 200 into the deep convolutional neural network 7. This allows deep image processing to be performed while mutually utilizing information on the same tomographic plane of the plurality of reconstructed images 11, thereby easily reducing noise in the plurality of reconstructed images 11 representing the same tomographic plane.

[0051] Furthermore, in this embodiment, as described above, the step of generating the plurality of noise-reduced reconstructed images 12 includes a step of generating the plurality of noise-reduced reconstructed images 12 by using the plurality of reconstructed images 11 as input and performing deep image processing while sharing structural information possessed by each of the plurality of reconstructed images 11 in the deep convolutional neural network 7. This allows deep image processing to be performed while mutually utilizing the structural information of the plurality of reconstructed images 11, making it possible to easily obtain a plurality of noise-reduced reconstructed images 12 that are low in noise and whose structural information is clearly identifiable.

[0052] Furthermore, in this embodiment, as described above, the step of generating the plurality of noise-reduced reconstructed images 12 includes a step of generating the plurality of noise-reduced reconstructed images 12 by performing deep image processing that uses the plurality of reconstructed images 11 as input, repeatedly updating the parameters of the deep convolutional neural network 7 while sharing structural information possessed by each of the plurality of reconstructed images 11 in the deep convolutional neural network 7. This allows common parameters to be automatically set for the plurality of reconstructed images 11 by the deep image processing, and therefore, unlike when the user sets correlation parameters between images to reduce noise, it is possible to reduce noise in the reconstructed images 11 while saving the user effort.

[0053] Furthermore, in this embodiment, as described above, the X-ray imaging method further includes a step of performing depth image processing on the virtual monochromatic image, effective atomic number image, and density distribution image acquired based on the multiple reconstructed images 11. As a result, by using the depth image processing, noise can be reduced from the virtual monochromatic image, effective atomic number image, and density distribution image, and images with improved image quality can be easily generated.

[0054] [Variations] The embodiments disclosed herein should be considered to be illustrative and not restrictive in all respects. The scope of the present invention is defined by the claims rather than the above description of the embodiments, and further includes all modifications (variations) within the meaning and scope of the claims.

[0055] For example, in the above embodiment, an example has been shown in which the subject is rotated relative to the imaging system including the X-ray source and the energy discrimination detector, but the present invention is not limited to this. In the present invention, X-ray CT imaging may be performed while changing the imaging position of the imaging system relative to the subject by rotating the imaging system including the X-ray source and the energy discrimination detector relative to the subject.

[0056] In the above embodiment, the control unit functions as an image processing unit, but the present invention is not limited to this. In the present invention, an image processing unit including a processor such as a GPU (Graphics Processing Unit) may be provided separately and independently from the control unit.

[0057] In addition, in the above embodiment, an example in which the deep image processing is DIP processing is shown, but the present invention is not limited to this. In the present invention, the deep image processing may be a similar processing other than DIP processing as long as it can acquire a noise-reduced reconstructed image.

[0058] In the above embodiment, an example was shown in which deep image processing was performed using a deep convolutional neural network that was not pre-trained using training data, but the present invention is not limited to this. In the present invention, deep image processing may be performed using a deep convolutional neural network that has been pre-trained using training data, as long as a noise-reduced reconstructed image can be obtained.

[0059] In addition, in the above embodiment, an example was shown in which the deep convolutional neural network was configured using U-Net, but the present invention is not limited to this. In the present invention, the deep convolutional neural network may be configured using a deep convolutional neural network other than U-Net.

[0060] In the above embodiment, an example was shown in which deep image processing was performed on the virtual monochromatic image, the effective atomic number image, and the density distribution image, but the present invention is not limited to this. In the present invention, deep image processing may be performed on at least one of the virtual monochromatic image, the effective atomic number image, and the density distribution image. Furthermore, deep image processing may not be performed on the virtual monochromatic image, the effective atomic number image, and the density distribution image.

[0061] [Aspect] It will be appreciated by those skilled in the art that the exemplary embodiments described above are examples of the following aspects.

[0062] (Item 1) irradiating an object with X-rays and obtaining a detection result of X-ray photons due to the X-rays transmitted through the object; generating a plurality of reconstructed images in different energy ranges based on the detection results of the X-ray photons; generating a plurality of noise-reduced reconstructed images by performing deep image processing using a deep convolutional neural network, the processing stopping before the deep convolutional neural network converges, with the plurality of reconstructed images as input to the deep convolutional neural network.

[0063] (Item 2) 2. The X-ray imaging method according to claim 1, wherein the step of generating the plurality of noise-reduced reconstructed images includes a step of generating the plurality of noise-reduced reconstructed images by performing the deep image processing on the deep convolutional neural network, where the plurality of reconstructed images have different degrees of noise.

[0064] (Item 3) 3. The X-ray imaging method according to claim 1, wherein the step of generating the plurality of noise-reduced reconstructed images includes a step of generating the plurality of noise-reduced reconstructed images by performing the deep image processing in which the plurality of reconstructed images representing the same cross-sectional plane of the same subject are input to the deep convolutional neural network.

[0065] (Item 4) 4. The X-ray imaging method according to claim 3, wherein the step of generating the plurality of noise-reduced reconstructed images includes the step of performing the deep image processing using the plurality of reconstructed images as inputs while sharing structural information possessed by each of the plurality of reconstructed images in the deep convolutional neural network, thereby generating the plurality of noise-reduced reconstructed images.

[0066] (Item 5) 5. The X-ray imaging method according to claim 4, wherein the step of generating the plurality of noise-reduced reconstructed images includes the step of performing the deep image processing in which the deep convolutional neural network repeatedly updates parameters of the deep convolutional neural network while sharing structural information of each of the plurality of reconstructed images by using the plurality of reconstructed images as input.

[0067] (Item 6) 6. The X-ray imaging method according to claim 1, further comprising a step of performing the deep image processing on at least one of a virtual monochromatic image, an effective atomic number image, and a density distribution image acquired based on the plurality of reconstructed images.

[0068] (Item 7) an X-ray irradiation unit that irradiates the subject with X-rays; an energy discrimination detector that detects X-ray photons due to the X-rays that have passed through the subject; an image processing unit that generates a plurality of reconstructed images in different energy ranges based on the detection result of the X-ray photons by the energy discrimination detector, the image processing unit is configured to generate a plurality of noise-reduced reconstructed images by using a deep convolutional neural network to perform deep image processing in which processing is stopped before the deep convolutional neural network converges, with the plurality of reconstructed images as inputs to the deep convolutional neural network. [Explanation of symbols]

[0069] 1 X-ray irradiation section 2. Energy discrimination detector 5. Control unit (image processing unit) 7 Deep Convolutional Neural Networks 11 Reconstructed image 12. Noise-reduced reconstructed image 100 X-ray equipment 200 subjects

Claims

1. a step of irradiating an object with X-rays and obtaining a detection result of X-ray photons due to the X-rays that have passed through the object; generating a plurality of reconstructed images in different energy ranges based on the detection results of the X-ray photons; and generating a plurality of noise-reduced reconstructed images, which are images in which noise has been reduced, from the plurality of input reconstructed images by performing deep image processing using a deep convolutional neural network, the processing being stopped before a processing result of the deep convolutional neural network converges, with each of the plurality of reconstructed images as an input to the deep convolutional neural network.

2. 2. The X-ray imaging method according to claim 1, wherein the step of generating the plurality of noise-reduced reconstructed images includes a step of generating the plurality of noise-reduced reconstructed images by performing the deep image processing on the deep convolutional neural network, the deep convolutional neural network inputting the plurality of reconstructed images having different degrees of noise.

3. 3. The X-ray imaging method according to claim 1, wherein the step of generating the plurality of noise-reduced reconstructed images includes a step of generating the plurality of noise-reduced reconstructed images by performing the deep image processing in which the plurality of reconstructed images representing the same slice plane of the same subject are input to the deep convolutional neural network.

4. A step of irradiating an object with X-rays and obtaining a detection result of X-ray photons due to the X-rays that have passed through the object; generating a plurality of reconstructed images in different energy ranges based on the detection results of the X-ray photons; generating a plurality of noise-reduced reconstructed images by performing deep image processing using a deep convolutional neural network, the processing being stopped before a processing result of the deep convolutional neural network converges, using the plurality of reconstructed images as inputs to the deep convolutional neural network; the step of generating the plurality of noise-reduced reconstructed images includes a step of generating the plurality of noise-reduced reconstructed images by performing the deep image processing in which the plurality of reconstructed images representing the same slice plane of the same subject are input to the deep convolutional neural network; the step of generating the plurality of noise-reduced reconstructed images includes a step of generating the plurality of noise-reduced reconstructed images by performing the deep image processing using the plurality of reconstructed images as input while sharing structural information possessed by each of the plurality of reconstructed images in the deep convolutional neural network.

5. 5. The X-ray imaging method according to claim 4, wherein the step of generating the plurality of noise-reduced reconstructed images includes the step of performing the deep image processing in which the deep convolutional neural network repeatedly updates parameters of the deep convolutional neural network while sharing structural information of each of the plurality of reconstructed images by using the plurality of reconstructed images as input.

6. A step of irradiating an object with X-rays and obtaining a detection result of X-ray photons due to the X-rays that have passed through the object; generating a plurality of reconstructed images in different energy ranges based on the detection results of the X-ray photons; generating a plurality of noise-reduced reconstructed images by performing deep image processing using a deep convolutional neural network, the processing being stopped before a processing result of the deep convolutional neural network converges, using the plurality of reconstructed images as inputs to the deep convolutional neural network; The X-ray imaging method further comprises a step of performing the depth image processing on at least one of a virtual monochromatic image, an effective atomic number image, and a density distribution image acquired based on the plurality of reconstructed images.

7. an X-ray irradiation unit that irradiates an object with X-rays; an energy discrimination detector that detects X-ray photons generated by the X-rays that have passed through the subject; an image processing unit that generates a plurality of reconstructed images in different energy ranges based on the detection result of the X-ray photons by the energy discrimination detector, the image processing unit is configured to perform deep image processing using a deep convolutional neural network, the processing being stopped before a processing result of the deep convolutional neural network converges, with each of the plurality of reconstructed images as an input to the deep convolutional neural network, thereby generating a plurality of noise-reduced reconstructed images, which are images in which noise has been reduced, from the plurality of input reconstructed images.

Images