Sparse pyramid beam CT imaging reconstruction method and device

Abstract

The invention relates to a sparse pyramid beam CT imaging reconstruction method and device, and the method comprises the steps: carrying out the FDK reconstruction of input projection data under a sparse angle condition, and carrying out the normalization processing, and obtaining an initial reconstruction image x0; inputting the initial reconstructed image x0 into a convolutional neural network for processing to obtain an initial reconstructed image x1; the initial reconstructed image x1 is input to a gradient item of a CBCT reconstruction model, a reconstructed image z is obtained, and the CBCT reconstruction model comprises a sparse item, a TV item, a gradient item and a near-end item; inputting the reconstructed image z into a convolutional neural network for processing to obtain an initial reconstructed image u; inputting the reconstructed image u into a near-end item and a TV item of a CBCT reconstruction model to obtain a reconstructed image x (n + 1); and calculating a loss function, and training to obtain a final CBCT reconstruction model. According to the method, the image quality and the detection efficiency of sparse pyramid beam CT imaging can be effectively improved.

Description

Technical Field

[0001] The present invention relates to a sparse cone-beam CT imaging reconstruction method and device. Background Art

[0002] Industrial X-ray cone-beam computed tomography (CBCT) is an important nondestructive testing method for three-dimensional visualization of the internal structure of workpieces. It has become a core technology for modern industrial quality inspection. CBCT imaging offers advantages such as high spatial resolution, high detection efficiency, and strong reliability. It is widely used in defect detection, dimensional measurement, and reverse engineering in aerospace, automotive manufacturing, electronic packaging, and other fields.

[0003] Although high-precision CBCT is of great value in industrial inspection, traditional scanning methods usually require a long acquisition time, which not only increases the cost of inspection but may also cause radiation effects on certain sensitive materials. To improve inspection efficiency, the current main methods are to reduce the projection angle to achieve sparse sampling reconstruction and adopt limited-angle scanning (such as local scanning within 180°). However, these methods will lead to incomplete projection data, which in turn affects the reconstruction quality. Traditional CBCT imaging reconstruction algorithms mainly include FDK, which is simple to implement but has strict requirements on data completeness. Reconstruction under sparse angle conditions will have serious stripe artifacts, resulting in a significant decrease in quality. It is difficult to meet the quality requirements of industrial inspection and has obvious limitations. Summary of the Invention

[0004] The object of the present invention is to provide a sparse cone-beam CT imaging reconstruction method and device, which can effectively improve the image quality and detection efficiency of sparse cone-beam CT imaging.

[0005] Based on the same inventive concept, the present invention has two independent technical solutions:

[0006] 1. A sparse cone-beam CT imaging reconstruction method, characterized by comprising the following steps:

[0007] Step 1: Perform FDK reconstruction under sparse angle conditions on the input projection data and perform normalization processing to obtain the initial reconstructed image x0;

[0008] Step 2: Input the initial reconstructed image x0 into the convolutional neural network for processing to obtain the initial reconstructed image x1;

[0009] Step 3: Input the initial reconstructed image x1 into the gradient term of the CBCT reconstruction model to obtain a reconstructed image z. The gradient term module updates the reconstructed image through a gradient descent operation. The CBCT reconstruction model includes a sparse term, a TV term, a gradient term, and a proximal term.

[0010] Step 4: Input the reconstructed image z into the convolutional neural network for processing to obtain the initial reconstructed image u;

[0011] Step 5: Input the reconstructed image u into the proximal term and TV term of the CBCT reconstruction model to obtain the reconstructed image x (n +1) ;

[0012] Step 6: Calculate the loss function and train to obtain the final CBCT reconstruction model.

[0013] Furthermore, in step 3, the reconstructed image x1 is input into the attention module to obtain the detail extracted image x_detail, and the attention module is used to extract details from the image;

[0014] In step 5, the reconstructed image x n+1 The final reconstructed image x_final is obtained by superimposing the detail extracted image x_detail.

[0015] Furthermore, the CBCT reconstruction model is as follows:

[0016]

[0017] In formula (1) is the data fidelity term, which forces the reconstruction result x to be consistent with the projection data y under the projection matrix A, where ||·||2 is the L2 norm, which represents the square root of the sum of the squares of the vector elements. ||·||1 is the L1 norm, which represents the sum of the absolute values ​​of the vector elements. y is the projection data, with a dimension of 768×768. x is the image to be reconstructed, with a dimension of 512×512×400, where 512 is the length and width of the reconstructed image and 400 is the thickness of the reconstructed layer. A is the projection matrix, A∈R M×N , where M is the dimension of the projection data (number of measurements × number of detector units), N is the dimension of the image to be reconstructed (total number of voxels = 512 × 512 × 400), and W represents the sparse transformation matrix (such as wavelet transform, etc.). Represent the gradients in the x, y, and z directions of the reconstructed image x, respectively. λ1 represents the regularization coefficient of the sparse transformation term. λ2, λ3, and λ4 represent the regularization coefficients of the x-direction gradient term, the y-direction gradient term, and the z-direction gradient term, respectively.

[0018] Furthermore, equation (1) is decomposed into two independent proximal operations, which are:

[0019]

[0020] Solving equation (2) can be transformed into solving the gradient term and the proximal term, as follows:

[0021] Gradient term:

[0022] z k =x k -tA T (Ax k -y) (4) Proximal term:

[0023]

[0024] In the formula In each iteration, the intermediate variable after the proximal term is updated and before the TV term is updated. t is the step size of the gradient descent, A T is the transpose of the projection matrix, z k is the reconstruction result after the gradient term is updated, x k+1 is the reconstruction result after the proximal term is updated, that is,

[0025] Furthermore, the Chambolle-Pock primal-dual optimization algorithm is used to solve equation (3) and obtain the reconstructed image x (n +1) .

[0026] Furthermore, based on the Chambolle-Pock primal-dual optimization algorithm, the optimization result of formula (3) is:

[0027]

[0028] Among them, p,q,r,s are all dual variables, and p (0) ,q (0) , r (0) , s (0) are all 0, x (0) is the original variable and The final iterative result, L is ||K||2, the K value can be calculated by formula (12), σ and τ are the iterative step sizes, and the value is always 1 / L. The final iterative result

[0029] Furthermore, based on the Chambolle-Pock primal-dual optimization algorithm, the optimization result of formula (3) is:

[0030]

[0031] Among them, SC_PA(·) is a pixel attention module that performs weighted sum processing on the feature map to improve the feature expression ability of a specific area, thereby effectively replacing the pixel attention module in formulas (28)-(30). Structure such asFigure 4 As shown; and Represents convolutional layers of different structures, Represents λ i ,σ,andτ,these parameters are Represents the inversion of the gradient in the x, y, and z directions respectively.

[0032] Furthermore, the system parameters include the distance between the ray source and the detection center, the distance between the ray source and the detector, the detector size, the detector interval size, the projection size, the reconstructed image size and the sampling interval. According to the system parameters, a set of high-quality 3D CT images are projected at 360° equal intervals to obtain the projection of the 3D CT image. Then, the FDK reconstructed image of the 360-angle projection is used as the label image u. k , 360 degrees of angle are sampled at equal intervals to obtain the sparse projection y k , thus forming the training set data pair {u k ,y k}, collect sparse angle projections according to system parameters to form a test set.

[0033] Furthermore, the attention module consists of a 3-dimensional channel attention module and a 3-dimensional spatial attention module, the 3-dimensional channel attention module is used to weight the channel attention, and the 3-dimensional spatial attention module is used to weight the spatial dimension.

[0034] 2. A sparse cone-angle beam CT imaging reconstruction device, used to execute the above-mentioned sparse cone-angle beam CT imaging reconstruction method.

[0035] The present invention has the beneficial effects:

[0036] The present invention performs FDK reconstruction under sparse angle conditions on the input projection data and performs normalization processing to obtain an initial reconstructed image x0; the initial reconstructed image x0 is input into a convolutional neural network for processing to obtain an initial reconstructed image x1; the initial reconstructed image x1 is input into the gradient term of the CBCT reconstruction model to obtain a reconstructed image z, and the gradient term module updates the reconstructed image through a gradient descent operation; the CBCT reconstruction model includes a sparse term, a TV term, a gradient term, and a proximal term; the reconstructed image z is input into a convolutional neural network for processing to obtain an initial reconstructed image u; the reconstructed image u is input into the proximal term and TV term of the CBCT reconstruction model to obtain a reconstructed image x (n+1) ; Calculate the loss function and train to obtain the final CBCT reconstruction model.

[0037] This method combines deep learning with iterative reconstruction algorithms. By expanding the complex iterative reconstruction process into a series of trainable neural network layers, it not only retains the high precision and robustness of iterative algorithms, but also leverages convolutional neural networks (CNNs) for efficient data processing and rapid convergence, significantly improving image quality and detail preservation. The iterative reconstruction subnetwork of this invention integrates the advantages of sparse representation and total variation (TV) regularization. This combination not only helps preserve the key structural features of 3D images but also effectively suppresses noise, resulting in clearer and more accurate reconstruction results.

[0038] The present invention is highly flexible and adaptable, can adjust the network structure according to specific application scenarios, and supports an end-to-end learning framework, simplifying the entire process from raw data to the final reconstructed image. Especially under sparse angle scanning conditions, the present invention can extract richer effective information from limited projection data, while ensuring reconstruction accuracy, significantly shortening the scanning time and effectively improving detection efficiency. The present invention has great engineering value for achieving efficient and accurate sparse angle industrial CBCT reconstruction, and can effectively meet the urgent needs of modern manufacturing for fast, accurate non-destructive testing. Especially in industrial application scenarios such as full inspection of high-value parts and precision assembly verification, it can greatly improve detection efficiency, reduce production costs, and provide strong technical support for intelligent manufacturing.

[0039] The present invention inputs the reconstructed image x1 into the attention module to obtain the detail extraction image x_detail, and the attention module is used to extract the details of the image; the reconstructed image x n+1 The detail extraction image x_detail is superimposed to obtain the final reconstructed image x_final. The attention module consists of a 3D channel attention module and a 3D spatial attention module. The 3D channel attention module is used to weight the channel attention, and the 3D spatial attention module is used to weight the spatial dimension. The present invention combines the iterative reconstruction subnetwork with the detail recovery subnetwork. The detail recovery subnetwork adopts a pair of interdependent spatial and channel attention branches, which can efficiently learn the discriminative features of the space and channels in the three-dimensional image data, and restore the detail information in the initial three-dimensional image in an adaptive manner. This strategy significantly improves the quality of CBCT reconstructed images under sparse angle conditions and improves the overall reconstruction efficiency.

[0040] The CBCT reconstruction model of the present invention is as follows:

[0041]

[0042] In formula (1) is the data fidelity term, which forces the reconstruction result x to be consistent with the projection data y under the projection matrix A, where ||·||2 is the L2 norm, which represents the square root of the sum of the squares of the vector elements. ||·||1 is the L1 norm, which represents the sum of the absolute values ​​of the vector elements. y is the projection data, with a dimension of 768×768. x is the image to be reconstructed, with a dimension of 512×512×400, where 512 is the length and width of the reconstructed image and 400 is the thickness of the reconstructed layer. A is the projection matrix, A∈R M×N , where M is the dimension of the projection data (number of measurements × number of detector units), N is the dimension of the image to be reconstructed (total number of voxels = 512 × 512 × 400), and W represents the sparse transformation matrix (such as wavelet transform, etc.). Represent the gradients in the x, y, and z directions of the reconstructed image x, respectively. λ1 represents the regularization coefficient of the sparse transformation term. λ2, λ3, and λ4 represent the regularization coefficients of the x-direction gradient term, the y-direction gradient term, and the z-direction gradient term, respectively.

[0043] Decompose Equation (1) into two independent proximal operations, which are:

[0044]

[0045] Solving equation (2) can be transformed into solving the gradient term and the proximal term, as follows:

[0046] Gradient term:

[0047] z k =x k -tA T (Ax k -y) (4) Proximal term:

[0048]

[0049] In the formula In each iteration, the intermediate variable after the proximal term is updated and before the TV term is updated. t is the step size of the gradient descent, A T is the transpose of the projection matrix, z k is the reconstruction result after the gradient term is updated, x k+1 is the reconstruction result after the proximal term is updated, that is,

[0050] Based on the Chambolle-Pock primal-dual optimization algorithm, the optimization result of formula (3) is:

[0051]

[0052]

[0053] Among them, SC_PA(·) is a pixel attention module that performs weighted sum processing on the feature map to improve the feature expression ability of a specific area, thereby effectively replacing the pixel attention module in formulas (28)-(30). Structure such as Figure 4 As shown; and Represents convolutional layers of different structures, Represents λ i ,σ,andτ,these parameters are Represents the inversion of the gradient in the x, y, and z directions respectively.

[0054] The present invention further ensures effective improvement of the image quality and detection efficiency of sparse cone-beam CT imaging through the above-mentioned CBCT reconstruction model and solution algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is the overall flow chart of the sparse cone-beam CT imaging reconstruction method of the present invention;

[0056] Figure 2 Schematic diagram of the structure of the reconstruction submodule and the attention submodule learning update in each iteration of the present invention;

[0057] Figure 3 This is a schematic diagram of the structure of the convolutional neural network after the TV item iteration process is expanded;

[0058] Figure 4 This is a schematic diagram of the pixel attention module SC-PA structure;

[0059] Figure 5 This is a diagram of the working principle of the three-dimensional U-Net convolutional neural network;

[0060] Figure 6 It is a structural diagram of the image restoration submodule. DETAILED DESCRIPTION

[0061] The present invention is described in detail below with reference to the various embodiments shown in the accompanying drawings, but it should be noted that these embodiments are not limitations of the present invention, and any equivalent transformations or substitutions in functions, methods, or structures made by ordinary technicians in this field based on these embodiments are all within the scope of protection of the present invention.

[0062] Example 1:

[0063] Sparse Cone-Beam CT Imaging Reconstruction Method

[0064] (1) CBCT reconstruction model

[0065] 1. The CBCT reconstruction model is as follows:

[0066]

[0067] In formula (1) is the data fidelity term, which forces the reconstruction result x to be consistent with the projection data y under the projection matrix A, where ||·||2 is the L2 norm, which represents the square root of the sum of the squares of the vector elements. ||·||1 is the L1 norm, which represents the sum of the absolute values ​​of the vector elements. y is the projection data, with a dimension of 768×768. x is the image to be reconstructed, with a dimension of 512×512×400, where 512 is the length and width of the reconstructed image and 400 is the thickness of the reconstructed layer. A is the projection matrix, A∈R M×N , where M is the dimension of the projection data (number of measurements × number of detector units), N is the dimension of the image to be reconstructed (total number of voxels = 512 × 512 × 400), and W represents the sparse transformation matrix (such as wavelet transform, etc.). Represent the gradients in the x, y, and z directions of the reconstructed image x, respectively. λ1 represents the regularization coefficient of the sparse transformation term. λ2, λ3, and λ4 represent the regularization coefficients of the x-direction gradient term, the y-direction gradient term, and the z-direction gradient term, respectively.

[0068] 2. Decomposition formula (1)

[0069] Formula (1) is decomposed into two independent proximal operations, specifically as follows:

[0070]

[0071] 3. Solve equation (2)

[0072] Solving equation (2) can be transformed into solving the gradient term and the proximal term, as follows:

[0073] Gradient term:

[0074] z k =x k -tA T (Ax k -y) (4)

[0075] Proximal term:

[0076]

[0077] In the formula In each iteration, the intermediate variable after the proximal term is updated and before the TV term is updated. t is the step size of the gradient descent, A T is the transpose of the projection matrix, z k is the reconstruction result after the gradient term is updated, x k+1 is the reconstruction result after the proximal term is updated, that is,

[0078] Where W is a sparse transform, which can be a wavelet transform or a discrete cosine transform (DCT).

[0079] 4. Solve equation (3)

[0080] Equation (3) is solved based on the Chambolle-Pock primal-dual optimization algorithm (CP algorithm). The CP algorithm can solve various convex optimization problems well, solving both the primal problem and the dual problem in one loop while ensuring that each subproblem has a closed-form solution.

[0081] The general primitive minimization form of the CP algorithm is as follows:

[0082] min x {F(Kx)+G(x)} (6)

[0083] The general form of the primitive dual maximization in formula (6) is as follows:

[0084] max y {-F * (yG * (-K T y)} (7)

[0085] In Equations (6) and (7), x and y are vectors in spaces X and Y, respectively; K is a linear transformation from X to Y; the CP framework requires that the two functions G and F be convex but not smooth, mapping their respective X and Y spaces to nonnegative real numbers. "T" denotes transpose, and "*" denotes convex conjugation in the dual maximization problem.

[0086] The conjugate convex function of any convex function H(Z) is defined as formula (8):

[0087] H * (Z)=max c {<z,c> -H(c)} (8)

[0088] Where <·,·> refers to the inner product of two vectors, and c is z′. Therefore, the elementary binary problem can be transformed into a saddle point optimization problem. Its form is as follows

[0089] min x max y {<Kx,y> Y +G(x)-F * (y)} (9)

[0090] Then maximize y and minimize x.

[0091] First, transform Equation (3) into Equation (10) to perform Chambolle-Pock primal-dual optimization solution:

[0092]

[0093] Formula (10) corresponds to Formula (6) according to the following series of transformations.

[0094]

[0095] This leads to the following equation:

[0096] F(y,e,h,o)=F1(y)+F2(e)+F3(h)+F4(o) (13)

[0097]

[0098] F2(e)=λ2||(|e|)||1, F3(h)=λ3||h|)||1, F4(o)=λ4||(|o|)||1 (15)

[0099] G(x)=0 (16)

[0100] In formula (13), F(y, e, h, o) is a convex function because it is the sum of four convex functions. According to formula (8), we can get

[0101]

[0102] δ Box(a) (|x|) is an indicator function, which is defined as follows:

[0103]

[0104] Approximate mapping is essentially an optimization problem, but since z is a variable, the result is a function of z.

[0105] According to formula (20), the approximate mapping between the above conjugate convex function and G(x) can be obtained as follows:

[0106]

[0107] prox τ G(x)=x, (25)

[0108] From the above formula, we can get F * An approximate mapping of :

[0109]

[0110] Therefore, the optimization result of formula (3) is:

[0111]

[0112] Among them, p,q,r,s are all dual variables, and p (0) ,q (0) , r (0) , s (0) are all 0, x (0) is the original variable and The final iterative result, L is ||K||2, the K value can be calculated by formula (12), σ and τ are the iterative step sizes, and the value is always 1 / L. The final iterative result

[0113] The final iteration result x (n+1) Right now This is the final reconstruction result.

[0114] During implementation, the system parameters include the distance between the ray source and the detection center, the distance between the ray source and the detector, the detector size, the detector interval size, the projection size, the reconstructed image size, and the sampling interval. Based on the system parameters, a set of high-quality 3D CT images are projected at 360° equal-interval angles to obtain the projection of the 3D CT image. Then, the FDK reconstructed image of the 360-angle projection is used as the label image u. k , 360 degrees of angle are sampled at equal intervals to obtain the sparse projection y k , thus forming the training set data pair {u k ,y k}, collect sparse angle projections according to system parameters to form a test set.

[0115] (2) Reconstruction method based on the above CBCT reconstruction model

[0116] like Figure 1 、 Figure 2 As shown, the sparse cone-beam CT imaging reconstruction method of the present invention includes the following steps:

[0117] Step 1: Perform FDK reconstruction under sparse angle conditions on the input projection data and perform normalization to obtain the initial reconstructed image x0.

[0118] Step 2: Input the initial reconstructed image x0 into the convolutional neural network for processing to obtain the initial reconstructed image x1.

[0119] During implementation, a 3D U-Net convolutional neural network is used. The initial reconstructed image x0 enters a 3D U-Net to obtain an updated initial reconstructed image x1, with the goal of improving the quality of the sparse FDK reconstructed image x0 in the image domain.

[0120] Step 3: Input the initial reconstructed image x1 into the gradient term of the CBCT reconstruction model to obtain a reconstructed image z, and the gradient term module updates the reconstructed image through a gradient descent operation;

[0121] The CBCT reconstruction model is as described above, including a sparse term, a TV term, a gradient term, and a proximal term.

[0122] The gradient term is shown in equation (4) of the CBCT reconstruction model above,

[0123] z k =x k -tA T (Ax k -y)

[0124] During implementation, in order to increase the flexibility of the network, the step size t can be set to be adaptive and change as the number of rounds of the network changes.

[0125] The reconstructed image x1 is simultaneously input into the attention module to obtain the detail extraction image x_detail, and the attention module is used to extract details from the image.

[0126] The attention module consists of a 3D channel attention module and a 3D spatial attention module. The 3D channel attention module is used to weight the channel attention, and the 3D spatial attention module is used to weight the spatial dimension.

[0127] Step 4: Input the reconstructed image z into the convolutional neural network for processing to obtain the initial reconstructed image u.

[0128] During implementation, the convolutional neural network uses a three-dimensional U-Net, which plays the same role as step 2. It can strengthen the information flow between different layers and achieve feature enhancement. At the same time, it is robust to noise and disturbance, improves the image quality of single-iteration reconstruction, and obtains a high-quality three-dimensional image u.

[0129] Step 5: Input the reconstructed image u into the proximal term and TV term of the above CBCT reconstruction model to obtain the reconstructed image x (n+1) .

[0130] The purpose of the proximal term of the CBCT reconstruction model is to remove the intermediate result z by performing thresholding in a certain transformation domain. k The noise and artifacts in the image are updated to be reconstructed as

[0131] The purpose of the TV term in the CBCT reconstruction model is to preserve the image edges while removing noise and smoothing unwanted details, and finally obtain the reconstructed image x (n+1) .

[0132] Among them, the TV item is expanded to the network as follows:

[0133]

[0134] Among them, SC_PA(·) is a pixel attention module, which performs weighted sum processing on the feature map to improve the feature expression ability of a specific area, thereby effectively replacing the and operate. and Indicates the convolutional layers of different structures. For the specific network diagram, see Figure 4 shown.

[0135] In the present invention, each iteration process is processed in sequence by the gradient term, the approximation term and the total variation (TV) regularization term, where the variable input to the TV term is The specific iteration method of the TV term is shown in Equations (27) to (32). To further improve the reconstruction efficiency and expression ability, the advantages of deep learning networks in modeling are utilized to replace some traditional modules and construct a network-based iterative framework. The iterative process is shown in Equations (33) to (37). In addition, a pixel attention mechanism is introduced to adaptively adjust the corresponding weights of features in spatial positions. Compared with the traditional max operation that only selects the maximum value, the pixel attention mechanism can more flexibly highlight the key feature areas through the weight distribution obtained by learning, thereby improving the feature representation ability and robustness while ensuring the sparsity constraint effect.

[0136] The reconstructed image x n+1 The final reconstructed image x_final is obtained by superimposing the detail extracted image x_detail.

[0137] Step 6: Calculate the loss function and train to obtain the final CBCT reconstruction model.

[0138] First, the MSE (mean square error) loss is used to minimize the pixel-level difference between the reconstructed 3D image and the standard reference image. The standard reference image is obtained by FDK reconstruction of industrial samples under standard inspection conditions using high-precision scanning data from all angles. Then, the SSIM (structural similarity) loss is used to improve the structural fidelity, fully retain the structural information, and enhance the detail effect. The initial learning rate is set to 10 -4 , gradually changes to 10 during training -6 Observe the reconstruction results of the training set and the change of the loss function, and save the final model when the model reaches the optimal value.

[0139] like Figure 5As shown in the figure, it is a three-dimensional U-Net, designed to enhance the interaction of feature information between different layers of the network, achieving feature enhancement. This module has excellent noise and interference immunity, effectively improving image quality during iterative reconstruction. When broken down into smaller modules, it also consists of two consecutive convolutions, downsampling, upsampling, feature fusion, and a final convolution. The basic CNN unit is composed of convolutional layers, activation functions, batch normalization layers, and bypass connections. The 3*3*3 convolutional layers extract feature maps, while the activation functions perform nonlinear transformations on the feature maps. The batch normalization layers introduce normalization and shifting steps during each nonlinear transformation, effectively addressing the transfer of internal covariates during network training and accelerating network convergence. Bypass connections prevent gradient vanishing and enhance feature reuse. The main architecture of the UNet is upsampling and downsampling. During downsampling, the number of channels in the feature map is typically increased through convolution operations, allowing the network to learn more feature representations (for example, the number of channels increases from 1 to 64, then from 64 to 128, then to 256, and even further to 512 after each downsampling). At the same time, downsampling also reduces the resolution, namely width, height and depth, which is achieved through the maximum pooling operation. Upsampling is the opposite of downsampling. The number of channels is reduced and the resolution is increased. Trilinear interpolation can be used to enlarge the size of the feature map. The activation functions used by this UNet are ReLU and Tanh. The ReLU activation function is applied after each convolutional layer to increase the nonlinear ability of the network and help the model learn more complex patterns. The Tanh activation function is applied to the last layer to limit the output to the range of [-1,1].

[0140] The image restoration submodule is a three-dimensional attention module based on the two-dimensional CBAM attention. It is an attention mechanism for computer vision tasks. It can help the model pay more attention to important areas when processing images, thereby improving the performance of the model. Figure 6As shown in Figure 1, it consists of two components: the 3D Channel Attention Module and the 3D Spatial Attention Module. These two modules weight the channel and spatial dimensions, respectively, thereby improving the model's ability to extract image features. The 3D Channel Attention Module obtains channel attention weights through global average pooling and global max pooling operations. These pooling results are processed using a shared fully connected layer (here, a 3D convolutional layer), and the final channel attention weights are generated using a sigmoid activation function. The 3D Spatial Attention Module calculates the average and maximum values ​​of the input tensor along the channel dimension, concatenates these two tensors along the channel dimension, and then processes them through a 3D convolutional layer to generate spatial attention weights. Finally, combining the channel and spatial attention mechanisms enhances the model's ability to capture 3D data features. By integrating the 3D CBAM module, our model effectively learns rich spatial and channel feature representations, significantly improving the quality of image restoration.

[0141] like Figure 3 As shown in Figure 1, the TV term iteration process is expanded into the convolutional neural network. One iteration consists of 5 modules (Block), PBlock, QBlock, RBlock, and SBlock are used to update the dual variable p respectively. (k) ,q (k) ,r (k) ,s (k) , while XBlock updates the original variable Different convolutional layers are used to simulate variables such as p, q, r, and s in the CP algorithm for calculation. and Denotes the convolutional layers of different structures, SC_PA(·) is the pixel attention module. First, the convolutional layer module is used to calculate the input feature map x (k) Gradient Specifically, The convolution kernel of the convolution layer is [-11], which is used to extract the gradient in the horizontal direction; The convolution kernel of the convolution layer is Used to extract the gradient in the vertical direction, The convolution kernel of the convolution layer is Used to extract the gradient in the depth direction. These gradient variables will be used to update the variable p in the subsequent steps. (k) ,q (k) , r (k) , s (k) . Then through the convolutional layer module and The purpose of this step is to perform convolution operations on the input features, extract important feature information, and prepare for the next step of the pixel attention mechanism. Next, the feature map output by the convolution layer is input into the pixel attention module SC_PA for weighted processing. The pixel attention mechanism improves the feature expression ability of a specific area by weighting and processing the feature map. It can adaptively adjust the feature weights at different positions to achieve an effect similar to the max operation. Subsequently, the convolution layer is used For the updated q (k+1) , r (k+1) , s (k+1) Perform transposed gradient calculation and get It is then input into the convolutional layer module In the process, feature extraction and linear combination operations are performed, and finally a new reconstructed image is obtained through parameter optimization. Automatically optimize λ during network training i , σ, and τ are parameters that ensure updating and optimization.

[0142] The pixel attention mechanism is used in the QBlock, RBlock, and SBlock modules, such as Figure 4 As shown in the figure, SC stands for Self-Calibrated Convolution and PA stands for Pixel Attention. This module realizes nonlinear mapping of input features by stacking SC-PA blocks, thereby increasing the expressive power of the model. Self-calibrated convolution technology uses the internal information of the input feature map to dynamically adjust the weights of the convolution kernel, improves the feature expression in an adaptive manner, and effectively captures and strengthens the complex features and patterns in the data. The pixel attention mechanism improves the attention to key pixel areas by weighted processing of different pixels in the feature map, thereby enhancing the accuracy and efficiency of the model in important feature extraction and task perception. The SC-PA module has the advantage of being stackable, and multiple SC-PA blocks can be used in series to form a deep feature extraction and processing network. This hierarchical feature processing mechanism enables the model to gradually optimize feature representation, improve the overall data interpretation ability and task execution effect, and enable the network to gradually optimize feature representation. U in the figure n-1 for The feature map after convolution block processing, U n It is the feature map processed by the SC-PA module, and the concat block is the concatenated tensor operation.

[0143] Example 2:

[0144] Sparse cone-beam CT imaging reconstruction device

[0145] The sparse angle cone-beam CT imaging reconstruction device is used to execute the above-mentioned sparse angle cone-beam CT imaging reconstruction method.

[0146] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent implementation methods or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

[0147] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations that come within the meaning and range of equivalents of the claims be embraced therein.

Claims

1. A sparse cone-beam CT imaging reconstruction method, characterized in that: The steps include: Step 1: Perform FDK reconstruction under sparse angle conditions on the input projection data and perform normalization processing to obtain the initial reconstructed image x0; Step 2: Input the initial reconstructed image x0 into the convolutional neural network for processing to obtain the initial reconstructed image x1; Step 3: Input the initial reconstructed image x1 into the gradient term of the CBCT reconstruction model to obtain a reconstructed image z. The gradient term module updates the reconstructed image through a gradient descent operation. The CBCT reconstruction model includes a sparse term, a TV term, a gradient term, and a proximal term. Step 4: Input the reconstructed image z into the convolutional neural network for processing to obtain the initial reconstructed image u; Step 5: Input the reconstructed image u into the proximal term and TV term of the CBCT reconstruction model to obtain the reconstructed image x (n+1) ; Step 6: Calculate the loss function and train to obtain the final CBCT reconstruction model.

2. The sparse cone-beam CT imaging reconstruction method according to claim 1, characterized in that: In step 3, the reconstructed image x1 is input into the attention module to obtain the detail extraction image x_detail. The attention module is used to extract details from the image. In step 5, the reconstructed image x n+1 The final reconstructed image x_final is obtained by superimposing the detail extracted image x_detail.

3. The sparse cone-beam CT imaging reconstruction method according to claim 1, characterized in that: The CBCT reconstruction model is as follows: Wherein, y is the data fidelity term, ||·||2 is the L2 norm, which represents the square root of the sum of the vector elements, and ||·||1 is the L1 norm, which represents the sum of the absolute values ​​of the vector elements; y is the projection data; x is the image to be reconstructed, A is the projection matrix, and A∈R M×N , where M is the dimension of the projection data, N is the dimension of the image to be reconstructed, and W represents the sparse transformation matrix. Represent the gradients in the x, y, and z directions of the reconstructed image x, respectively. λ1 represents the regularization coefficient of the sparse transformation term. λ2, λ3, and λ4 represent the regularization coefficients of the x-direction gradient term, the y-direction gradient term, and the z-direction gradient term, respectively.

4. The sparse cone-beam CT imaging reconstruction method according to claim 3, characterized in that: Decompose Equation (1) into two independent proximal operations, which are: Solving equation (2) can be transformed into solving the gradient term and the proximal term, as follows: Gradient term: z k =x k -tA T (Ax k -y) (4) Proximal term: In the formula is the intermediate variable after the proximal term is updated and before the TV term is updated in each iteration; t is the step length of the gradient descent, A T is the transpose of the projection matrix, z k is the reconstruction result after the gradient term is updated, x k+1 is the reconstruction result after the proximal term is updated, that is, 5. The sparse cone-beam CT imaging reconstruction method according to claim 4, characterized in that: Based on the Chambolle-Pock primal-dual optimization algorithm, we solve equation (3) and obtain the reconstructed image x (n+1) .

6. The sparse cone-beam CT imaging reconstruction method according to claim 5, characterized in that: Based on the Chambolle-Pock primal-dual optimization algorithm, we solve equation (3): First, transform Equation (3) into Equation (10) to perform Chambolle-Pock primal-dual optimization solution: In formula (10), The optimization result of formula (3) is: Among them, p,q,r,s are all dual variables, and p (0) ,q (0) , r (0) , s (0) are all 0, x (0) is the original variable and The final iterative result, L is ||K||2, the K value can be calculated by formula (12), σ and τ are the iterative step sizes, and the value is always 1 / L. The final iterative result 7. The sparse cone-beam CT imaging reconstruction method according to claim 5, characterized in that: Based on the Chambolle-Pock primal-dual optimization algorithm, the optimization result of formula (3) is: Among them, SC_PA(·) is a pixel attention module that performs weighted sum processing on the feature map to improve the feature expression ability of a specific area, replacing the formula (28) to the formula (30) and Represents convolutional layers of different structures, Represents λ i ,σ,andτparameters, Represents the inversion of the gradient in the x, y, and z directions respectively.

8. The sparse cone-beam CT imaging reconstruction method according to claim 7, characterized in that: System parameters include the distance between the ray source and the detection center, the distance between the ray source and the detector, the detector size, the detector interval size, the projection size, the reconstructed image size and the sampling interval, etc. According to the system parameters, a set of high-quality 3D CT images are projected at 360° equal intervals to obtain the projection of the 3D CT image; then the 360-angle projection is reconstructed by FDK as the label image u k , 360 degrees of angle are sampled at equal intervals to obtain the sparse projection y k , thus forming the training set data pair {u k ,y k }, collect sparse angle projections according to system parameters to form a test set.

9. The sparse cone-beam CT imaging reconstruction method according to claim 2, characterized in that: The attention module consists of a 3D channel attention module and a 3D spatial attention module. The 3D channel attention module is used to weight the channel attention, and the 3D spatial attention module is used to weight the spatial dimension.

10. A sparse cone-beam CT imaging reconstruction device, characterized by: Used to perform the method according to any one of claims 1 to 9.

Images