A Deep Learning-Based Intelligent Reconstruction Method for the CT Full Chain

Abstract

This deep learning-based intelligent CT reconstruction method processes raw measurement data sequentially through a pixel-by-pixel intelligent correction network, a trained angle-by-angle intelligent filtering network, and a trained backprojection tensor intelligent reconstruction network to obtain the final CT reconstructed image. First, the pixel-by-pixel intelligent correction network learns the variance of each pixel to perform pixel-by-pixel adaptive Gaussian filtering on low-dose raw measurement data. Second, the angle-by-angle intelligent filtering network learns the filtering kernel for each projection angle to perform angle-by-angle adaptive filtering on chordogram data. Finally, the backprojection tensor intelligent reconstruction network learns the mapping from the backprojection tensor to the target image, ensuring image quality while reducing scan dose.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and computed tomography (CT) technology, and in particular to a deep learning-based intelligent reconstruction method for the entire CT chain. Background Technology

[0002] Computed tomography (CT) technology is widely used in important fields such as medical imaging, security inspection, and industrial inspection due to its non-invasive, high-resolution three-dimensional imaging capabilities. However, the high radiation dose from medical CT imaging raises concerns about the potential cancer risks associated with X-rays. Therefore, minimizing radiation dose while maintaining image quality has become a common goal. Reducing the scan dose inevitably introduces significant noise artifacts into the reconstructed images, thus affecting clinical diagnostic performance. Therefore, improving the quality of low-dose CT images has always been one of the key research technologies in the field of medical imaging.

[0003] In recent years, deep learning technology has been widely applied in medical image reconstruction and has achieved unprecedented success. Deep learning-based low-dose CT data reconstruction algorithms have been extensively developed, enhancing image quality by learning key modules in the CT reconstruction pipeline in a data-driven manner. These include chordogram domain networks, image domain networks, and dual-domain networks. Specifically, chordogram domain networks directly recover chordogram data affected by noise and then reconstruct the CT image through filtered backprojection. Image domain networks, belonging to image post-processing techniques, suppress noise by learning the nonlinear mapping from low-dose images to target images. Dual-domain networks construct differentiable filtered backprojection layers between the chordogram and image domains, achieving joint optimization of the chordogram and image sub-networks. This fully utilizes the integral information of the chordogram data and the organizational structure information of the image data to further improve image quality. Although the above methods can suppress image noise well, they still face the following shortcomings: First, due to the privacy protection of low-dose raw measurement data, existing methods are based on the logarithmically transformed chord graph data for simulation and correction, without fully considering the statistical characteristics of low-dose raw measurement data, which greatly limits their imaging performance in real clinical scenarios; Second, existing models do not fully consider the data correction and reconstruction process of commercial CT reconstruction pipelines, resulting in poor model interpretability and a lack of a robust intelligent reconstruction model for the entire CT chain.

[0004] Therefore, it is essential to provide a deep learning-based intelligent reconstruction method for the entire CT process to address the shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based intelligent reconstruction method for the entire CT scan chain. This deep learning-based intelligent reconstruction method for the entire CT scan chain can maintain image quality while reducing scan dose.

[0006] The above-mentioned objectives of the present invention are achieved through the following technical measures:

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

[0008] S1. The original measurement data is input into the trained pixel-by-pixel intelligent correction network and processed to obtain the corrected measurement data; the original measurement data includes two-dimensional projection data and three-dimensional projection data.

[0009] S2. The corrected measurement data obtained in S1 is processed by negative logarithmic transformation to obtain chord diagram data;

[0010] S3. Input the chord diagram data obtained in S2 into the trained angle-by-angle intelligent filtering network and process it to obtain the filtered chord diagram data.

[0011] S4. The filtered chord graph data obtained in S3 is processed by angle-by-angle back projection to obtain back projection tensor data.

[0012] S5. Sort the back projection tensor data obtained in S4 by angle and direction to obtain sorted back projection tensor data.

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

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

[0015] σ=PNet(I,θ p ...Equation (1);

[0016] k=K(σ)……Equation (2);

[0017]

[0018] in, This is two-dimensional projection data, where U is the number of detector channels, V is the number of projection angles, and I... 3d For three-dimensional projection data, θ p The network parameters for the pixel-wise intelligent correction network PNet are defined as follows: σ represents the standard deviation of the pixel-wise Gaussian filter kernel, K is the window width mapping function of the Gaussian filter kernel, and k is the window width of the pixel-wise Gaussian filter kernel. For pixel-by-pixel 3D Gaussian filter kernel, For the calibrated measurement data, 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 above negative logarithmic transformation process is represented by equation (5);

[0020]

[0021] Where s represents the chord graph data.

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

[0023]

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

[0025]

[0026] in, For the one-dimensional Fourier transform of the chord data along the detector direction, θ s For the network parameters of the angle-by-angle intelligent filtering network SNet, H v (w) is the window function of the chord graph data filtering kernel, P(w) is the ramp filtering kernel, w m R is the bandwidth of the ramp filter kernel. v (w) is the angle-by-angle filtering kernel. This represents the one-dimensional inverse Fourier transform of the chord graph data along the detector direction. The data is the filtered chordogram data, and

[0027] Preferably, the above angle-by-angle back-projection processing is represented by equation (10);

[0028]

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

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

[0031]

[0032] in, This is the sorted backprojection tensor data. For sorting operations, : represents all backprojection tensors in the angular direction.

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

[0034]

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

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

[0037] Preferably, the training database is configured with multiple pairs of paired normal dose measurement data, low dose measurement data, and normal dose image data reconstructed from the normal dose measurement data. A low dose measurement data and a normal dose measurement data form a paired data, and the doses of the low dose data are all lower than those of the normal dose data. 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 end-to-end training described above is performed through the following steps: First, the pixel-by-pixel intelligent correction network is pre-trained using a 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 using the pre-trained pixel-by-pixel intelligent correction network and a second training method to obtain a pre-trained angle-by-angle intelligent filtering network; next, the back-projection tensor intelligent reconstruction network is pre-trained using 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 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 retrained using a fourth training method to obtain the trained pixel-by-pixel intelligent correction network, the trained angle-by-angle intelligent filtering network, and the trained back-projection tensor intelligent reconstruction network.

[0039] Preferably, the first training method described above is performed through the following steps:

[0040] 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 and process them to obtain the corrected measurement data.

[0041] A2. Update the parameters of the pixel-by-pixel intelligent correction network of A1 using the first loss function, which is represented by equation (13);

[0042]

[0043] In equation (13), Loss p Here, N represents the loss value of the first loss function, and N is the number of training samples. The corrected measurement data obtained from A1 These are normal dose measurement data within the same paired dataset;

[0044] A3. Determine if the training epoch has reached ω1. If yes, proceed to A4; otherwise, return to A1, where ω1 is an integer greater than 2.

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

[0046] Preferably, the second training method described above is performed through the following steps:

[0047] 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 and obtain the corrected measurement data through correction processing;

[0048] B2. The corrected measurement data obtained in B1 is processed by negative logarithmic transformation to obtain chord graph data;

[0049] B3. Input the chord diagram data obtained in B2 into the angle-by-angle intelligent filtering network for filtering processing to obtain the filtered chord diagram data;

[0050] B4. Update the parameters of the angle-by-angle intelligent filtering network of B3 using the second loss function, which is represented by equation (14).

[0051]

[0052] In equation (14), Loss s The loss value of the second loss function. To obtain the chord diagram data for B3, This is the chord graph data after normal dose filtering;

[0053] B5. Determine if the training epoch has reached ω2 times. If yes, proceed to B6; otherwise, return to B1, and ω2 is an integer greater than 2.

[0054] B6. Loss all loss values 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 described above is performed through the following steps:

[0056] 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 and obtain the corrected measurement data through correction processing;

[0057] C2. The corrected measurement data obtained in C1 is processed by negative logarithmic transformation to obtain chord graph data;

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

[0059] C4. The filtered chord graph data obtained in C3 is processed by angle-by-angle back projection to obtain back projection tensor data;

[0060] C5. Sort the backprojection tensor data obtained in C4 by angle and direction to obtain sorted backprojection tensor data.

[0061] C6. Input the sorted backprojection tensor data obtained in C5 into the backprojection tensor intelligent reconstruction network for reconstruction to obtain the CT reconstructed image.

[0062] C7. Update the parameters of the back projection tensor intelligent reconstruction network of C6 through the third loss function, which is represented by equation (15);

[0063]

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

[0065] C8. Determine if the training epoch has reached ω3. If yes, proceed to C9; otherwise, return to C1, and ω3 is an integer greater than 2.

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

[0067] Preferably, the fourth training method described above is performed through the following steps:

[0068] 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 and obtain the corrected measurement data through correction processing;

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

[0070] D3. Input the chord graph data obtained in D2 into the pre-trained angle-by-angle intelligent filtering network for filtering processing to obtain the filtered chord graph data;

[0071] D4. The filtered chord graph data obtained in D3 is processed by angle-by-angle back projection to obtain back projection tensor data.

[0072] D5. Sort the back projection tensor data obtained in D4 by angle and direction to obtain sorted back projection tensor data.

[0073] D6. Input the sorted backprojection tensor data obtained in D5 into the pre-trained backprojection tensor intelligent reconstruction network for reconstruction to obtain the CT reconstruction image.

[0074] D7. 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 are updated by the fourth loss function, which is represented by equation (16).

[0075]

[0076] In equation (16), Loss is the loss value of the fourth loss function. The corrected measurement data obtained from D1 For normal dose measurement data in the same paired dataset, To obtain the chord diagram data for D3, The chord graph data is the filtered normal dose measurement data from the same paired dataset; The CT reconstructed image obtained from D6. Normal dose image data corresponding to normal dose measurement data in the same paired data;

[0077] D8. Determine if the training epoch has reached ω4. If yes, proceed to D9; otherwise, return to D1, and ω4 is an integer greater than 2.

[0078] D9. Use the minimum value among all loss values ​​as the parameters of the pre-trained backprojection tensor intelligent reconstruction network, the pre-trained angle-by-angle intelligent filtering network, and the pre-trained backprojection tensor intelligent reconstruction network, and end the training.

[0079] This invention discloses a deep learning-based intelligent reconstruction method for the entire CT chain, comprising the following steps: S1, inputting the original measurement data into a trained pixel-by-pixel intelligent correction network for correction processing to obtain corrected measurement data; the original measurement data includes two-dimensional projection data and three-dimensional projection data; S2, processing the corrected measurement data obtained in S1 through negative logarithmic transformation to obtain chord graph data; S3, inputting the chord graph data obtained in S2 into a trained angle-by-angle intelligent filtering network for filtering processing to obtain filtered chord graph data; S4, processing the filtered chord graph data obtained in S3 through angle-by-angle backprojection to obtain backprojection tensor data; S5, sorting the backprojection tensor data obtained in S4 through angle direction sorting to obtain sorted backprojection tensor data; S6, inputting the sorted backprojection tensor data obtained in S5 into a trained backprojection tensor intelligent reconstruction network for reconstruction processing to obtain the final CT reconstructed image. The pixel-by-pixel intelligent correction network provided by this invention can perform pixel-by-pixel adaptive Gaussian filtering on low-dose original measurement data by learning the variance of each pixel. Furthermore, the angle-by-angle intelligent filtering network can perform angle-by-angle adaptive filtering on chordogram data by learning the filtering kernel for each projection angle. Finally, the back-projection tensor intelligent reconstruction network can reduce scanning dose while maintaining image quality by learning the mapping from the back-projection tensor to the target image. Attached Figure Description

[0080] The invention will be further described with reference to the accompanying drawings, but the contents of the drawings do not constitute any limitation on the invention.

[0081] Figure 1 This is a flowchart of a deep learning-based intelligent reconstruction method for the entire CT process.

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

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

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

[0085] Example 1

[0086] A deep learning-based intelligent reconstruction method for the entire CT scan chain, such as... Figure 1 As shown, it includes the following steps:

[0087] S1. Input the original measurement data into the trained pixel-by-pixel intelligent correction network and process it to obtain the corrected measurement data; the original measurement data includes two-dimensional projection data and three-dimensional projection data.

[0088] S2. The corrected measurement data obtained in S1 is processed by negative logarithmic transformation to obtain chord diagram data;

[0089] S3. Input the chord diagram data obtained in S2 into the trained angle-by-angle intelligent filtering network and process it to obtain the filtered chord diagram data.

[0090] S4. The filtered chord graph data obtained in S3 is processed by angle-by-angle back projection to obtain back projection tensor data.

[0091] S5. Sort the back projection tensor data obtained in S4 by angle and direction to obtain sorted back projection tensor data.

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

[0093] It should be noted that the original measurement data of this invention are 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 ...Equation (1);

[0096] k=K(σ)……Equation (2);

[0097]

[0098] in, This is two-dimensional projection data, where U is the number of detector channels, V is the number of projection angles, and I... 3d For three-dimensional projection data, θ p The network parameters for the pixel-wise intelligent correction network PNet are defined as follows: σ represents the standard deviation of the pixel-wise Gaussian filter kernel, K is the window width mapping function of the Gaussian filter kernel, and k is the window width of the pixel-wise Gaussian filter kernel. For pixel-by-pixel 3D Gaussian filter kernel, For the calibrated measurement data, 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] Where s represents the chord graph 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)……Equation (8);

[0105]

[0106] in, For the one-dimensional Fourier transform of the chord data along the detector direction, θ s For the network parameters of the angle-by-angle intelligent filtering network SNet, H v R(w) is the window function of the chord graph data filtering kernel, and R(w) is the ramp filtering kernel. m R is the bandwidth of the ramp filter kernel. v (w) is the angle-by-angle filtering kernel. This represents the one-dimensional inverse Fourier transform of the chord graph data along the detector direction. The data is the filtered chordogram data, and

[0107] The angle-by-angle back projection processing of the present invention is represented by equation (10);

[0108]

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

[0110] The angle direction sorting process of the present invention is as follows: the back projection tensor data are sorted in the angle direction according to the back projection tensor values ​​from small to large, as expressed by equation (11);

[0111]

[0112] in, This is the sorted backprojection tensor data. For sorting operations, : represents all backprojection tensors in the angular direction.

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

[0114]

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

[0116] This invention utilizes a pixel-by-pixel intelligent correction network, an angle-by-angle intelligent filtering network, and a back-projection tensor intelligent reconstruction network, trained end-to-end using a pre-established training database. The training database contains multiple pairs of paired normal dose measurement data, low dose measurement data, and normal dose image data reconstructed from the normal dose measurement data. Each low dose measurement data pair consists of one normal dose measurement data pair, and the doses of the low dose data are all lower than those of the normal dose data. Each low dose measurement data pair includes both two-dimensional and three-dimensional projection data, and the doses of all low dose measurement data pairs are not necessarily identical.

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

[0118] End-to-end training is performed through the following steps: First, the pixel-wise intelligent correction network is pre-trained using the first training method to obtain the pre-trained pixel-wise intelligent correction network; then, the angle-wise intelligent filtering network is pre-trained using the pre-trained pixel-wise intelligent correction network and the second training method to obtain the pre-trained angle-wise intelligent filtering network; next, the back-projection tensor intelligent reconstruction network is pre-trained using the pre-trained pixel-wise intelligent correction network, the pre-trained angle-wise intelligent filtering network, and the third training method to obtain the pre-trained back-projection tensor intelligent reconstruction network; finally, the pre-trained pixel-wise intelligent correction network, the pre-trained angle-wise intelligent filtering network, and the pre-trained back-projection tensor intelligent reconstruction network are retrained using the fourth training method to obtain the trained pixel-wise intelligent correction network, the trained angle-wise intelligent filtering network, and the trained back-projection tensor intelligent reconstruction network.

[0119] The first training method is carried out through 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 and obtain the corrected measurement data through correction processing;

[0121] A2. The parameters of the pixel-by-pixel intelligent correction network of A1 are updated by the first loss function, which is represented by equation (13).

[0122]

[0123] In equation (13), Loss p Here, N represents the loss value of the first loss function, and N is the number of training samples. The corrected measurement data obtained from A1 These are normal dose measurement data within the same paired dataset;

[0124] A3. Determine if the training epoch has reached ω1. If yes, proceed to A4; otherwise, return to A1, where ω1 is an integer greater than 2.

[0125] A4, Loss p The minimum value 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 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 and obtain the corrected measurement data through correction processing.

[0128] B2. The corrected measurement data obtained in B1 is processed by negative logarithmic transformation to obtain chord graph data;

[0129] B3. Input the chord diagram data obtained in B2 into the angle-by-angle intelligent filtering network for filtering processing to obtain the filtered chord diagram data;

[0130] B4. Update the parameters of the angle-by-angle intelligent filtering network of B3 through the second loss function, which is represented by equation (14).

[0131]

[0132] In equation (14), Loss s The loss value of the second loss function. To obtain the chord diagram data for B3, This is the chord graph data after normal dose filtering;

[0133] B5. Determine if the training epoch has reached ω2 times. If yes, proceed to B6; otherwise, return to B1, and ω2 is an integer greater than 2.

[0134] B6. Loss all loss values 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 carried out through the following steps:

[0136] C1. Input the two-dimensional 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 and obtain the corrected measurement data through correction processing;

[0137] C2. The corrected measurement data obtained in C1 is processed by negative logarithmic transformation to obtain chord graph data;

[0138] C3. Input the chord diagram data obtained in C2 into the angle-by-angle intelligent filtering network for filtering processing to obtain the filtered chord diagram data;

[0139] C4. The filtered chord graph data obtained in C3 is processed by angle-by-angle back projection to obtain back projection tensor data;

[0140] C5. Sort the backprojection tensor data obtained in C4 by angle and direction to obtain sorted backprojection tensor data.

[0141] C6. Input the sorted backprojection tensor data obtained in C5 into the backprojection tensor intelligent reconstruction network for reconstruction to obtain the CT reconstructed image.

[0142] C7. Update the parameters of the back projection tensor intelligent reconstruction network of C6 through the third loss function, which is represented by equation (15).

[0143]

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

[0145] C8. Determine if the training epoch has reached ω3. If yes, proceed to C9; otherwise, return to C1, and ω3 is an integer greater than 2.

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

[0147] The fourth training method is carried out through the following steps:

[0148] D1. Input the two-dimensional 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 and obtain the corrected measurement data through correction processing.

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

[0150] D3. Input the chord diagram data obtained in D2 into the pre-trained angle-by-angle intelligent filtering network for filtering processing to obtain the filtered chord diagram data;

[0151] D4. The filtered chord graph data obtained in D3 is processed by angle-by-angle back projection to obtain back projection tensor data.

[0152] D 5. Sort the back projection tensor data obtained in D4 by angle direction to obtain sorted back projection tensor data.

[0153] D6. Input the sorted backprojection tensor data obtained in D5 into the pre-trained backprojection tensor intelligent reconstruction network for reconstruction to obtain the CT reconstructed image.

[0154] D7. 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 are updated by the fourth loss function. The fourth loss function is represented by equation (16).

[0155]

[0156] In equation (16), Loss is the loss value of the fourth loss function. The corrected measurement data obtained from D1 For normal dose measurement data in the same paired dataset, To obtain the chord diagram data for D3, The chord graph data is the filtered normal dose measurement data from the same paired dataset; The CT reconstructed image obtained from D6. Normal dose image data corresponding to normal dose measurement data in the same paired data;

[0157] D8. Determine if the training epoch has reached ω4. If yes, proceed to D9; otherwise, return to D1, and ω4 is an integer greater than 2.

[0158] D9. Use the minimum value among all loss values ​​as the parameters of the pre-trained backprojection tensor intelligent reconstruction network, the pre-trained angle-by-angle intelligent filtering network, and the pre-trained backprojection tensor intelligent reconstruction network, and end the training.

[0159] In this embodiment, the specific values ​​of ω1, ω2, ω3, and ω4 are determined based on factors such as the size of the computer's computing power network and the amount of data. In this embodiment, ω1, ω2, ω3, and ω4 are specifically 100, and the learning rate is 1×10⁻⁶. -4 The network optimizer is Adam.

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

[0161] This deep learning-based intelligent reconstruction method for the entire CT process chain firstly employs a pixel-by-pixel intelligent correction network that learns the variance of each pixel to perform pixel-by-pixel adaptive Gaussian filtering on low-dose raw measurement data. Secondly, a pixel-by-angle intelligent filtering network learns the filtering kernel for each projection angle to perform angle-by-angle adaptive filtering on chordogram data. Finally, a back-projection tensor intelligent reconstruction network learns the mapping from the back-projection tensor to the target image, enabling it to maintain image quality while reducing scan dose.

[0162] Example 2

[0163] A deep learning-based intelligent reconstruction method for the entire CT process is presented. Other features are the same as in Example 1. In this example, before reconstruction and training, the imaging system parameters of the CT device are obtained, including: number of detector channels nb = 736, detector size ds = 1.2858 mm, source-to-detector distance dsd = 1085.6 mm, source-to-rotation center distance dso = 595 mm, number of projection angles na = 1152, number of image pixels nx = 512, and image pixel size dx = 0.6934 mm.

[0164] The normal projection data in this embodiment includes normal dose projection data for the head, chest, and abdomen, and image data. 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 backprojection tensor intelligent reconstruction network RNet The dimensions are 512×512×1152. The number of input channels in the pixel-wise intelligent correction network PNet and the angle-wise intelligent filtering network SNet is 1, while the number of input channels in the backprojection tensor intelligent reconstruction network RNet is 1152. Furthermore, the window width of the pixel-wise Gaussian filter kernel is k∈{3,5,7,9}.

[0167] Figure 2Low-dose image data is obtained by reconstructing low-dose measurement data, with a dose of 1 / 16 of the normal dose. Figure 3 It is by Figure 2 The low-dose measurement data are 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 Low-dose images contain severe noise and streak artifacts. Figure 3 The results of this invention have essentially removed 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 solutions of this invention and not to limit the scope of protection of this invention. Although this invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this invention without departing from the essence and scope of this invention.

Claims

1. A deep learning-based intelligent reconstruction method for the entire CT scan chain, characterized in that, Includes the following steps: S1. Input the raw measurement data (below the normal dose) into the trained pixel-by-pixel intelligent correction network and process it to obtain the corrected measurement data; the raw measurement data includes two-dimensional projection data and three-dimensional projection data. S2. The corrected measurement data obtained in S1 is processed by negative logarithmic transformation to obtain chord diagram data; S3. Input the chord diagram data obtained in S2 into the trained angle-by-angle intelligent filtering network and process it to obtain the filtered chord diagram data. S4. The filtered chord graph data obtained in S3 is processed by angle-by-angle back projection to obtain back projection tensor data. S5. Sort the back projection tensor data obtained in S4 by angle and direction to obtain sorted back projection tensor data. S6. Input the sorted backprojection tensor data obtained in S5 into the trained backprojection tensor intelligent reconstruction network for reconstruction processing to obtain the final CT reconstruction image. The correction process is represented by equations (1) to (4); ...Equation (1); ...Equation (2); ...Equation (3); ...Equation (4); in, For two-dimensional projection data, U This represents the number of detector channels. V The number of projection angles. For three-dimensional projection data, For pixel-by-pixel intelligent correction network For network parameters, The standard deviation of the pixel-by-pixel Gaussian filter kernel. The window width mapping function for the Gaussian filter kernel. The window width of the pixel-by-pixel Gaussian filter kernel. For pixel-by-pixel 3D Gaussian filter kernel, For the corrected measurement data, x , y and z For spatial coordinates, u Let be the coordinates in the direction of the detector channel. v Let be the coordinates in the angular direction; The filtering process is represented by equations (6) to (9); ...Equation (6); ...Equation (7); ...Equation (8); ...Equation (9); in, This represents the one-dimensional Fourier transform of the chord data along the detector direction. For angle-by-angle intelligent filtering network Network parameters, The window function for filtering chord graph data. for Filter kernel, for The bandwidth of the filter kernel Angle-by-angle filtering kernel, This represents the one-dimensional inverse Fourier transform of the chord graph data along the detector direction. The data is the filtered chordogram data, and .

2. The deep learning-based intelligent reconstruction method for the entire CT chain according to claim 1, characterized in that: The negative logarithmic transformation process is represented by equation (5); ...Equation (5); in, This is chord diagram data.

3. The deep learning-based intelligent reconstruction method for the entire CT chain according to claim 2, characterized in that: The angle-by-angle back projection process is represented by equation (10); ...Equation (10); in, For back-projection tensor data, and , W The width of the image, J For the height of the image, V The number of projection angles. For linear interpolation operations, For the first The rotation angle corresponding to each angle projection. This is the back projection weighting factor.

4. The deep learning-based intelligent reconstruction method for the entire CT chain according to claim 3, characterized in that, The angle direction sorting process is as follows: the back projection tensor data are sorted in the angle direction according to the back projection tensor values ​​from small to large, as expressed by equation (11); ...Equation (11); in, This is the sorted backprojection tensor data. For sorting operations, : represents all backprojection tensors in the angular direction.

5. The deep learning-based intelligent reconstruction method for the entire CT chain according to claim 4, characterized in that: The reconstruction process is represented by equation (12); ...Equation (12); in, For the final CT reconstructed image, and , Intelligent Reconstruction Network for Back-Projection Tensors Network parameters.

6. The deep learning-based intelligent reconstruction method for the entire CT chain according to any one of claims 1 to 5, 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 through a pre-established training database to obtain the trained pixel-by-pixel intelligent correction network, the trained angle-by-angle intelligent filtering network, and the trained back-projection tensor intelligent reconstruction network. The training database contains multiple pairs of paired normal dose measurement data, low dose measurement data, and normal dose image data reconstructed from the normal dose measurement data. A low dose measurement data and a normal dose measurement data form a paired data set, and the doses of the low dose data are all lower than those of the normal dose data. Each low dose measurement data set includes two-dimensional projection data and three-dimensional projection data, and the doses of all low dose measurement data sets are not all the same.

7. The deep learning-based intelligent reconstruction method for the entire CT chain according to claim 6, characterized in that, The end-to-end training is performed through the following steps: First, the pixel-by-pixel intelligent correction network is pre-trained using a 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 using the pre-trained pixel-by-pixel intelligent correction network and a second training method to obtain a pre-trained angle-by-angle intelligent filtering network; next, the back-projection tensor intelligent reconstruction network is pre-trained using 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 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 retrained using a 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.

8. The deep learning-based intelligent reconstruction method for the entire CT chain according to claim 7, characterized in that, The first training method is performed through the following steps: 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 and process them to obtain the corrected measurement data. A2. Update the parameters of the pixel-by-pixel intelligent correction network of A1 using the first loss function, which is represented by equation (13); ...Equation (13); In equation (13), The loss value of the first loss function. The number of training samples. The corrected measurement data obtained from A1 These are normal dose measurement data within the same paired dataset; A3. Determine if the training epoch has been reached. Next, if yes, proceed to A4; otherwise, return to A1, and Integers greater than 2; A4 The minimum value is used as the parameter of the pre-trained pixel-by-pixel intelligent correction network, and the training ends; The second training method is carried out through the following steps: 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 and obtain the corrected measurement data through correction processing; B2. The corrected measurement data obtained in B1 is processed by negative logarithmic transformation to obtain chord graph data; B3. Input the chord diagram data obtained in B2 into the angle-by-angle intelligent filtering network for filtering processing to obtain the filtered chord diagram data; B4. Update the parameters of the angle-by-angle intelligent filtering network of B3 through the second loss function, which is represented by equation (14); ...Equation (14); In equation (14), The loss value of the second loss function. To obtain the chord diagram data for B3, This is the chord graph data after normal dose filtering; B5. Determine if the training epoch has been reached. Next, if yes, proceed to B6; otherwise, return to B1, and It is an integer greater than 2; B6. All loss values 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 through the following steps: 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 and obtain the corrected measurement data through correction processing; C2. The corrected measurement data obtained in C1 is processed by negative logarithmic transformation to obtain chord graph data; C3. Input the chord diagram data obtained in C2 into the angle-by-angle intelligent filtering network for filtering processing to obtain the filtered chord diagram data; C4. The filtered chord graph data obtained in C3 is processed by angle-by-angle back projection to obtain back projection tensor data; C5. Sort the backprojection tensor data obtained in C4 by angle and direction to obtain sorted backprojection tensor data. C6. Input the sorted backprojection tensor data obtained in C5 into the backprojection tensor intelligent reconstruction network for reconstruction to obtain the CT reconstructed image. C7. Update the parameters of the back projection tensor intelligent reconstruction network of C6 through the third loss function, which is represented by equation (15); ...Equation (15); In equation (15), The loss value of the third loss function. The CT reconstructed image obtained from C6, Normal dose image data corresponding to normal dose measurement data in the same paired data; C8. Determine if the training epoch has been reached. Next, if yes, proceed to C9; otherwise, return to C1, and It is an integer greater than 2; C9. All loss values The minimum value is used as the parameter of the pre-trained backprojection tensor intelligent reconstruction network, and the training ends. The fourth training method is performed through the following steps: 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 and obtain the corrected measurement data through correction processing; D2. The corrected measurement data obtained in D1 is processed by negative logarithmic transformation to obtain chord diagram data; D3. Input the chord graph data obtained in D2 into the pre-trained angle-by-angle intelligent filtering network for filtering processing to obtain the filtered chord graph data; D4. The filtered chord graph data obtained in D3 is processed by angle-by-angle back projection to obtain back projection tensor data. D5. Sort the back projection tensor data obtained in D4 by angle and direction to obtain sorted back projection tensor data. D6. Input the sorted backprojection tensor data obtained in D5 into the pre-trained backprojection tensor intelligent reconstruction network for reconstruction to obtain the CT reconstruction image. D7. 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 are updated by the fourth loss function, which is represented by equation (16). ...Equation (16); In equation (16), The loss value of the fourth loss function. The corrected measurement data obtained from D1 For normal dose measurement data in the same paired dataset, To obtain the chord diagram data for D3, The chord graph data is the filtered normal dose measurement data from the same paired dataset; The CT reconstructed image obtained from D6. Normal dose image data corresponding to normal dose measurement data in the same paired data; D8. Determine if the training epoch has been reached. Next, if yes, proceed to D9; otherwise, return to D1, and It is an integer greater than 2; D9. All loss values The minimum value is used as the parameter of the pre-trained backprojection tensor intelligent reconstruction network, the pre-trained angle-by-angle intelligent filtering network, and the pre-trained backprojection tensor intelligent reconstruction network, and the training ends.

Images