Tomographic γ-scan image reconstruction method based on deep learning

Abstract

The present invention discloses a tomographic gamma scanning image reconstruction method based on deep learning, which uses a two-dimensional convolutional neural network trained with prior data to perform radioactive waste TGS detection image reconstruction using sparse scan data, improving the detection efficiency and the image reconstruction accuracy. The method includes the steps of: S1, determining a sparse scan mode according to a voxel division model; S2, using software related to the Monte Carlo method to establish a simulation model according to the TGS system size; S3, conducting a simulation experiment and establishing a transmission image training data set; S4, building a deep learning neural network and training the deep learning neural network; S5, reconstructing a high-resolution transmission image and an emission image and fusing them; S6, establishing a fused image training data set and training the deep learning neural network; reconstructing the fused image; S7, designing an analysis and processing system to implement the reconstruction of the transmission image and the fused image. Using this method can effectively reduce the number of scans and improve the detection efficiency; improve the quality of the reconstructed image.

Description

Technical Field

[0001] The present invention relates to the field of tomographic gamma scanning measurement of nuclear waste, and specifically discloses a tomographic gamma scanning image reconstruction method based on deep learning. Background Art

[0002] As is well known: According to the relevant regulations on nuclear waste management and disposal, it is required to use standard barrels to encapsulate radioactive waste. Before the temporary storage, transportation and disposal of radioactive waste, it is necessary to accurately measure the composition, content and distribution law of radioactive nuclides in the waste barrel.

[0003] Due to the radioactivity of the waste barrel, non-destructive testing methods are mostly used for measurement. Tomographic Gamma Scanning (TGS) is one of the currently mature non-destructive testing methods. According to the voxel division method of the waste barrel, a gamma detector is used to scan the waste barrel at multiple angles and multiple positions in layers, and the line attenuation coefficient distribution image (transmission image) and the nuclide activity distribution image (emission image) inside the barrel are reconstructed. It can accurately analyze the types, contents and spatial position distributions of radioactive nuclides contained in the sample without any change in the physical and chemical forms of the sample. Usually, the more the number of voxels divided, the more accurate the detection result and the more detailed the information contained in the image. However, fine voxel division requires more scans, increasing a large amount of detection time and reducing work efficiency. While large voxel division, although requiring fewer scans and shorter detection time, the accuracy of the detection result will decrease to a certain extent. Therefore, balancing the improvement of detection efficiency while ensuring the accuracy of the detection result is the key technical problem that the current TGS technology urgently needs to solve. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a tomographic gamma scanning image reconstruction method based on deep learning that uses a two-dimensional convolutional neural network trained with prior data to perform radioactive waste TGS detection image reconstruction, improving detection efficiency and image reconstruction accuracy.

[0005] The technical solution adopted by the present invention to solve its technical problems is: A method for reconstructing a tomographic gamma scanning image of radioactive waste based on deep learning, comprising the following steps:

[0006] S1. Determine the sparse scanning method according to the voxel division model;

[0007] S2. Use software related to the Monte Carlo method to establish a simulation model according to the size of the TGS system;

[0008] S3. Conduct a simulation experiment and establish a transmission image training dataset;

[0009] S4. Build a deep learning neural network; and train the deep learning neural network with the transmission image training dataset obtained in step 3.

[0010] S5. Reconstruct and fuse high-resolution transmission images and emission images through the deep learning neural network trained in step S4.

[0011] S6. Establish a fused image training dataset and train a deep learning neural network; reconstruct the fused image through the trained deep learning neural network.

[0012] S7. Design an analysis and processing system to realize the reconstruction of transmission images and fused images.

[0013] Specifically, in step S1, the voxel division model is a polar coordinate voxel division model or a rectangular grid voxel division model.

[0014] In the Cartesian coordinate system, the voxel is divided into a cube shape; in the polar coordinate system, the voxel is divided into a sector shape; each division method longitudinally divides the waste bin into I tomographies, and each tomography is further divided into J voxel blocks.

[0015] Specifically, in step S2, the simulation model is consistent with the size of the TGS system, and the detector size and type in the simulation model are the same as those of the TGS detector.

[0016] Specifically, in step S2, the Geant4 or MCNP5 Monte Carlo method software is used. In step S3, an external transmission source uses gamma rays of multiple energies to perform transmission scanning on the waste bin. The waste bin is filled with radioactive sources of various shapes, material media, and various types and activities. The ART algorithm is used to reconstruct the large-voxel low-resolution transmission image, and a corresponding small-voxel high-resolution transmission image is established according to the simulation settings. Multiple groups of image pairs are obtained through multiple experiments to establish a transmission image training dataset.

[0017] Furthermore, the simulation experiment in step S3 includes transmission scanning and emission scanning.

[0018] The process of the transmission scanning is as follows:

[0019] 1) The external transmission source and the detector are moved to the initial position, and the external transmission source is turned on to start transmission scanning. The detector detects gamma rays to form an energy spectrum.

[0020] 2) According to the specified sparse scanning method, rotate the waste bin to perform scanning at the next angle until all angles are scanned.

[0021] 3) The transmission source and the detector are moved to the next position, and then all-angle scanning is performed until all angles and positions are scanned.

[0022] The emission scanning process is as follows: turn off the external transmission source, and the waste bin rotation, detector movement, and scanning sequence are the same as those in the transmission scanning. The detector detects the self-emitted gamma rays from the radioactive substances in the waste bin.

[0023] Furthermore, in step S3, the transmission image training data set includes low-resolution reconstructed transmission images and high-resolution reference transmission images;

[0024] The low-resolution reconstructed transmission image is reconstructed using the ART iterative algorithm according to the transmission measurement equation, and the transmission measurement equation is as follows:

[0025]

[0026] In the formula, I i (E) is the intensity of the gamma rays after attenuation through the absorbing material during the i-th scan, and x i,j is the path length of the gamma rays passing through the j-th voxel during the i-th scan. If the gamma rays do not pass through the voxel, then x i,j = 0. J is the number of voxel blocks in each layer of the waste bin. Therefore, μ j is the attenuation coefficient of the j-th voxel for gamma rays with energy E;

[0027] Let the transmittance P i (E) = I i (E) / I0(E), and define the logarithmic transmittance V i (E) = -lnP i (E). Then

[0028]

[0029] Assume that the i-th scan is performed on each layer of the waste bin during a single transmission measurement. Then, through (2), the matrix formula

[0030] X·U(E) = V(E) (3)

[0031] In the formula, X is the path matrix of the gamma rays passing through the single-layer waste bin, U(E) is the attenuation coefficient matrix of the voxel block for gamma rays with energy E, and U(E) = [μ1(E), μ2(E), … μ j (E)] T ; V(E) is the logarithmic transmittance matrix V(E) = [v1, v2, …, v i T ; Obtain the attenuation coefficients in the attenuation coefficient matrix U(E) and convert them into gray values to obtain the low-resolution reconstructed transmission image of the large voxel;

[0032] The ART algorithm is expressed as follows:

[0033] ​

[0034] where k is the number of iterative measurements, i is the serial number of the transmission source scan, j is the serial number of the voxel, λ is the relaxation factor which is a manually set constant (0 < λ < 1), is the attenuation coefficient of the j-th voxel in the k-th iteration, and is the updated attenuation coefficient obtained after the k-th iteration. The iterative process of the ART algorithm can be described as the following six steps:

[0035] 1) Preset the initial attenuation coefficient values of J voxels,

[0036] 2) For the i-th scan, calculate the logarithmic transmittance of the γ-ray after passing through the preset attenuation coefficient,

[0037] 3) Compare the calculated logarithmic transmittance with the experimentally obtained logarithmic transmittance and find the difference: Δ i = v i - v i ';

[0038] 4) Update and correct the attenuation coefficient value of the j-th voxel:

[0039] 5) Repeat steps 2 to 4 until the required number of iterations is reached or the difference is less than the set value;

[0040] 6) Output the attenuation coefficient matrix U(E).

[0041] Specifically, in step S4, the deep learning neural network adopts any one of the recurrent neural network (RNN), convolutional neural network (CNN), and generative adversarial network (GAN).

[0042] Furthermore, in step S5, the trained neural network is used to reconstruct a high-resolution transmission image, and on this basis, an emission image is reconstructed, and finally fused into the initial TGS image;

[0043] Reconstructing the high-resolution transmission image in step S5 is to use the deep learning neural network in step S4 to reconstruct the large-voxel low-resolution reconstructed image in the transmission image training dataset again to improve the resolution of the transmission image. On this basis, the emission image is reconstructed and fused according to the emission measurement equation. The emission measurement equation is as follows:

[0044]

[0045] where A ij(E) is the radioactivity in the j-th voxel during the i-th measurement, N ij (E) is the number of γ photons with energy E emitted in the j-th voxel detected by the detector per unit time. α(E) is the γ-ray branching ratio for energy E. It can be easily obtained from reference materials. ρ ij is the detection efficiency of the detector for γ photons emitted in voxel j and can be obtained by efficiency calibration through experimental methods. η ij (E) is the attenuation factor of the γ photons emitted in voxel j before reaching the detector and can be described as:

[0046]

[0047] In the formula, K is the voxel through which the γ rays with energy E emitted in voxel j pass on the way to the detector during the i-th emission measurement. μ k (E) is the attenuation coefficient of the k-th voxel for this ray. The transmission image of γ rays with energy E, i.e., the attenuation coefficient distribution, can be obtained by fitting the transmission images of γ rays with multiple energies. x ik is the track length of the γ rays passing through voxel k;

[0048] According to Equation (5), the total number of γ photons with energy E emitted in voxel J obtained by the detector during the i-th emission measurement is

[0049]

[0050] Let h i (E) = N ij (E) / α(E) and w ij (E) = ρ ij η ij , then the matrix equation

[0051] A·W = H (8)

[0052] In the formula, A is the radioactivity matrix in voxel J, A = [A1(E), A2(E), …, A j (E)], W is a J×I order attenuation correction matrix, H = [h1(E), h2(E), …, h i (E)];

[0053] Similarly, the ART algorithm is used to iteratively solve the radioactivity matrix A, and the radioactivity in the matrix is converted into gray values to obtain the emission image.

[0054] It has that in step S6, according to the filling setting in the waste bin in the simulation, a TGS reference image and the corresponding TGS initial image are constructed to form an image pair. Multiple simulations are performed to construct a fused image training dataset to train the neural network.

[0055] Further, in the image fusion described in step S5, the transmission image and each pixel in the emission image are respectively placed into any two channels of the RGB image, and the weighted sum of the pixels in the two images is placed into the third channel to obtain the TGS image.

[0056] Further, in step S6, based on all the transmission images in the transmission image training dataset, combined with the data of the simulated emission scan in step S3, the emission images corresponding to all the transmission images are reconstructed using the method in S5 and fused to obtain the initial TGS image.

[0057] Further, according to the settings of the position and shape of the radiation source in the simulated emission scan in step S3, a reference emission image is established. The reference emission image and the reference image in the transmission image training dataset in step S3 are fused into a TGS reference image using the image fusion method in step S5. The reconstructed initial TGS image and the corresponding TGS reference image are saved in pairs to establish a fused image training dataset. After the fused image training dataset is constructed, a deep learning neural network 2 is built again to remove the noise and artifacts in the initial TGS image and reconstruct the initial TGS image into a high-quality TGS image. Preferably, the deep learning neural network 2 is the same as that in step S4 and is trained using the fused image training dataset.

[0058] Further, in step S7, a data automation software is written using the C++ or Java computer language to complete the analysis and processing system. The data processing of the analysis and processing system includes the following processes:

[0059] 1) Input the transmission scan data and calculate the transmission γ photon count;

[0060] 2) Use the ART algorithm to reconstruct the large-voxel low-resolution transmission image;

[0061] 3) The deep learning neural network trained in step S4 reconstructs the large-voxel low-resolution transmission image into a small-voxel high-resolution transmission image;

[0062] 4) Input the emission scan data, judge the type of radionuclide in the bucket, and the characteristic γ-ray energy;

[0063] 5) Fit the transmission image of the characteristic γ-ray, combine the reconstructed small-voxel high-resolution transmission image, and reconstruct and fuse the emission image using the ART algorithm into the initial TGS image;

[0064] 6) The deep learning neural network trained in step 6 reconstructs the initial TGS image;

[0065] 7) Output the reconstructed high-quality high-resolution TGS image.

[0066] The beneficial effects of the present invention are as follows: The method for reconstructing tomographic γ-scan images of radioactive waste based on deep learning according to the present invention has the following advantages:

[0067] 1. By sparse sampling, the number of scans is reduced, and the detection efficiency is improved. High-resolution detection images require a very large number of transmission and emission scans, which reduces the detection efficiency. By reducing the angular and position scans, the number of scans is effectively reduced, and the detection efficiency is improved.

[0068] 2. Use deep learning methods to reconstruct low-resolution images created from undersampled data, and improve the resolution of the images. Through the training of prior data, the characteristic information of the detection images is deeply mined, thereby reducing the dependence on detection data. High-resolution transmission images are reconstructed using a small amount of data, and on this basis, the entire detection image is reconstructed.

[0069] 3. Compared with traditional iterative methods and analytical methods, deep learning can reconstruct high-resolution images faster and with higher quality. The trained neural network can be directly used to directly reconstruct the input low-resolution images. The reconstruction process does not require multiple iterations, which speeds up the reconstruction speed. Experiments have proven that the images reconstructed by deep learning have higher quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic flow diagram of the method for reconstructing tomographic γ-scan images based on deep learning according to the present invention;

[0071] Figure 2 It is a schematic diagram of the TGS system according to an embodiment of the present invention;

[0072] Figure 3 It is a schematic diagram of the polar coordinate voxel division method provided by an embodiment of the present invention;

[0073] Figure 4 It is a schematic diagram of the transmission measurement according to an embodiment of the present invention;

[0074] Figure 5 It is a schematic diagram of the emission measurement according to an embodiment of the present invention;

[0075] Figure 6 It is a schematic diagram of the transmission image training data set according to an embodiment of the present invention;

[0076] Figure 7 It is a schematic diagram of the fused image training data set according to an embodiment of the present invention;

[0077] In the figure: 10 - vertical mechanical transmission device, 20 - transmission source, 30 - detector system, 40 - horizontal and rotary mechanical transmission device, 50 - waste bin, 60 - control and analysis system, 70 - voxel division layer, 80 - voxel block, 1 - detector collimator, 2 - detector. Detailed implementation manners

[0078] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0079] Combined with the attached Figures 1 to 7 As shown, the reconstruction method of the tomographic γ scan image of radioactive waste based on deep learning according to the present invention includes the following steps:

[0080] S1. Determine the sparse scanning method according to the voxel division model;

[0081] In the step S1, the voxel division model is a polar coordinate voxel division model or a rectangular grid voxel division model;

[0082] In the rectangular coordinate system, the voxel is divided into a cube shape; in the polar coordinate system, the voxel is divided into a sector shape; each division method longitudinally divides the waste bin into I tomographies, and each tomography is further divided into J voxel blocks.

[0083] S2. Use the software related to the Monte Carlo method to establish a simulation model according to the size of the TGS system;

[0084] In the step S2, the simulation model is consistent with the size of the TGS system, and the size and type of the detector in the simulation model are the same as those of the TGS detector; use the Geant4 or MCNP5 Monte Carlo method software,

[0085] S3. Conduct a simulation experiment and establish a transmission image training data set;

[0086] The external transmission source uses γ rays of multiple energies to perform transmission scanning on the waste bin. The waste bin is filled with radioactive sources of various shapes, material media, and various types and activities. Use the ART algorithm to reconstruct the large-voxel low-resolution transmission image, and establish the corresponding small-voxel high-resolution transmission image according to the simulation settings; multiple groups of image pairs are obtained through multiple experiments to establish a transmission image training data set.

[0087] The simulation experiment includes transmission scanning and emission scanning;

[0088] The process of the transmission scanning is as follows:

[0089] 1) The external transmission source and the detector are moved to the initial position, and the external transmission source is turned on to start transmission scanning. The detector detects γ rays to form an energy spectrum;

[0090] 2) According to the formulated sparse scanning method, rotate the waste bin to perform scanning at the next angle, and complete scanning at all angles;

[0091] 3) The transmission source and the detector are moved to the next position, and then scanning at all angles is performed until scanning at all angles and positions is completed;

[0092] The emission scanning process is as follows: turn off the external transmission source, and the waste bin rotation, detector movement, and scanning sequence are the same as those in transmission scanning. The detector detects the self-emitted gamma rays from the radioactive substances in the waste bin.

[0093] The transmission image training dataset includes low-resolution reconstructed transmission images and high-resolution reference transmission images;

[0094] The low-resolution reconstructed transmission images are reconstructed using the ART iterative algorithm according to the transmission measurement equation, and the transmission measurement equation is as follows:

[0095]

[0096] In the formula, I i (E) is the intensity of gamma rays after attenuation through the absorbing material during the i-th scan, and x i,j is the track length of gamma rays passing through the j-th voxel during the i-th scan. If it does not pass through the voxel, then x i,j = 0, J is the number of voxel blocks in each layer of the waste bin, so μ j is the attenuation coefficient of the j-th voxel for gamma rays with energy E;

[0097] Let the transmittance P i (E) = I i (E) / I0(E), and define the logarithmic transmittance V i (E) = -lnP i (E), then

[0098]

[0099] Assume that each layer of the waste bin is scanned i times in a single transmission measurement. Then, through (2), the matrix formula

[0100] X·U(E) = V(E) (3)

[0101] In the formula, X is the track matrix of gamma rays passing through a single layer of the waste bin, U(E) is the attenuation coefficient matrix of voxel blocks for gamma rays with energy E, U(E) = [μ1(E), μ2(E), … μ j (E)] T ; V(E) is the logarithmic transmittance matrix V(E) = [v1, v2, …, v i T ; Obtain the attenuation coefficients in the attenuation coefficient matrix U(E) and convert them into gray values to obtain the low-resolution reconstructed transmission images of large voxels;

[0102] The ART algorithm is expressed as follows:

[0103] ​

[0104] Where k is the number of iterative measurements, i is the serial number of the transmission source scan, j is the serial number of the voxel, and λ is the relaxation factor, which is a manually set constant (0 < λ < 1). Then it is the attenuation coefficient of the j-th voxel in the k-th iteration, and is the attenuation coefficient obtained after update after the k-th iteration. The iterative process of the ART algorithm can be described as the following six steps:

[0105] 1) Preset the initial attenuation coefficient values of J voxels.

[0106] 2) For the i-th scan, calculate the logarithmic transmittance of the γ-ray after passing through the preset attenuation coefficient.

[0107] 3) Compare the calculated logarithmic transmittance with the experimentally obtained logarithmic transmittance and find the difference: Δ i = v i - v i ';

[0108] 4) Update and correct the attenuation coefficient value of the j-th voxel:

[0109] 5) Repeat steps 2 to 4 until the required number of iterations is reached or the difference is less than the set value.

[0110] 6) Output the attenuation coefficient matrix U(E).

[0111] S4. Build a deep learning neural network; and train the deep learning neural network with the transmission image training dataset obtained in step 3; the deep learning neural network adopts any one of the recurrent neural network (RNN), convolutional neural network (CNN), and generative adversarial network (GAN).

[0112] Use the trained neural network to reconstruct a high-resolution transmission image, and on this basis, reconstruct an emission image, and finally fuse them into the initial TGS image;

[0113] Reconstructing the high-resolution transmission image in step S5 is to use the deep learning neural network in step S4 to reconstruct the large-voxel low-resolution reconstructed image in the transmission image training dataset again to improve the resolution of the transmission image. On this basis, reconstruct the emission image according to the emission measurement equation and perform fusion. The emission measurement equation is as follows:

[0114]

[0115] where A ij (E) is the radioactivity in the j-th voxel during the i-th measurement, N ij (E) is the number of γ photons with energy E emitted from the j-th voxel detected by the detector per unit time. α(E) is the γ-ray branching ratio for energy E, which can be easily obtained from reference materials. ρ ij is the detection efficiency of the detector for γ photons emitted from the j-th voxel, which can be obtained by efficiency calibration through experimental methods. η ij (E) is the attenuation factor of the γ photons emitted from the j-th voxel before reaching the detector, which can be described as:

[0116]

[0117] where K is the voxel through which the γ rays with energy E emitted from voxel j pass on the way to the detector during the i-th emission measurement. μ k (E) is the attenuation coefficient of the k-th voxel for this ray, and the transmission image of γ rays with energy E, i.e., the attenuation coefficient distribution, can be obtained by fitting the transmission images of γ rays with multiple energies. x ik is the path length of the γ rays passing through voxel k;

[0118] According to Equation (5), the total number of γ photons with energy E emitted from J voxels obtained by the detector during the i-th emission measurement is

[0119]

[0120] Let h i (E) = N ij (E) / α(E) and w ij (E) = ρ ij η ij , then the matrix equation

[0121] A·W = H (8)

[0122] where A is the radioactivity matrix in J voxels, A = [A1(E), A2(E), …, A j (E)], W is a J×I order attenuation correction matrix, H = [h1(E), h2(E), …, h i (E)];

[0123] Similarly, the ART algorithm is used to iteratively solve for the radioactivity matrix A, and the emission image can be obtained by converting the radioactivity in the matrix into gray values.

[0124] S5. Reconstruct and fuse the high-resolution transmission image and emission image through the deep learning neural network trained in step S4;

[0125] S6. Establish a fused image training dataset and train a deep learning neural network; reconstruct the fused image through the trained deep learning neural network; construct a TGS reference image and the corresponding initial TGS image according to the filling settings in the waste bin during simulation to form an image pair, and conduct multiple simulations to construct the fused image training dataset and train the neural network.

[0126] S7. Design an analysis and processing system to achieve the reconstruction of transmission images and fused images;

[0127] Use C++ or Java computer language to write data automation software to complete the analysis and processing system. The data processing of the analysis and processing system includes the following processes:

[0128] 1) Input the transmission scan data and calculate the transmission γ photon count;

[0129] 2) Use the ART algorithm to reconstruct the large-voxel low-resolution transmission image;

[0130] 3) The deep learning neural network trained in step S4 reconstructs the large-voxel low-resolution transmission image into a small-voxel high-resolution transmission image;

[0131] 4) Input the emission scan data, determine the types of radionuclides in the bucket, and the characteristic γ-ray energy;

[0132] 5) Fit the transmission image of the characteristic γ-ray, combine the reconstructed small-voxel high-resolution transmission image, and reconstruct the emission image through the ART algorithm and fuse it into the initial TGS image;

[0133] 6) The deep learning neural network trained in step 6 reconstructs the initial TGS image;

[0134] 7) Output the reconstructed high-quality and high-resolution TGS image.

[0135] Embodiment

[0136] As Figure 1 shown, this embodiment provides a method for reconstructing tomographic γ-scan images based on deep learning, including the following steps:

[0137] S1: Determine the sparse scan method according to the voxel division model;

[0138] S2: Use Monte Carlo method-related software such as Geant4 to establish a simulation model according to the TGS system size;

[0139] S3: Conduct simulation experiments and establish a transmission image training dataset;

[0140] S4: Build and train the deep learning neural network 1;

[0141] S5: Reconstruct the high-resolution transmission image and emission image and fuse them;

[0142] S6: Establish a fused image training dataset and train the deep learning neural network 2;

[0143] S7: Design an analysis and processing system to achieve automated data processing.

[0144] In this embodiment, a self-developed TGS system is used as Figure 2 shown, which includes a vertical mechanical transmission device A, a transmission source B, a detector system C, a horizontal and rotary mechanical transmission device D, a control and analysis system F, a detector collimator 1, a detector 2, etc. The polar coordinate division method is used as Figure 2 , and the waste bin is vertically divided into 3 layers, and each layer is divided into 96 voxel blocks. According to the voxel division method, each layer of the waste bin is scanned at 4 horizontal positions, and 24 angular positions at each horizontal position.

[0145] A 1:1 simulation model of the TGS system is established for simulation experiments. In the simulation model, the waste bin sample is filled with media of various shapes and materials, and the filling positions are randomly selected. Radioactive sources of different sizes, shapes, and radioactivities are randomly filled at any position in the waste bin. A variety of energy γ-rays are selected as the external transmission source to perform transmission measurements on the waste bin.

[0146] The simulation experiment in this embodiment includes two parts: transmission scanning and emission scanning.

[0147] The process of transmission scanning is as follows:

[0148] 1) The external transmission source and the detector are moved to the initial position as Figure 4 shown, the external transmission source is turned on to start transmission scanning, and the detector detects γ-rays to form an energy spectrum;

[0149] 2) The waste bin rotates 15°, and the next angle is scanned. A total of 360° is rotated to complete 24 scans;

[0150] 3) The transmission source and the detector are moved to the next horizontal position, and then all-angle scans are performed to complete all-angle scans at 4 horizontal positions, a total of 96 scans;

[0151] 4) The transmission source and the detector are moved to the next vertical position, and steps 2) and 3) are repeated until all-angle and all-position scans are completed, a total of 288 scans;

[0152] The process of emission scanning is: turn off the external transmission source, and the rotation of the waste bin, the movement and scanning sequence of the detector are the same as those of the transmission scanning. The detector detects the γ-rays spontaneously emitted by the radioactive source in the waste bin, as Figure 5 shown.

[0153] Reconstruct the large-voxel low-resolution transmission image using the ART iterative algorithm according to the transmission measurement equation, which is as follows:

[0154]

[0155] where I i (E) is the intensity of the γ-ray after attenuation through the absorbing material during the i-th scan, and x i,j is the track length of the γ-ray passing through the j-th voxel during the i-th scan. If it does not pass through the voxel, x i,j is recorded as 0. J is the number of the voxel block in each layer of the waste bin. Therefore, μ j is the attenuation coefficient of the j-th voxel for the γ-ray with energy E.

[0156] Let the transmittance P i (E) = I i (E) / I0(E), and define the logarithmic transmittance V i (E) = -lnP i (E). Then

[0157]

[0158] In one transmission measurement, 96 scans are performed for each layer of the waste bin. Then, through (10), the matrix formula

[0159] X·U(E) = V(E) (3)

[0160] can be obtained. Here, X is the track matrix of the γ-ray passing through the single-layer waste bin, U(E) is the attenuation coefficient matrix of the voxel block for the γ-ray with energy E, and U(E) = [μ1(E), μ2(E), … μ j (E)] T ; V(E) is the logarithmic transmittance matrix V(E) = [v1, v2, …, v i T . Obtain the attenuation coefficients in the attenuation coefficient matrix U(E) and convert them into gray values to obtain the large-voxel low-resolution reconstructed transmission image as shown in the left figure below. Figure 6 As shown in the left figure.

[0161] The ART algorithm mentioned above can be expressed as follows:

[0162]

[0163] where k is the number of iteration measurements, k = 200, i is the serial number of the transmission source scan, j is the serial number of the voxel, λ is the relaxation factor, λ = 0.5, is the attenuation coefficient of the j-th voxel in the k-th iteration, and ​That is the attenuation coefficient obtained after updating in the k-th iteration. The iterative process of the ART algorithm can be described by the following six steps:

[0164] 1) Preset the initial attenuation coefficient values of J voxels.

[0165] 2) For the i-th scan, calculate the logarithmic transmittance of the γ-ray after passing through the preset attenuation coefficient.

[0166] 3) Compare the calculated logarithmic transmittance with the experimentally obtained logarithmic transmittance and find the difference: Δ i = v i - v' i ;

[0167] 4) Update and correct the attenuation coefficient value of the j-th voxel:

[0168] 5) Repeat steps 2 to 4 until the required number of iterations is reached or the difference is less than the set value.

[0169] 6) Output the attenuation coefficient matrix U.

[0170] Use the same voxel division method as in step S1 to divide the waste bin into small voxels again. Each large voxel is further divided into 16 small voxels. Based on the attenuation coefficients, positions, and shapes of the filling medium and radiation source in the simulation, establish a high-resolution reference transmission image of the small voxels as shown in the right figure below. Save the reconstructed image and the reference image in pairs, and perform 10,000 simulations to construct a training dataset of transmission images. Figure 6 As shown in the right figure. Save the reconstructed image and the reference image in pairs, and perform 10,000 simulations to construct a training dataset of transmission images.

[0171] In this embodiment, the deep learning neural network selects the Residual Network (ResNet). The neural network is trained with the training dataset of transmission images to reconstruct the low-resolution transmission images of the large voxels in the training dataset of transmission images again, improve the resolution of the transmission images, and on this basis, reconstruct and fuse the emission images according to the emission measurement equation as follows:

[0172]

[0173] In the formula, A ij (E) is the radioactivity in the j-th voxel during the i-th measurement, N ij (E) is the number of γ photons with energy E emitted from the j-th voxel detected by the detector per unit time. α(E) is the γ-ray branching ratio with energy E, which can be easily obtained by referring to data. ρ ij is the detection efficiency of the detector for the γ photons emitted in the j-th voxel, which can be obtained by efficiency calibration through experimental methods. η ij(E) is the attenuation factor of the γ photons emitted from the j-th voxel before reaching the detector, which can be described as:

[0174]

[0175] In the formula, K is the voxel that the γ ray with energy E emitted from the j-th voxel passes through on the way to the detector during the i-th emission measurement, and μ k (E) is the attenuation coefficient of the k-th voxel for this ray. The transmission image of the γ ray with energy E, that is, the attenuation coefficient distribution, can be obtained by fitting the transmission images of γ rays with multiple energies. x ik is the track length of the γ ray passing through the k-th voxel.

[0176] According to Equation (12), the total number of γ photons with energy E emitted from the J voxels obtained by the detector during the i-th emission measurement can be obtained as

[0177]

[0178] Let h i (E) = N ij (E) / α(E) and w ij (E) = ρ ij η ij , then the matrix equation

[0179] A·W = H (15)

[0180] In the formula, A is the radioactivity matrix in the J voxels, A = [A1(E), A2(E), …, A j (E)], W is a J×I order attenuation correction matrix, H = [h1(E), h2(E), …, h i (E)]. Similarly, the ART algorithm can be used to iteratively solve the radioactivity matrix A, and the emission image can be obtained by converting the radioactivity in the matrix into gray values.

[0181] Put each pixel in the transmission image and the emission image into any two channels of the RGB image respectively, and put the weighted sum of the pixels in the two images into the third channel to obtain the initial TGS image as Figure 7 shown in the left figure.

[0182] Based on all the transmission images in the transmission image training dataset, combined with the data of the simulated emission scan in step S3, reconstruct the emission images corresponding to all the transmission images and fuse them to obtain the initial TGS image. According to the settings of the position and shape of the radiation source in the simulated emission scan in step S3, establish a reference emission image, and use the image fusion method in step S5 to fuse the reference emission image with the reference transmission image in the transmission image training dataset into a TGS reference image as Figure 7As shown in the right figure. The reconstructed initial TGS images and the corresponding TGS reference images are saved in pairs to establish a fused image training dataset. After the fused image training dataset is constructed, a deep learning neural network 2 is built again to remove the noise and artifacts in the initial TGS images and reconstruct the initial TGS images into high-quality TGS images. The deep learning neural network 2 also selects the Residual Network (ResNet) and is trained using the fused image training dataset.

[0183] Use Python to write data automation software to complete the analysis and processing system. The data processing includes the following processes:

[0184] 1) Input the transmission scan data and calculate the transmission γ photon count;

[0185] 2) Use the ART algorithm to reconstruct the large-voxel low-resolution transmission image;

[0186] 3) The trained deep learning neural network 1 reconstructs the large-voxel low-resolution transmission image into a small-voxel high-resolution transmission image;

[0187] 4) Input the emission scan data, determine the types of radionuclides in the bucket, and the characteristic γ-ray energy;

[0188] 5) Fit the transmission image of the characteristic γ-ray, combine the reconstructed small-voxel high-resolution transmission image, and reconstruct and fuse the emission image into the initial TGS image through the ART algorithm;

[0189] 6) The trained deep learning neural network 2 reconstructs the initial TGS image;

[0190] 7) Output the reconstructed high-quality and high-resolution TGS image.

[0191] So far, the entire tomographic γ-scan image reconstruction method based on deep learning is completed. During detection, use the TGS system to scan according to the sparse scan method in the simulation, input the transmission scan data and the emission scan data into the analysis and processing system, and feedback to obtain the reconstructed high-quality and high-resolution TGS image.

Claims

1. A method for reconstructing a tomographic γ-scan image of radioactive waste based on deep learning, characterized in that, It includes the following steps: S1. Determine the sparse scanning method according to the voxel division model; S2. Use software related to the Monte Carlo method to establish a simulation model according to the size of the TGS system; The related software uses Geant4 or MCNP5 Monte Carlo method software; S3. Conduct a simulation experiment and establish a transmission image training dataset; The external transmission source uses γ rays of multiple energies to perform transmission scanning on the waste bin. The waste bin is filled with various media of different shapes and materials and various types and activities of radiation sources. Use the ART algorithm to reconstruct the low-resolution transmission image of large voxels, and establish the corresponding high-resolution transmission image of small voxels according to the simulation settings; Multiple experiments obtain multiple groups of image pairs to establish a transmission image training dataset; The simulation experiment includes transmission scanning and emission scanning; The process of the transmission scanning is as follows: 1) Move the external transmission source and the detector to the initial position, turn on the external transmission source to start transmission scanning, and the detector detects γ rays to form an energy spectrum; 2) According to the formulated sparse scanning method, rotate the waste bin to perform scanning at the next angle until all angles are scanned; 3) Move the transmission source and the detector to the next position and then perform scanning at all angles until scanning at all angles and positions is completed; The process of the emission scanning is: turn off the external transmission source, and the rotation of the waste bin, the movement of the detector, and the scanning sequence are the same as those of the transmission scanning. The detector detects the self-emitted γ rays of the radiation source in the waste bin; The transmission image training dataset includes a low-resolution reconstructed transmission image and a high-resolution reference transmission image; The low-resolution reconstructed transmission image is reconstructed using the ART iterative algorithm according to the transmission measurement equation. The transmission measurement equation is as follows: Where, I i (E) is the intensity of γ-rays after attenuation through the absorbing material during the i-th scan, and x i,j is the path length of the γ-rays through the j-th voxel during the i-th scan. If the γ-rays do not pass through the voxel, x i,j is recorded as 0. J is the number of the voxel block in each waste bin layer. Therefore, μ j is the attenuation coefficient of the j-th voxel for γ-rays with energy E; Let the transmittance be P i (E) = I i (E) / I0(E), and define the logarithmic transmittance V i (E) = -lnP i (E), then Assume that in one transmission measurement, the waste bin is scanned i times for each layer. Then, through (2), the matrix formula can be obtained X·U(E)=V(E) (3) Wherein, X is the track matrix of γ-rays passing through the single-layer waste bin, U(E) is the attenuation coefficient matrix of the voxel block for γ-rays with energy E, and U(E) = [μ1(E), μ2(E), … μ j (E)] T ; V(E) is the logarithmic transmittance matrix V(E) = [v1, v2, …, v i T ; Obtaining the attenuation coefficients in the attenuation coefficient matrix U(E) and converting them into gray values to obtain the reconstructed transmission image of the large voxel with low resolution;​ The ART algorithm is expressed as follows: where k is the number of iterative measurements, i is the serial number of the transmission source scan, j is the serial number of the voxel, λ is the relaxation factor which is a constant set artificially (0 < λ < 1), is the attenuation coefficient of the j-th voxel in the k-th iteration, and is the attenuation coefficient updated after the k-th iteration; The iterative process of the ART algorithm can be described by the following six steps: 1) Preset the initial attenuation coefficient value of J voxels, 2) For the i-th scan, calculate the logarithmic transmittance of the gamma rays after passing through the preset attenuation coefficient. 3) Compare the calculated logarithmic transmittance with the experimentally obtained logarithmic transmittance and find the difference: Δ i = v i - v i '; 4) Update and correct the attenuation coefficient value of the j-th voxel: 5) Repeat steps 2 to 4 until the required number of iterations is reached or the difference is less than the set value; 6) Output the attenuation coefficient matrix U(E); S4. Build a deep learning neural network; And train the deep learning neural network through the transmission image training dataset obtained in step 3; S5. Reconstruct and fuse the high-resolution transmission image and the emission image through the deep learning neural network trained in step S4; S6. Establish a fused image training dataset and train the deep learning neural network; Reconstruct the fused image through the trained deep learning neural network; S7. Design an analysis and processing system to realize the reconstruction of the transmission image and the fused image.

2. The reconstruction method of the tomographic γ scanning image of radioactive waste based on deep learning according to claim 1, characterized in that: The voxel division model in step S1 is a polar coordinate voxel division model or a rectangular grid voxel division model; In the rectangular coordinate system, the voxels are divided into cube shapes; in the polar coordinate system, the voxels are divided into fan-shaped bodies; Each division method divides the waste bin longitudinally into I tomographies, and each tomography is further divided into J voxel blocks.

3. The reconstruction method of the tomographic γ-scan image of radioactive waste based on deep learning according to claim 2, characterized in that: In step S2, the simulation model is consistent with the size of the TGS system, and the size and type of the detector in the simulation model are the same as those of the TGS detector.

4. The reconstruction method of the tomographic γ-scan image of radioactive waste based on deep learning according to claim 1, characterized in that: In step S4, the deep learning neural network adopts any one of the neural networks, namely, Recurrent Neural Network (RNN), Convolutional Neural Networks (CNN), and Generative Adversarial Network (GAN).

5. The reconstruction method of the tomographic γ-scanning image of radioactive waste based on deep learning according to claim 4, characterized in that: In step S5, the trained neural network is used to reconstruct the high-resolution transmission image, and on this basis, the emission image is reconstructed, and finally fused into the initial TGS image; In step S5, to reconstruct the high-resolution transmission image, the deep learning neural network in step S4 is used to reconstruct the large-voxel low-resolution reconstructed image in the transmission image training dataset again to improve the resolution of the transmission image. On this basis, the emission image is reconstructed and fused according to the emission measurement equation. The emission measurement equation is as follows: Where A ij (E) is the radioactivity in the j-th voxel at the i-th measurement, N ij (E) is the number of γ photons with energy E emitted in the j-th voxel detected by the detector per unit time. The γ-ray branching ratio α(E) for energy E can be easily obtained from reference materials. ρ ij is the detection efficiency of the detector for the γ photons emitted in the j-th voxel and can be obtained by efficiency calibration through experimental methods. η ij (E) is the attenuation factor of the γ photons emitted in the j-th voxel before reaching the detector, described as: where K is the voxel that the γ-ray with energy E emitted from voxel j passes through on its way to the detector during the i-th emission measurement, and μ k (E) is the attenuation coefficient of the k-th voxel for this ray. The transmission image of γ-rays with energy E, that is, the attenuation coefficient distribution, can be obtained by fitting the transmission images of γ-rays with multiple energies. x ik is the track length of the γ-ray traveling through voxel k; According to Equation (5), in the i-th emission measurement, the total number of γ photons with energy E emitted by the detector in J voxels is Let h i (E)=N ij (E) / α(E) and w ij (E)=ρ ij η ij , then the matrix equation is obtained: A·W = H (8) where, A is the radioactivity matrix in the J voxel, A = [A1(E), A2(E), …, A j (E)], W is a J×I order attenuation correction matrix, H = [h1(E), h2(E), …, h i (E)]; Similarly, the ART algorithm is used to iteratively solve the activity matrix A, and the activity in the matrix is converted into gray values to obtain the emission image.

6. The reconstruction method of the tomographic γ scanning image of radioactive waste based on deep learning according to claim 5, characterized in that: In step S6, according to the settings of the fillers in the waste bin in the simulation, the TGS reference image and the corresponding initial TGS image are constructed to form an image pair. Multiple simulations are performed to construct a fused image training dataset to train the neural network.

7. The reconstruction method of the tomographic γ-scan image of radioactive waste based on deep learning according to claim 5, characterized in that: In step S7, a data automation software is written using the C++ or Java computer language to complete the analysis and processing system. The data processing of the analysis and processing system includes the following processes: 1) Input the transmission scan data and calculate the transmission γ photon count; 2) Use the ART algorithm to reconstruct the large-voxel low-resolution transmission image; 3) The trained deep learning neural network in step S4 reconstructs the large-voxel low-resolution transmission image into a small-voxel high-resolution transmission image; 4) Input the emission scan data, and determine the types of radionuclides in the bucket and the characteristic γ-ray energy; 5) Fit the transmission image of the characteristic γ-ray, and combine the reconstructed small-voxel high-resolution transmission image to reconstruct the emission image by the ART algorithm and fuse it into the initial TGS image; 6) The trained deep learning neural network in step 6 reconstructs the initial TGS image; 7) Output the reconstructed high-quality high-resolution TGS image.

Images