CT full-chain intelligent reconstruction method based on deep learning

Abstract

According to the CT full-chain intelligent reconstruction method based on deep learning, original measurement data is sequentially processed by a trained pixel-by-pixel intelligent correction network, a trained angle-by-angle intelligent filtering network and a trained back projection tensor intelligent reconstruction network to finally obtain a final CT reconstruction image. According to the method, firstly, the pixel-by-pixel intelligent correction network can perform pixel-by-pixel adaptive Gaussian filtering on low-dose original measurement data by learning the variance of each pixel point; and the angle-by-angle intelligent filtering network can carry out angle-by-angle adaptive filtering on chordal graph data by learning the filtering kernel of each projection angle. And finally, the back projection tensor intelligent reconstruction network learns the mapping of the back projection tensor to the target image, so that the image quality can be ensured while the scanning dose is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and computer tomography, and in particular to a CT full-chain intelligent reconstruction method based on deep learning. Background Art

[0002] Computed tomography (CT) is widely used in important fields such as medical imaging, security inspection, and industrial inspection due to its non-invasive three-dimensional high-resolution imaging capabilities. However, the high radiation dose of medical CT imaging has caused people to worry about the potential cancer risks brought by X-rays. Under the premise of ensuring image quality, the principle of the lowest radiation dose possible has become a common goal pursued by everyone. Reducing the scanning dose will inevitably introduce serious noise artifacts in the reconstructed image, which in turn affects the clinical diagnostic performance. Therefore, improving the quality of low-dose CT images has always been one of the key technologies in the field of medical imaging.

[0003] In recent years, deep learning technology has been widely used in medical image reconstruction and has achieved unprecedented success. Low-dose CT data reconstruction algorithms based on deep learning have been developed in large quantities, and key modules in the CT reconstruction pipeline are learned in a data-driven way to enhance image quality, including chord domain network, image domain network and dual-domain network. Among them, the chord domain network directly restores the chord data affected by noise, and then reconstructs the CT image through filtered back projection; the image domain network belongs to the image post-processing technology, and realizes noise suppression by learning the nonlinear mapping from low-dose images to target images; the dual-domain network realizes the joint optimization of the chord sub-network and the image sub-network by constructing a differentiable filtered back projection layer between the chord domain and the image domain, making full use of the integral information of the chord data and the organizational structure information of the image data to further improve the image quality. Although the above methods can suppress image noise well, they still face the following defects: First, due to the privacy protection of low-dose original measurement data, the existing methods are all based on the simulation and correction of the chord diagram data after logarithmic transformation, without fully considering the statistical characteristics of the low-dose original measurement data, which greatly limits its imaging performance in real clinical scenarios; Second, the existing model does not fully consider the data correction and reconstruction process of the commercial CT reconstruction pipeline, the model interpretability is poor, and there is a lack of a robust CT full-chain intelligent reconstruction model.

[0004] Therefore, in view of the shortcomings of existing technologies, it is necessary to provide a CT full-chain intelligent reconstruction method based on deep learning to solve the shortcomings of existing technologies. Summary of the invention

[0005] The purpose of the present invention is to avoid the shortcomings of the prior art and provide a CT full-chain intelligent reconstruction method based on deep learning. The CT full-chain intelligent reconstruction method based on deep learning can reduce the scanning dose while ensuring image quality.

[0006] The above-mentioned purpose of the present invention is achieved by the following technical measures:

[0007] A CT full-chain intelligent reconstruction method based on deep learning is provided, comprising the following steps:

[0008] S1. Inputting the original measurement data into the trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data; the original measurement data is provided with two-dimensional projection data and three-dimensional projection data;

[0009] S2, processing the corrected measurement data obtained in S1 by negative logarithmic transformation to obtain chord diagram data;

[0010] S3, inputting the chord diagram data obtained in S2 into the trained angle-by-angle intelligent filtering network for filtering to obtain filtered chord diagram data;

[0011] S4, processing the filtered chord diagram data obtained in S3 by back-projection per angle to obtain back-projection tensor data;

[0012] S5, sorting the back-projection tensor data obtained in S4 by angle direction to obtain sorted back-projection tensor data;

[0013] S6. Input the sorted back-projection tensor data obtained in S5 into the trained back-projection tensor intelligent reconstruction network for reconstruction processing to obtain a final CT reconstructed image.

[0014] Preferably, the correction process is represented by equations (1) to (4);

[0015] σ=PNet(I,θ p )……Formula (1);

[0016] k=K(σ)……Formula (2);

[0017]

[0018] in, is the two-dimensional projection data, U is the number of detector channels, V is the number of projection angles, I 3d is the three-dimensional projection data, θ p is the network parameter of the pixel-by-pixel intelligent correction network PNet, σ is the standard deviation of the pixel-by-pixel Gaussian filter kernel, K is the window width mapping function of the Gaussian filter kernel, k is the window width of the pixel-by-pixel Gaussian filter kernel, is a pixel-by-pixel three-dimensional Gaussian filter kernel, is the measured data after correction, where x, y and z are spatial coordinates, u is the coordinate in the detector channel direction, and v is the coordinate in the angular direction.

[0019] Preferably, the negative logarithmic transformation process is represented by formula (5);

[0020]

[0021] Among them, s is the chord diagram data.

[0022] Preferably, the above filtering process is represented by equations (6) to (9);

[0023]

[0024] R v (w) = H v (w).*R(w)……Formula (8);

[0025]

[0026] in, is the one-dimensional Fourier transform of the chord data along the detector direction, θ s is the network parameter of the angle-by-angle smart filtering network SNet, H v (w) is the window function of the chord data filter kernel, P(w) is the ramp filter kernel, and w m is the bandwidth of the ramp filter kernel, R v (w) is the angle-by-angle filter kernel, is the one-dimensional inverse Fourier transform of the chord data along the detector direction, is the filtered chord data, and

[0027] Preferably, the above-mentioned angle-by-angle back-projection process is expressed by formula (10);

[0028]

[0029] in, is the back-projected tensor data, and I is the width of the image, J is the height of the image, INT is the linear interpolation operation, θ v is the rotation angle corresponding to the v-th angle projection, is the back-projection weighting factor.

[0030] Preferably, the above-mentioned angular direction sorting process is: sorting the back-projection tensor data in the angular direction according to the back-projection tensor values ​​from small to large, as expressed by formula (11);

[0031]

[0032] in, is the sorted back-projected tensor data, is the sorting operation, : is all the back-projected tensors in the angle direction.

[0033] Preferably, the above reconstruction process is expressed by equation (12);

[0034]

[0035] Where μ is the final CT reconstructed image, and θ r The network parameters of the back-projected tensor intelligent reconstruction network RNet.

[0036] Preferably, the above-mentioned pixel-by-pixel intelligent correction network, the angle-by-angle intelligent filtering network and the back-projection tensor intelligent reconstruction network are trained end-to-end through a pre-established training database to obtain a trained pixel-by-pixel intelligent correction network, a trained angle-by-angle intelligent filtering network and a trained back-projection tensor intelligent reconstruction network.

[0037] Preferably, the training database is provided with a plurality of pairs of paired normal dose measurement data, low dose measurement data, and normal dose image data reconstructed according to the normal dose measurement data, wherein one low dose measurement data and one normal dose measurement data constitute one paired data, and the dose of the low dose data is lower than that of the normal dose data, wherein each low dose measurement data includes two-dimensional projection data and three-dimensional projection data, and the doses of all low dose measurement data are not all the same.

[0038] Preferably, the above-mentioned end-to-end training is carried out through the following steps: first, the pixel-by-pixel intelligent correction network is pre-trained by the first training method to obtain a pre-trained pixel-by-pixel intelligent correction network; then, the angle-by-angle intelligent filtering network is pre-trained according to the pre-trained pixel-by-pixel intelligent correction network and the second training method to obtain a pre-trained angle-by-angle intelligent filtering network; then, the back-projection tensor intelligent reconstruction network is pre-trained according to the pre-trained pixel-by-pixel intelligent correction network, the pre-trained angle-by-angle intelligent filtering network and the third training method to obtain a pre-trained back-projection tensor intelligent reconstruction network; finally, the pre-trained pixel-by-pixel intelligent correction network, the pre-trained angle-by-angle intelligent filtering network and the pre-trained back-projection tensor intelligent reconstruction network are re-trained by the fourth training method to obtain a trained pixel-by-pixel intelligent correction network, a trained angle-by-angle intelligent filtering network and a trained back-projection tensor intelligent reconstruction network.

[0039] Preferably, the first training method is performed by the following steps:

[0040] A1. Inputting the two-dimensional projection data and the three-dimensional projection data of the same low-dose measurement data in the training database into a pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data;

[0041] A2, updating the parameters of the pixel-by-pixel intelligent correction network of A1 by using a first loss function, wherein the first loss function is represented by formula (13);

[0042]

[0043] In formula (13), Loss p is the loss value of the first loss function, N is the number of training samples, is the corrected measurement data obtained by A1, It is the normal dose measurement data in the same paired data;

[0044] A3, determine whether the training epoch has reached ω1 times. If yes, go to A4; if no, return to A1, and ω1 is an integer greater than 2;

[0045] A4. Loss p The minimum value of is used as the parameter of the pre-trained pixel-by-pixel intelligent correction network, and the training is ended.

[0046] Preferably, the second training method is performed by the following steps:

[0047] B1, inputting the two-dimensional projection data and the three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data;

[0048] B2, the corrected measurement data obtained in B1 is processed by negative logarithmic transformation to obtain chord diagram data;

[0049] B3, inputting the chord diagram data obtained in B2 into the angle-by-angle intelligent filtering network for filtering, thereby obtaining filtered chord diagram data;

[0050] B4, updating the parameters of the angle-by-angle intelligent filtering network of B3 by using a second loss function, wherein the second loss function is represented by formula (14);

[0051]

[0052] In formula (14), Loss s is the loss value of the second loss function, Get the chord diagram data for B3, It is the chord data after normal dose filtering;

[0053] B5. Determine whether the training epoch has reached ω2 times. If yes, go to B6; if no, return to B1, and ω2 is an integer greater than 2;

[0054] B6. All loss values ​​Loss sThe minimum value is used as the parameter of the pre-trained angle-by-angle intelligent filtering network, and the training ends.

[0055] Preferably, the third training method is performed by the following steps:

[0056] C1. Inputting the two-dimensional projection data and the three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data;

[0057] C2, the corrected measurement data obtained in C1 is processed by negative logarithmic transformation to obtain chord diagram data;

[0058] C3, inputting the chord diagram data obtained in C2 into the angle-by-angle intelligent filtering network for filtering to obtain filtered chord diagram data;

[0059] C4, processing the filtered chord diagram data obtained in C3 by back-projection per angle to obtain back-projection tensor data;

[0060] C5, sorting the back-projection tensor data obtained in C4 by angle direction to obtain sorted back-projection tensor data;

[0061] C6, inputting the sorted back-projection tensor data obtained in C5 into the back-projection tensor intelligent reconstruction network for reconstruction to obtain a CT reconstructed image;

[0062] C7, updating the parameters of the back-projection tensor intelligent reconstruction network of C6 through a third loss function, wherein the third loss function is represented by formula (15);

[0063]

[0064] In formula (15), Loss r is the loss value of the third loss function, The CT reconstructed image obtained for C6, It is the normal dose image data corresponding to the normal dose measurement data in the same paired data;

[0065] C8, determine whether the training epoch has reached ω3 times. If yes, enter C9; if not, return to C1, and ω3 is an integer greater than 2;

[0066] C9, all loss values ​​Loss r The minimum value is used as the parameter of the pre-trained back-projection tensor intelligent reconstruction network, and the training ends.

[0067] Preferably, the fourth training method is performed by the following steps:

[0068] D1, inputting the two-dimensional projection data and the three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data;

[0069] D2, the corrected measurement data obtained in D1 is processed by negative logarithmic transformation to obtain chord diagram data;

[0070] D3, inputting the chord diagram data obtained in D2 into the pre-trained angle-by-angle intelligent filtering network for filtering to obtain filtered chord diagram data;

[0071] D4, processing the filtered chord diagram data obtained in D3 by back-projection per angle to obtain back-projection tensor data;

[0072] D5, sorting the back-projection tensor data obtained in D4 by angle direction to obtain sorted back-projection tensor data;

[0073] D6, inputting the sorted back-projection tensor data obtained in D5 into the pre-trained back-projection tensor intelligent reconstruction network for reconstruction to obtain a CT reconstructed image;

[0074] D7, respectively updating the parameters of the pre-trained back-projection tensor intelligent reconstruction network of D1, the parameters of the pre-trained angle-by-angle intelligent filtering network of D3, and the parameters of the pre-trained back-projection tensor intelligent reconstruction network of D6 through a fourth loss function, wherein the fourth loss function is represented by formula (16);

[0075]

[0076] In formula (16), Loss is the loss value of the fourth loss function, is the corrected measurement data obtained for D1, is the normal dose measurement data in the same paired data, Get the chord diagram data for D3, The chord diagram data is the filtered normal dose measurement data in the same paired data; The CT reconstructed image obtained for D6, It is the normal dose image data corresponding to the normal dose measurement data in the same paired data;

[0077] D8, determine whether the training epoch has reached ω4 times. If yes, go to D9; if no, return to D1, and ω4 is an integer greater than 2;

[0078] D9. The minimum value of all loss values ​​Loss is used as the parameters of the pre-trained back-projection tensor intelligent reconstruction network, the parameters of the pre-trained angle-by-angle intelligent filtering network, and the parameters of the pre-trained back-projection tensor intelligent reconstruction network, and the training is ended.

[0079] A CT full-chain intelligent reconstruction method based on deep learning of the present invention comprises the following steps: S1, inputting the original measurement data into the trained pixel-by-pixel intelligent correction network for correction processing to obtain the corrected measurement data; the original measurement data is provided with two-dimensional projection data and three-dimensional projection data; S2, the corrected measurement data obtained by S1 is processed by negative logarithmic transformation to obtain chord data; S3, the chord data obtained by S2 is input into the trained angle-by-angle intelligent filtering network for filtering processing to obtain filtered chord data; S4, the filtered chord data obtained by S3 is processed by angle-by-angle back-projection to obtain back-projection tensor data; S5, the back-projection tensor data obtained by S4 is processed by angular direction sorting to obtain sorted back-projection tensor data; S6, the sorted back-projection tensor data obtained by S5 is input into the trained back-projection tensor intelligent reconstruction network for reconstruction processing to obtain the final CT reconstructed image. The pixel-by-pixel intelligent correction network provided by the present invention can perform pixel-by-pixel adaptive Gaussian filtering on low-dose original measurement data by learning the variance of each pixel point. Moreover, the angle-by-angle intelligent filtering network can perform angle-by-angle adaptive filtering on the chord diagram data by learning the filter kernel of each projection angle. Finally, the back-projection tensor intelligent reconstruction network can reduce the scanning dose while ensuring the image quality by learning the mapping from the back-projection tensor to the target image. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The present invention is further described with reference to the accompanying drawings, but the contents in the accompanying drawings do not constitute any limitation to the present invention.

[0081] Figure 1 This is a flow chart of a CT full-chain intelligent reconstruction method based on deep learning.

[0082] Figure 2 It is low-dose image data.

[0083] Figure 3 This is the final CT reconstructed image after reconstruction according to the present invention. DETAILED DESCRIPTION

[0084] The technical solution of the present invention is further described in conjunction with the following embodiments.

[0085] Example 1

[0086] A CT full-chain intelligent reconstruction method based on deep learning, such as Figure 1 As shown, the steps include:

[0087] S1. Inputting the original measurement data into the trained pixel-by-pixel intelligent correction network for correction processing to obtain the corrected measurement data; the original measurement data is provided with two-dimensional projection data and three-dimensional projection data;

[0088] S2, processing the corrected measurement data obtained in S1 by negative logarithmic transformation to obtain chord diagram data;

[0089] S3, inputting the chord diagram data obtained in S2 into the trained angle-by-angle intelligent filtering network for filtering to obtain filtered chord diagram data;

[0090] S4, processing the filtered chord diagram data obtained in S3 by back-projection per angle to obtain back-projection tensor data;

[0091] S5, sorting the back-projection tensor data obtained in S4 by angle direction to obtain sorted back-projection tensor data;

[0092] S6. Input the sorted back-projection tensor data obtained in S5 into the trained back-projection tensor intelligent reconstruction network for reconstruction processing to obtain a final CT reconstructed image.

[0093] It should be noted that the original measurement data of the present invention is lower than the normal dose measurement data.

[0094] The correction process of the present invention is represented by equations (1) to (4);

[0095] σ=PNet(I,θ p )……Formula (1);

[0096] k=K(σ)……Formula (2);

[0097]

[0098] in, is the two-dimensional projection data, U is the number of detector channels, V is the number of projection angles, I 3d is the three-dimensional projection data, θ p is the network parameter of the pixel-by-pixel intelligent correction network PNet, σ is the standard deviation of the pixel-by-pixel Gaussian filter kernel, K is the window width mapping function of the Gaussian filter kernel, k is the window width of the pixel-by-pixel Gaussian filter kernel, is a pixel-by-pixel three-dimensional Gaussian filter kernel, is the measured data after correction, where x, y and z are spatial coordinates, u is the coordinate in the detector channel direction, and v is the coordinate in the angular direction.

[0099] The negative logarithmic transformation process of the present invention is represented by equation (5);

[0100]

[0101] Among them, s is the chord diagram data.

[0102] The filtering process of the present invention is represented by equations (6) to (9);

[0103]

[0104] R v (w) = H v (w).*R(w)……Formula (8);

[0105]

[0106] in, is the one-dimensional Fourier transform of the chord data along the detector direction, θ s is the network parameter of the angle-by-angle smart filtering network SNet, H v (w) is the window function of the chord data filter kernel, R(w) is the ramp filter kernel, and w m is the bandwidth of the ramp filter kernel, R v (w) is the angle-by-angle filter kernel, is the one-dimensional inverse Fourier transform of the chord data along the detector direction, is the filtered chord data, and

[0107] The angle-by-angle back-projection process of the present invention is expressed by equation (10):

[0108]

[0109] in, is the back-projected tensor data, and I is the width of the image, J is the height of the image, INT is the linear interpolation operation, θ v is the rotation angle corresponding to the v-th angle projection, is the back-projection weighting factor.

[0110] The angular direction sorting process of the present invention is: sorting the back-projection tensor data in the angular direction from small to large according to the back-projection tensor value, as expressed by formula (11);

[0111]

[0112] in, is the sorted back-projected tensor data, is the sorting operation, : is all the back-projected tensors in the angle direction.

[0113] The reconstruction process of the present invention is expressed by equation (12):

[0114]

[0115] Where μ is the final CT reconstructed image, and θ r The network parameters of the back-projected tensor intelligent reconstruction network RNet.

[0116] The pixel-by-pixel intelligent correction network, angle-by-angle intelligent filtering network and back-projection tensor intelligent reconstruction network of the present invention are trained end-to-end through a pre-established training database to obtain a trained pixel-by-pixel intelligent correction network, a trained angle-by-angle intelligent filtering network and a trained back-projection tensor intelligent reconstruction network. The training database is provided with multiple pairs of paired normal dose measurement data, low dose measurement data, and normal dose image data reconstructed according to the normal dose measurement data, one low dose measurement data and one normal dose measurement data form a paired data, and the dose of the low dose data is lower than the normal dose data, wherein each low dose measurement data includes two-dimensional projection data and three-dimensional projection data, and the doses of all low dose measurement data are not all the same.

[0117] It should be noted that the low-dose measurement data in the training database is obtained by simulation based on the normal-dose measurement data.

[0118] End-to-end training is carried out through the following steps: first, the pixel-by-pixel intelligent correction network is pre-trained by the first training method to obtain a pre-trained pixel-by-pixel intelligent correction network; then, the angle-by-angle intelligent filtering network is pre-trained according to the pre-trained pixel-by-pixel intelligent correction network and the second training method to obtain a pre-trained angle-by-angle intelligent filtering network; then, the back-projection tensor intelligent reconstruction network is pre-trained according to the pre-trained pixel-by-pixel intelligent correction network, the pre-trained angle-by-angle intelligent filtering network and the third training method to obtain a pre-trained back-projection tensor intelligent reconstruction network; finally, the pre-trained pixel-by-pixel intelligent correction network, the pre-trained angle-by-angle intelligent filtering network and the pre-trained back-projection tensor intelligent reconstruction network are re-trained by the fourth training method to obtain a trained pixel-by-pixel intelligent correction network, a trained angle-by-angle intelligent filtering network and a trained back-projection tensor intelligent reconstruction network.

[0119] The first training method is performed by the following steps:

[0120] A1. Input the two-dimensional projection data and three-dimensional projection data of the same low-dose measurement data in the training database into the pixel-by-pixel intelligent correction network for correction processing to obtain the corrected measurement data;

[0121] A2, update the parameters of the pixel-by-pixel intelligent correction network of A1 through the first loss function, and the first loss function is expressed by formula (13);

[0122]

[0123] In formula (13), Loss p is the loss value of the first loss function, N is the number of training samples, is the corrected measurement data obtained by A1, It is the normal dose measurement data in the same paired data;

[0124] A3, determine whether the training epoch has reached ω1 times. If yes, go to A4; if no, return to A1, and ω1 is an integer greater than 2;

[0125] A4. Loss p The minimum value of is used as the parameter of the pre-trained pixel-by-pixel intelligent correction network, and the training ends;

[0126] The second training method is carried out through the following steps:

[0127] B1. Input the two-dimensional projection data and three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain the corrected measurement data;

[0128] B2, the corrected measurement data obtained in B1 is processed by negative logarithmic transformation to obtain chord diagram data;

[0129] B3, inputting the chord diagram data obtained in B2 into the angle-by-angle intelligent filtering network for filtering, thereby obtaining filtered chord diagram data;

[0130] B4, updating the parameters of the angle-by-angle intelligent filtering network of B3 through the second loss function, the second loss function is expressed by formula (14);

[0131]

[0132] In formula (14), Loss s is the loss value of the second loss function, Get the chord diagram data for B3, The chord data after normal dose filtering;

[0133] B5. Determine whether the training epoch has reached ω2 times. If yes, go to B6; if no, return to B1, and ω2 is an integer greater than 2;

[0134] B6. All loss values ​​Loss s The minimum value is used as the parameter of the pre-trained angle-by-angle intelligent filtering network, and the training ends;

[0135] The third training method is performed by the following steps:

[0136] C1. Input the two-dimensional projection data and three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain the corrected measurement data;

[0137] C2, the corrected measurement data obtained in C1 is processed by negative logarithmic transformation to obtain chord diagram data;

[0138] C3, inputting the chord diagram data obtained in C2 into the angle-by-angle intelligent filtering network for filtering, and obtaining filtered chord diagram data;

[0139] C4, processing the filtered chord diagram data obtained in C3 by back-projection per angle to obtain back-projection tensor data;

[0140] C5, sorting the back-projection tensor data obtained in C4 by angle direction to obtain sorted back-projection tensor data;

[0141] C6, inputting the sorted back-projection tensor data obtained in C5 into the back-projection tensor intelligent reconstruction network for reconstruction to obtain a CT reconstructed image;

[0142] C7, update the parameters of the back-projection tensor intelligent reconstruction network of C6 through the third loss function, and the third loss function is expressed by formula (15);

[0143]

[0144] In formula (15), Loss r is the loss value of the third loss function, The CT reconstructed image obtained for C6, It is the normal dose image data corresponding to the normal dose measurement data in the same paired data;

[0145] C8, determine whether the training epoch has reached ω3 times. If yes, enter C9; if not, return to C1, and ω3 is an integer greater than 2;

[0146] C9, all loss values ​​Loss r The minimum value is used as the parameter of the pre-trained back-projection tensor intelligent reconstruction network, and the training ends;

[0147] The fourth training method is performed by the following steps:

[0148] D1. Input the two-dimensional projection data and three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain the corrected measurement data;

[0149] D2, the corrected measurement data obtained in D1 is processed by negative logarithmic transformation to obtain chord diagram data;

[0150] D3, inputting the chord diagram data obtained in D2 into the pre-trained angle-by-angle intelligent filtering network for filtering, and obtaining filtered chord diagram data;

[0151] D4, processing the filtered chord diagram data obtained in D3 by back-projection per angle to obtain back-projection tensor data;

[0152] D 5. Sorting the back-projected tensor data obtained in D4 by angle direction to obtain sorted back-projected tensor data;

[0153] D6, inputting the sorted back-projection tensor data obtained in D5 into the pre-trained back-projection tensor intelligent reconstruction network for reconstruction to obtain a CT reconstructed image;

[0154] D7, respectively update the parameters of the pre-trained back-projection tensor intelligent reconstruction network of D1, the parameters of the pre-trained angle-by-angle intelligent filtering network of D3, and the parameters of the pre-trained back-projection tensor intelligent reconstruction network of D6 through the fourth loss function, and the fourth loss function is represented by formula (16);

[0155]

[0156] In formula (16), Loss is the loss value of the fourth loss function, is the corrected measurement data obtained for D1, is the normal dose measurement data in the same paired data, Get the chord diagram data for D3, The chord diagram data is the filtered normal dose measurement data in the same paired data; The CT reconstructed image obtained by D6, It is the normal dose image data corresponding to the normal dose measurement data in the same paired data;

[0157] D8, determine whether the training epoch has reached ω4 times. If yes, go to D9; if no, return to D1, and ω4 is an integer greater than 2;

[0158] D9. The minimum value of all loss values ​​Loss is used as the parameters of the pre-trained back-projection tensor intelligent reconstruction network, the parameters of the pre-trained angle-by-angle intelligent filtering network, and the parameters of the pre-trained back-projection tensor intelligent reconstruction network, and the training is ended.

[0159] The specific values ​​of ω1, ω2, ω3 and ω4 in this embodiment are determined according to factors such as the size of the computer computing power network and the amount of data. In this embodiment, ω1, ω2, ω3 and ω4 are specifically 100, and the learning rate is 1×10 -4 , the network optimizer is Adam.

[0160] It should be noted that one epoch is to traverse all the low-dose measurement data in the training database once. In the present invention, the pixel-by-pixel intelligent correction network, the angle-by-angle intelligent filtering network and the back-projection tensor intelligent reconstruction network are all conventional networks, such as UNet, residual network, etc. Before reconstruction or training, the present invention also needs to obtain the imaging system parameters of the CT device, including the number of detector channels nb, the detector size ds, the source-to-detector distance dsd, the source-to-rotation center distance dso, the number of projection angles na, the number of image pixels nx, and the image pixel size dx.

[0161] This CT full-chain intelligent reconstruction method based on deep learning first uses a pixel-by-pixel intelligent correction network to learn the variance of each pixel point, and can perform pixel-by-pixel adaptive Gaussian filtering on low-dose raw measurement data. In addition, the angle-by-angle intelligent filtering network can perform angle-by-angle adaptive filtering on chord diagram data by learning the filter kernel of each projection angle. Finally, the back-projection tensor intelligent reconstruction network can reduce the scanning dose while ensuring image quality by learning the mapping of the back-projection tensor to the target image.

[0162] Example 2

[0163] A CT full-chain intelligent reconstruction method based on deep learning, the other features are the same as those of Example 1. In this embodiment, before reconstruction and training, the imaging system parameters of the CT equipment are first obtained, wherein the number of detector channels nb=736, the detector size ds=1.2858 mm, the source-to-detector distance dsd=1085.6 mm, the source-to-rotation center distance dso=595 mm, the number of projection angles na=1152, the number of image pixels nx=512, and the image pixel size dx=0.6934 mm.

[0164] The normal projection data of this embodiment includes normal dose projection data and image data of the head, chest and abdomen. The projection size is 736×1152, and the image size is 512×512.

[0165] The low-dose measurement data in the training database of this embodiment includes Low-dose measurement data at four different dose levels.

[0166] In the back-projection tensor intelligent reconstruction network RNet The dimension size is 512×512×1152. The number of input channels in the pixel-by-pixel intelligent correction network PNet and the angle-by-angle intelligent filtering network SNet is 1, and the number of input channels in the back-projection tensor intelligent reconstruction network RNet is 1152. The window width of the pixel-by-pixel Gaussian filter kernel is k∈{3,5,7,9}.

[0167] Figure 2The low-dose image data is reconstructed from low-dose measurement data, and its dose is 1 / 16 of the normal dose. Figure 3 Is Figure 2 The low-dose measurement data is used to obtain the final CT reconstructed image through the CT full-chain intelligent reconstruction method of this embodiment, such as Figure 3 . And through Figure 2 and Figure 3 visible, Figure 2 The low-dose images contain severe noise and streak artifacts. Figure 3 The result of the present invention has basically removed the noise artifacts and restored the tissue structure in the image. Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the protection scope of the present invention. Although the present invention is described in detail with reference to the preferred embodiments, ordinary technicians in this field should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A CT full-chain intelligent reconstruction method based on deep learning, characterized in that: The steps include: S1. Inputting the original measurement data into the trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data; the original measurement data is provided with two-dimensional projection data and three-dimensional projection data; S2, processing the corrected measurement data obtained in S1 by negative logarithmic transformation to obtain chord diagram data; S3, inputting the chord diagram data obtained in S2 into the trained angle-by-angle intelligent filtering network for filtering to obtain filtered chord diagram data; S4, processing the filtered chord diagram data obtained in S3 by back-projection per angle to obtain back-projection tensor data; S5, sorting the back-projection tensor data obtained in S4 by angle direction to obtain sorted back-projection tensor data; S6. Input the sorted back-projection tensor data obtained in S5 into the trained back-projection tensor intelligent reconstruction network for reconstruction processing to obtain a final CT reconstructed image.

2. The CT full-chain intelligent reconstruction method based on deep learning according to claim 1, characterized in that: The correction process is represented by equations (1) to (4); σ = PNet(I, θ p )... Equation (1); k=K(σ)……Formula (2); in, is the two-dimensional projection data, U is the number of detector channels, V is the number of projection angles, I 3d is the three-dimensional projection data, θ p is the network parameter of the pixel-by-pixel intelligent correction network PNet, σ is the standard deviation of the pixel-by-pixel Gaussian filter kernel, K is the window width mapping function of the Gaussian filter kernel, k is the window width of the pixel-by-pixel Gaussian filter kernel, is a pixel-by-pixel three-dimensional Gaussian filter kernel, is the measured data after correction, where x, y and z are spatial coordinates, u is the coordinate in the detector channel direction, and v is the coordinate in the angular direction.

3. The CT full-chain intelligent reconstruction method based on deep learning according to claim 2 is characterized in that: The negative logarithmic transformation process is expressed by equation (5); Among them, s is the chord diagram data.

4. The CT full-chain intelligent reconstruction method based on deep learning according to claim 3 is characterized in that: The filtering process is represented by equations (6) to (9); in, is the one-dimensional Fourier transform of the chord data along the detector direction, θ s is the network parameter of the angle-by-angle smart filtering network SNet, H v (w) is the window function of the chord data filter kernel, R(w) is the ramp filter kernel, and w m is the bandwidth of the ramp filter kernel, R v (w) is the angle-by-angle filter kernel, is the one-dimensional inverse Fourier transform of the chord data along the detector direction, is the filtered chord data, and 5. The CT full-chain intelligent reconstruction method based on deep learning according to claim 4 is characterized in that: The angle-by-angle back-projection process is expressed by equation (10); in, is the back-projected tensor data, and I is the width of the image, J is the height of the image, V is the number of projection angles, INT is the linear interpolation operation, θ v is the rotation angle corresponding to the v-th angle projection, is the back-projection weighting factor.

6. The CT full-chain intelligent reconstruction method based on deep learning according to claim 5 is characterized in that: The angular direction sorting process is: sorting the back-projection tensor data in the angular direction from small to large according to the back-projection tensor values, as expressed by formula (11); in, is the sorted back-projected tensor data, is the sorting operation, : is all the back-projected tensors in the angle direction.

7. The CT full-chain intelligent reconstruction method based on deep learning according to claim 6, characterized in that: The reconstruction process is expressed by equation (12): Where μ is the final CT reconstructed image, and θ r The network parameters of the back-projected tensor intelligent reconstruction network RNet.

8. The CT full-chain intelligent reconstruction method based on deep learning according to any one of claims 1 to 7, characterized in that: The pixel-by-pixel intelligent correction network, the angle-by-angle intelligent filtering network and the back-projection tensor intelligent reconstruction network are trained end-to-end using a pre-established training database to obtain a trained pixel-by-pixel intelligent correction network, a trained angle-by-angle intelligent filtering network and a trained back-projection tensor intelligent reconstruction network; The training database is provided with a plurality of pairs of paired normal dose measurement data, low dose measurement data, and normal dose image data reconstructed according to the normal dose measurement data, wherein one low dose measurement data and one normal dose measurement data constitute one paired data, and the dose of the low dose data is lower than that of the normal dose data, wherein each low dose measurement data includes two-dimensional projection data and three-dimensional projection data, and the doses of all low dose measurement data are not all the same.

9. The CT full-chain intelligent reconstruction method based on deep learning according to claim 8, characterized in that: The end-to-end training is carried out through the following steps: first, the pixel-by-pixel intelligent correction network is pre-trained by the first training method to obtain a pre-trained pixel-by-pixel intelligent correction network; then, the angle-by-angle intelligent filtering network is pre-trained according to the pre-trained pixel-by-pixel intelligent correction network and the second training method to obtain a pre-trained angle-by-angle intelligent filtering network; then, the back-projection tensor intelligent reconstruction network is pre-trained according to the pre-trained pixel-by-pixel intelligent correction network, the pre-trained angle-by-angle intelligent filtering network and the third training method to obtain a pre-trained back-projection tensor intelligent reconstruction network; finally, the pre-trained pixel-by-pixel intelligent correction network, the pre-trained angle-by-angle intelligent filtering network and the pre-trained back-projection tensor intelligent reconstruction network are re-trained by the fourth training method to obtain a trained pixel-by-pixel intelligent correction network, a trained angle-by-angle intelligent filtering network and a trained back-projection tensor intelligent reconstruction network.

10. The CT full-chain intelligent reconstruction method based on deep learning according to claim 9, characterized in that: The first training method is performed by the following steps: A1. Inputting the two-dimensional projection data and the three-dimensional projection data of the same low-dose measurement data in the training database into a pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data; A2, updating the parameters of the pixel-by-pixel intelligent correction network of A1 by using a first loss function, wherein the first loss function is represented by formula (13); In formula (13), Loss p is the loss value of the first loss function, N is the number of training samples, is the corrected measurement data obtained by A1, It is the normal dose measurement data in the same paired data; A3, determine whether the training epoch has reached ω1 times. If yes, go to A4; if no, return to A1, and ω1 is an integer greater than 2; A4. Loss p The minimum value of is used as the parameter of the pre-trained pixel-by-pixel intelligent correction network, and the training ends; The second training method is performed by the following steps: B1, inputting the two-dimensional projection data and the three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data; B2, the corrected measurement data obtained in B1 is processed by negative logarithmic transformation to obtain chord diagram data; B3, inputting the chord diagram data obtained in B2 into the angle-by-angle intelligent filtering network for filtering, thereby obtaining filtered chord diagram data; B4, updating the parameters of the angle-by-angle intelligent filtering network of B3 by using a second loss function, wherein the second loss function is represented by formula (14); In formula (14), Loss s is the loss value of the second loss function, Get the chord diagram data for B3, The chord data after normal dose filtering; B5. Determine whether the training epoch has reached ω2 times. If yes, go to B6; if no, return to B1, and ω2 is an integer greater than 2; B6. All loss values ​​Loss s The minimum value is used as the parameter of the pre-trained angle-by-angle intelligent filtering network, and the training ends; The third training method is performed by the following steps: C1. Inputting the two-dimensional projection data and the three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data; C2, the corrected measurement data obtained in C1 is processed by negative logarithmic transformation to obtain chord diagram data; C3, inputting the chord diagram data obtained in C2 into the angle-by-angle intelligent filtering network for filtering to obtain filtered chord diagram data; C4, processing the filtered chord diagram data obtained in C3 by back-projection per angle to obtain back-projection tensor data; C5, sorting the back-projection tensor data obtained in C4 by angle direction to obtain sorted back-projection tensor data; C6, inputting the sorted back-projection tensor data obtained in C5 into the back-projection tensor intelligent reconstruction network for reconstruction to obtain a CT reconstructed image; C7, updating the parameters of the back-projection tensor intelligent reconstruction network of C6 through a third loss function, wherein the third loss function is represented by formula (15); In formula (15), Loss r is the loss value of the third loss function, The CT reconstructed image obtained for C6, It is the normal dose image data corresponding to the normal dose measurement data in the same paired data; C8, determine whether the training epoch has reached ω3 times. If yes, enter C9; if not, return to C1, and ω3 is an integer greater than 2; C9, all loss values ​​Loss r The minimum value is used as the parameter of the pre-trained back-projection tensor intelligent reconstruction network, and the training ends; The fourth training method is performed by the following steps: D1, inputting the two-dimensional projection data and the three-dimensional projection data of the low-dose measurement data in the training database into the pre-trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data; D2, the corrected measurement data obtained in D1 is processed by negative logarithmic transformation to obtain chord diagram data; D3, inputting the chord diagram data obtained in D2 into the pre-trained angle-by-angle intelligent filtering network for filtering to obtain filtered chord diagram data; D4, processing the filtered chord diagram data obtained in D3 by back-projection per angle to obtain back-projection tensor data; D 5. Sorting the back-projected tensor data obtained in D4 by angle direction to obtain sorted back-projected tensor data; D6, inputting the sorted back-projection tensor data obtained in D5 into the pre-trained back-projection tensor intelligent reconstruction network for reconstruction to obtain a CT reconstructed image; D7, respectively updating the parameters of the pre-trained back-projection tensor intelligent reconstruction network of D1, the parameters of the pre-trained angle-by-angle intelligent filtering network of D3, and the parameters of the pre-trained back-projection tensor intelligent reconstruction network of D6 through a fourth loss function, wherein the fourth loss function is represented by formula (16); In formula (16), Loss is the loss value of the fourth loss function, is the corrected measurement data obtained for D1, is the normal dose measurement data in the same paired data, Get the chord diagram data for D3, The chord diagram data is the filtered normal dose measurement data in the same paired data; The CT reconstructed image obtained by D6, It is the normal dose image data corresponding to the normal dose measurement data in the same paired data; D8, determine whether the training epoch has reached ω4 times. If yes, go to D9; if no, return to D1, and ω4 is an integer greater than 2; D9. The minimum value of all loss values ​​Loss is used as the parameters of the pre-trained back-projection tensor intelligent reconstruction network, the parameters of the pre-trained angle-by-angle intelligent filtering network, and the parameters of the pre-trained back-projection tensor intelligent reconstruction network, and the training is ended.

Images