A Spectral CT Imaging Method Based on Deep Convolutional Sparse Representation Reconstruction Network

Abstract

This invention discloses a method for spectral CT imaging based on a deep convolutional sparse representation reconstruction network, belonging to the field of spectral CT reconstruction technology. The core steps of this invention include: first, establishing neural network priors and sparse representation priors; then, integrating different priors into the reconstruction model and decomposing and solving the reconstruction model; next, networking the decomposed models to form a deep convolutional sparse representation reconstruction network and training it; finally, the user can input spectral CT projection data into the trained deep convolutional sparse representation reconstruction network, outputting the reconstructed spectral CT image. This invention addresses the problem of poor reconstruction quality of spectral CT images by combining the advantages of neural network priors and sparse representation priors to establish a learnable parameter-adaptive reconstruction network. This network can reduce noise in the spectral CT reconstruction process, improving image reconstruction quality and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of computed tomography imaging technology, and more specifically, relates to an energy spectrum CT imaging method based on a deep convolutional sparse representation reconstruction network. Background Technology

[0002] Since its inception in 1971 by Godfrey Hounsfield, X-ray computed tomography (CT) has undergone several generations of upgrades in both hardware and software, achieving superior temporal, spatial, and density resolution, making it a leading representative of modern high-end medical imaging equipment. However, conventional medical CT imaging is limited to revealing the anatomical morphology of patients and cannot accurately provide information on the material composition of the examined organs, greatly restricting its wider clinical application. Addressing the shortcomings of conventional medical CT, as early as the 1970s, scholars proposed the concept of spectral CT imaging based on the energy spectrum characteristics of X-rays and their physical mechanisms of interaction with matter. However, due to the limitations of hardware technology at the time, it could not be put into practical use. In recent years, with innovations in X-ray tubes, detectors, and core algorithms, spectral CT has fundamentally changed the traditional CT imaging method that uses a single grayscale CT value as the standard. Conventional CT examinations are inevitably affected by metal artifacts and beam hardening artifacts, which can lead to missed diagnoses and misdiagnoses in severe cases. Spectral CT single-energy channel imaging and artifact removal techniques can correct the low signal problem caused by "photon starvation" after X-ray scanning of metals. It provides accurate projection data for metals and surrounding tissues, effectively suppressing common metal artifacts and other radiation-hardening artifacts, thus improving diagnostic accuracy. Spectral CT can simultaneously acquire up to 101 different keV single-energy images and material separation images ranging from 40 to 140 keV. The optimal keV single-energy images for different tissues and organs have the best contrast-to-noise ratio, which is beneficial for detecting solid lesions, improving the detection rate of small lesions, and achieving precise lesion localization.

[0003] In recent years, with the increasing use of spectral CT in clinical practice, some limitations of the existing technology have gradually become apparent. First, the image quality generated by many spectral CT scans under low photon number conditions fails to meet clinical requirements; some spectral CT examinations even generate higher radiation doses, which contradicts the internationally agreed-upon goal of "As Low As Reasonably Achievable" (ALARA). Second, spectral CT is often used in conjunction with traditional single-voltage CT, which not only risks further increasing radiation doses but also increases the workload for physicians interpreting images, both of which pose challenges to clinical application. To address these issues, recent research has mainly focused on two aspects: Firstly, from a data processing perspective, research explores how to fully utilize the statistical regularities of X-rays and prior information about the patient to recover better images under the same data conditions, such as statistical reconstruction in iterative reconstruction algorithms and total variation (TV) minimization optimization reconstruction. Secondly, from a data acquisition perspective, research focuses on how to further enhance spectral separation, reduce noise and radiation dose during data acquisition, and accelerate data acquisition time. One representative of this direction is energy-based spectral CT technology using a photon counting detector (PCD).

[0004] In recent years, with the improvement of computing power and the generation of large-scale data, data-driven learning algorithms have demonstrated excellent performance in many fields, and corresponding research has expanded to signal and image processing, computer vision, pattern recognition, artificial intelligence, medical physics, and other fields. For spectral CT, which is essentially a type of multispectral image, besides the image itself being a two-dimensional matrix, the energy spectrum can be considered as a third-dimensional data. Current spectral CT imaging methods simply process the projection data of each energy spectrum segment separately, or unfold images of different energy spectrum segments into one-dimensional signals and then stitch them together to form a two-dimensional matrix before processing. This approach ignores the inherent connections between images of different energy spectrum segments and does not fully utilize the energy spectrum information. Introducing a dilution representation-based approach can more effectively utilize the redundant information between images of different energy spectrum segments, improving the imaging quality of spectral CT under pathological data measurement conditions and ensuring the usability of the images. Furthermore, the trend of combining data-driven deep learning technology with CT image reconstruction is becoming increasingly apparent, and the reconstruction results achieved by deep learning in image reconstruction are significant. To address this, this invention proposes to start with convolutional sparse coding in the field of sparse learning, combine it with the ideas of deep neural networks, establish a joint constraint model of sparse representation and neural network restoration, realize the constraint of different feature information in the reconstructed image, and at the same time complete the trainingable network to form a deep convolutional sparse representation reconstruction network specifically for energy spectrum CT reconstruction, thereby improving the energy spectrum CT reconstruction effect.

[0005] A search revealed that Chinese patent application number 202111337950.X discloses a method and apparatus for energy spectrum CT reconstruction based on tensor nuclear norm and transform Lp norm. This method utilizes the frame set tensor nuclear norm to explore the correlation between images in different channels and improves the objective function using a constructed nonlocal similarity tensor, establishing a new reconstruction model. Through continuous iteration and a loop mechanism, the reconstruction effect is improved. This invention belongs to the statistical iterative reconstruction class, which has many parameter adjustments. Furthermore, the prior information utilized is only the image itself, without fully extracting prior information from large datasets. In contrast, the method of this invention achieves adaptive mapping between energy spectrum projection data and high-quality energy spectrum image data, belonging to the category of deep reconstruction algorithms. The reconstructed image has certain advantages in suppressing image artifacts and preserving details.

[0006] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0007] 1. The problem to be solved

[0008] To address the shortcomings of existing spectral CT reconstruction methods, this invention proposes a spectral CT imaging method based on a deep convolutional sparse representation reconstruction network, termed Deep Convolutional Sparse Representation Reconstruction Network (DCSR-Net). This method aims to improve reconstructed image quality, reduce noise artifacts, enhance the differentiation of different tissues, and achieve parameter-adaptive reconstruction. Within the statistical iterative reconstruction framework, this invention leverages the advantages of combining neural network priors and sparse representation priors to establish a learnable parameter-adaptive reconstruction network. This network can reduce noise during the spectral CT reconstruction process, thereby improving image reconstruction quality and efficiency.

[0009] 2. Technical Solution

[0010] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0011] The present invention provides a spectral CT imaging method based on a deep convolutional sparse representation reconstruction network, comprising the following steps:

[0012] Step 1: Construct the training dataset and the test set.

[0013] Step 2: Design and construct the neural network prior N(u) and sparse representation prior S(u) for energy spectrum CT reconstruction;

[0014] Step 3: Incorporate the prior knowledge into the reconstruction model and decompose and solve it to obtain the decomposed energy spectrum CT reconstruction formula;

[0015] Step 4: Network the decomposed energy spectrum CT reconstruction to form a deep convolutional sparse representation reconstruction network, and train the network using the training dataset to obtain the trained network parameters.

[0016] Step 5: Input the energy spectrum CT projection data from the test set into the trained deep convolutional sparse representation reconstruction network, output the reconstructed energy spectrum CT image, and verify the reconstruction performance.

[0017] Furthermore, the training dataset in step 1 is constructed as follows: high-quality energy-spectral CT images are obtained by collecting multi-channel energy-spectral CT images and processing them. s High-quality energy-spectral CT images were reprojected and noise was added to obtain multi-channel energy-spectral CT projection data p that had noise intensity close to that of the test set data. s , forming training data pairs {p s u s The test set was constructed by collecting multi-channel energy spectral CT projection data. t This is used to test the algorithm.

[0018] Furthermore, the neural network prior N(u) in step 2 is a three-layer lightweight residual network, consisting of two consecutive convolutions and ReLU operations, one convolution operation, and a skip summation connection.

[0019] Furthermore, the sparse representation prior S(u) in step 2, using a convolutional sparse coding model, can be expressed as:

[0020]

[0021] In the above formula, u is the energy spectrum CT image to be reconstructed, * is the convolution operator, and {f n} n＝1，2，…，N For the convolution kernel group, {M n} n＝1，2，…，N For each convolutional kernel, there are different feature map sets, and β is the regularization parameter.

[0022] Furthermore, the reconstruction model in step 3, incorporating prior knowledge, is as follows:

[0023]

[0024] In the above formula, u is the energy spectrum CT image to be reconstructed, p is the scanned and preprocessed projection data, A is the projection matrix, N(u) is the neural network prior, S(u) is the sparse representation prior, and λ1 and λ2 are regularization parameters.

[0025] Furthermore, in step 3, the reconstruction model is decomposed and solved. The decomposed energy spectrum CT reconstruction formula is as follows:

[0026]

[0027]

[0028]

[0029] In the above formula, u is the energy spectrum CT image to be reconstructed, p is the scanned and preprocessed projection data, A is the projection matrix, N(·) is the neural network prior, λ1 and λ2 are regularization parameters, α is the Lagrange multiplier, t = 1, 2, ..., T is the number of reconstruction iterations, k = 1, 2, ..., K is the number of convolution sparse iterations, * is the convolution operator, and M is {M n} n＝1，2，…，N The matrix form, F is {f n} n＝1，2，…，N In matrix form, C is less than F H F is the largest eigenvalue constant, I is the identity matrix, z is the dual variable of u, and h θ (·) represents the soft thresholding operation, and θ is the threshold.

[0030] Furthermore, the deep convolutional sparse representation reconstruction network in step 4 comprises four sub-modules: a parameter network module, a restoration neural network module, a convolutional sparse representation network module, and an image update module. The deep convolutional sparse representation reconstruction network is obtained by concatenating the restoration neural network module, the convolutional sparse representation network module, and the image update module T times. The parameter network module is used to learn the hyperparameters {α, β, λ1, λ2} in the reconstruction. The network structure consists of two consecutive convolutional layers and a Softplus operation, with a 1×1 kernel size. The input is the initial reconstructed image. The restoration neural network module is the neural network prior N(u), used to solve equation (3) and return z. t The value of . The convolutional sparse representation network module is used to learn the sparse representation prior S(u), which is used to solve equation (2) and return . The network structure is as follows: first, a convolutional layer; then, a skip summation connection followed by K ReLU operations and a convolutional layer; next, a convolutional layer, a ReLU operation, and finally a convolutional layer. All convolutional kernels are 3×3, and all have 32 channels. The image update module calculates the reconstructed energy spectrum CT image to solve equation (1) and returns u. t+1 The value of is expressed as follows:

[0031]

[0032] In the above formula, u is the energy spectrum CT image to be reconstructed, p is the scanned and preprocessed projection data, and A is the projection matrix. T Let λ1 be the back projection matrix, α be the regularization parameter, α be the Lagrange multiplier, * be the convolution operator, t = 1 - 2, ..., T be the number of reconstruction iterations, k = 1, 2, ..., K be the number of convolution sparse iterations, * be the convolution operator, and M be {M n} n＝1，2，…，N The matrix form, F is {f n} n＝1，2，…，N The matrix form of u is given, where z is the dual variable of u, and I is a vector consisting entirely of 1s. The initial reconstructed image is obtained by the FBP reconstruction algorithm.

[0033] Furthermore, in step 4, the loss function L for reconstructing the network during training is defined. dcsr for:

[0034] L dcsr =MSE(u T u s )+ηMSE(P R (u T ), P R (u s ))

[0035] In the above formula, u s For high-quality energy spectrum CT images and u T To output the reconstructed energy spectrum CT image, MSE(g) is the mean square error calculation function, P R (g) is the perception calculation function, and η is the loss weight.

[0036] Furthermore, step 4 iteratively updates the network model parameters using a mini-batch stochastic gradient descent algorithm. Training stops when the change in the loss function value before and after the training cycle is within 5%, thus obtaining the network model parameters. In step 5, the trained network is used to implement the test set energy spectrum CT projection data p. t The reconstruction process is as follows: the projection data in the test set are fed into the trained deep convolutional sparse representation reconstruction network one by one, and the network is used to calculate and output the reconstructed high-quality energy spectrum CT image.

[0037] 3. Beneficial effects

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] (1) The present invention provides a spectral CT imaging method based on deep convolutional sparse representation reconstruction network. Starting from convolutional sparse coding in the field of sparse learning, and combining the idea of ​​deep neural network, a joint constraint model of sparse representation and neural network restoration is established, which can realize the constraint of different feature information in the reconstructed image and improve the quality of the reconstructed image.

[0040] (2) The energy spectrum CT imaging method based on a deep convolutional sparse representation reconstruction network of the present invention adopts the idea of ​​deep learning to expand and optimize the statistical iterative reconstruction model and map each part to the network process. It is a learnable reconstruction that realizes the iterative solution process by using an end-to-end training method. In addition, the parameter network in the present invention can realize the adaptive selection of each hyperparameter during the reconstruction process.

[0041] (3) The present invention provides a spectral CT imaging method based on a deep convolutional sparse representation reconstruction network. The relevant experimental results verify that in the reconstruction of spectral CT images of different energy ranges under photon counting scanning, the method of the present invention has better visual effects and contrast than the reconstruction method based on tensor kernel norm and transform Lp norm (referred to as Tensor-Lp reconstruction) and the spectral CT reconstruction method based solely on deep learning (referred to as DSIR reconstruction). It has achieved satisfactory results in artifact suppression and preservation of anatomical tissue boundaries. It has great advantages in clinical screening, diagnosis and treatment applications, and can improve examination efficiency and reduce radiation damage to patients. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the overall process of an energy spectrum CT imaging method based on a deep convolutional sparse representation reconstruction network according to an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of a deep convolutional sparse representation reconstruction network according to an embodiment of the present invention;

[0044] Figure 3 This is an example image of a simulated abdominal CT scan in the embodiment;

[0045] Figure 4 The following are the energy spectrum CT reconstruction results of the 80 keV energy range simulation data in the embodiments (a: reference image; b: Tensor-Lp reconstruction result; c: DSIR reconstruction result; d: DCSR-Net reconstruction result).

[0046] Figure 5 The following are the energy spectrum CT reconstruction results of the 100 keV energy range simulation data in the embodiments (a: reference image; b: Tensor-Lp reconstruction result; c: DSIR reconstruction result; d: DCSR-Net reconstruction result).

[0047] Figure 6 The following are the energy spectrum CT reconstruction results of the 120 keV energy range simulation data in the embodiments (a: reference image; b: Tensor-Lp reconstruction result; c: DSIR reconstruction result; d: DCSR-Net reconstruction result).

[0048] Figure 7 The following are error maps of the energy spectrum CT reconstruction results of the 100 keV energy range simulation data in the examples (a: Tensor-Lp reconstruction error map; b: DSIR reconstruction error map; c: DCSR-Net reconstruction error map);

[0049] Figure 8 The example shows the profile curve of the region of interest in the energy spectrum CT reconstruction result of the 100 keV energy range simulation data in the embodiment. Detailed Implementation

[0050] This invention proposes a method for spectral CT imaging based on a deep convolutional sparse representation reconstruction network, called Deep Convolutional Sparse Representation Reconstruction Network (DCSR-Net). Within a statistical iterative reconstruction framework, it leverages the advantages of combining neural network priors and sparse representation priors to establish a learnable, adaptive reconstruction network. This network reduces noise during spectral CT reconstruction, improving image reconstruction quality and efficiency. The core steps include: first, constructing neural network priors and sparse representation priors; then, integrating the priors into the reconstruction model and decomposing and solving the model; next, networking the decomposed models to form the deep convolutional sparse representation reconstruction network and training it; finally, the user can input spectral CT projection data into the trained deep convolutional sparse representation reconstruction network, which outputs the reconstructed spectral CT image. This invention addresses the problem of poor image reconstruction quality in spectral CT images by combining the advantages of neural network priors and sparse representation priors to establish a reconstruction network with learnable adaptive parameters. This network can reduce noise in the spectral CT reconstruction process and improve image reconstruction quality and efficiency.

[0051] The inventors of this application have been committed to the research of CT image reconstruction and processing technology and have achieved certain results. For example, application No. 202210765963.5 discloses a low-dose CT reconstruction method based on residual domain iterative optimization network. This application establishes a multi-objective optimization function for low-dose CT reconstruction in the form of image domain and projection domain residuals. The residual domain update part is implemented by convolutional sparse coding network, which belongs to the image reconstruction method. However, this invention focuses on multi-channel spectral CT reconstruction. Its input is spectral CT projection data. It combines the advantages of sparse prior and network prior to construct a reconstruction network. At the same time, it can also realize parameter adaptive learning and realize end-to-end reconstruction based on spectral CT projection data. In the process of spectral CT reconstruction, it has the advantages of high reconstructed image quality and parameter adaptiveness.

[0052] The present invention will be further described below with reference to specific embodiments.

[0053] Example

[0054] like Figure 1 As shown in the figure, the specific steps of the energy spectrum CT imaging method based on deep convolutional sparse representation reconstruction network in this embodiment are as follows:

[0055] Step 1: Construct the training dataset and the test set.

[0056] Specifically, the training set construction method in step 1 is as follows: High-quality spectral CT images are obtained by collecting multi-channel energy-dispersive CT images and processing them. s High-quality energy-spectral CT images were reprojected and noise was added to obtain multi-channel energy-spectral CT projection data p that had noise intensity close to that of the test set data. s , forming training data pairs {p s u s The test set was constructed by collecting multi-channel energy spectral CT projection data. t This is used to test the algorithm.

[0057] Step 2: Design and construct the neural network prior N(u) and sparse representation prior S(u) for energy spectrum CT reconstruction;

[0058] Specifically, the neural network prior N(u) in step 2 is a three-layer lightweight residual network, including two consecutive convolutions followed by ReLU operations, one convolution operation, and a skip summation connection. The sparse representation prior S(u) adopts a convolutional sparse coding model, which can be expressed as:

[0059]

[0060] In the above formula, u is the energy spectrum CT image to be reconstructed, * is the convolution operator, and {f n} n＝1，2，…，N For the convolution kernel group, {M n} n＝1，2，…，N For each convolutional kernel, there are different feature map sets, and β is the regularization parameter.

[0061] Step 3: Incorporate the prior knowledge into the reconstruction model and decompose and solve it to obtain the decomposed energy spectrum CT reconstruction formula;

[0062] Specifically, by incorporating the neural network prior N(u) and the sparse representation prior S(u) into the iterative reconstruction model, we can obtain:

[0063]

[0064] In the above equation, u is the spectral CT image to be reconstructed, p is the scanned and preprocessed projection data, A is the projection matrix, N(u) is the neural network prior, S(u) is the sparse representation prior, and λ1 and λ2 are regularization parameters. Next, the reconstruction model is decomposed and solved. The decomposed spectral CT reconstruction equation is:

[0065]

[0066]

[0067]

[0068] In the above formula, u is the energy spectrum CT image to be reconstructed, p is the scanned and preprocessed projection data, A is the projection matrix, N(·) is the neural network prior, λ1 and λ2 are regularization parameters, α is the Lagrange multiplier, t = 1, 2, ..., T is the number of reconstruction iterations, k = 1, 2, ..., K is the number of convolution sparse iterations, * is the convolution operator, and M is {M n} n＝1，2，…，N The matrix form, F is {f n} n＝1，2，…，N In matrix form, C is less than F H F is the largest eigenvalue constant, I is the identity matrix, z is the dual variable of u, and h θ (·) represents the soft thresholding operation, and θ is the threshold.

[0069] Step 4: Network the decomposed energy spectrum CT reconstruction to form a deep convolutional sparse representation reconstruction network, and train the network using the training dataset to obtain the trained network parameters.

[0070] Specifically, the decomposed spectral CT reconstruction is networked, and the resulting deep convolutional sparse representation reconstruction network includes four sub-modules: a parameter network module, a restoration neural network module, a convolutional sparse representation network module, and an image update module. The deep convolutional sparse representation reconstruction network is obtained by T-fold cascading of the restoration neural network module, the convolutional sparse representation network module, and the image update module. The parameter network module is used to learn the hyperparameters {α, β, λ1, λ2} in the reconstruction. The network structure consists of two consecutive convolutional layers and a Softplus operation, with a 1×1 kernel size and the initial reconstruction image as input. The restoration neural network module is the neural network prior N(u), used to solve equation (3) and return z. t The value of . The convolutional sparse representation network module is used to learn the sparse representation prior S(u), which is used to solve equation (2) and return . The network structure is as follows: first, a convolutional layer; then, a skip summation connection followed by K ReLU operations and a convolutional layer; next, a convolutional layer, a ReLU operation, and finally a convolutional layer. All convolutional kernels are 3×3, and all have 32 channels. The image update module calculates the reconstructed energy spectrum CT image to solve equation (1) and returns u. t+1 The value of is expressed as follows:

[0071]

[0072] In the above formula, u is the energy spectrum CT image to be reconstructed, P is the scanned and preprocessed projection data, and A is the projection matrix. T Let λ1 be the back projection matrix, α be the regularization parameter, α be the Lagrange multiplier, * be the convolution operator, t = 1, 2, ..., T be the number of reconstruction iterations, k = 1, 2, ..., K be the number of convolution sparse iterations, and M be {M n} n＝1，2，…，N The matrix form, F is {f n} n＝1，2，…，N The matrix form of u is given, where z is the dual variable of u, and I is a vector of all 1s. The initial reconstructed image is obtained by the FBP reconstruction algorithm. The loss function L during the training of the depthwise convolutional sparse representation reconstruction network is defined as follows: dcsr for:

[0073] L dcsr =MSE(u T u s )+ηMSE(P R (u T ), P R (u s ))

[0074] In the above formula, u s For high-quality energy spectrum CT images and u T To output the reconstructed energy spectrum CT image, MSE(g) is the mean square error calculation function, P R (g) is the perceptuality calculation function, and η is the loss weight. During training, the network model parameters are iteratively updated using the mini-batch stochastic gradient descent algorithm. Training stops when the change in the loss function value before and after the training cycle is within 5%, and the network model parameters are obtained.

[0075] Step 5: Input the energy spectrum CT projection data from the test set into the trained deep convolutional sparse representation reconstruction network, output the reconstructed energy spectrum CT image, and verify the reconstruction performance.

[0076] Specifically, the trained network is used to implement the test collection of spectral CT projection data p tThe reconstruction process is as follows: the projection data in the test set are fed into the trained deep convolutional sparse representation reconstruction network one by one, and the network is used to calculate and output the reconstructed high-quality energy spectrum CT image.

[0077] To verify the feasibility of the method of the present invention, in this embodiment, high-quality CT image data of 6 patients, approximately 3,000 CT images, were collected. All collected CT image data came from the same device (GE Discovery dual-energy CT750HD). Virtual multi-channel spectral CT images (energy range of 60 keV to 120 keV, each segment with 20 keV energy, denoted as 80 keV, 100 keV, and 120 keV) were obtained using simulation software. After processing with a discriminative feature representation algorithm, these images were used as high-quality spectral CT image data. s Projection data was obtained by reprojecting multi-channel energy-spectral CT images at a projection angle of 1200 revolutions per circle. There were 736 detectors, each with a physical size of 1.28 mm. The distances from the line source to the object center and the detector center were 595 mm and 1086 mm, respectively. Additionally, Poisson noise and electronic noise were added to the projection data (the photon intensity in Poisson noise was 10). 5 With an electronic noise variance of 5), energy spectrum CT projection data were obtained, and 95% of them were selected as training data, i.e., {p s u s The remainder is used for testing. During network training, the loss function weight parameter η is set to 6.5, the number of convolution sparse iterations K is 5, and the number of reconstruction iterations T is 10.

[0078] Reconstruction effect evaluation

[0079] In the attached figures, all CT images show attenuation coefficient windows [0.16, 0.23]. Figures 4-6 The simulation results of different methods are presented under simulated data at energy ranges of 80keV, 100keV, and 120keV. Columns one through four show the reference image, Tensor-Lp reconstruction result, DSIR reconstruction result, and DCSR-Net reconstruction result, respectively. By comparing the visual effects of the reconstructed spectral CT images, it can be seen that the image reconstructed by the DCSR-Net network of this invention is superior to the Tensor-Lp and DSIR reconstruction methods. Common artifacts and noise in spectral scanning are well suppressed and removed, and the reconstructed image has high detail resolution. The Tensor-Lp reconstructed image shows good noise removal and some restoration of edge structures, but it cannot overcome large sparse angle artifacts. The DSIR method reconstructs tissue edge structures to a certain extent, but the resulting image is too smooth, and the details in the image are blurred.

[0080] Figure 7 Error maps of spectral CT reconstruction results from simulated data in the 100 keV energy range are shown. From left to right, they represent the error maps of reconstruction results using Tensor-Lp, DSIR, and DCSR-Net methods. By comparing the error maps, the image reconstructed by the method of this invention has less residual noise, indicating that the reconstruction process eliminates most artifacts, resulting in a clear structure that is very close to the reference image. Figure 8 For along Figure 5 The white lines in the diagram represent profile curves showing the intensity of different attenuation coefficients. By comparing the profile curves of the selected region of interest, it can be found that the CT values ​​of the reconstructed image are closer to those of the reference image, with the smallest intensity deviation.

[0081] High-quality energy-spectral CT images were used as reference images for quantitative evaluation. The Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) of the reconstructed energy-spectral CT images were calculated using different methods to quantitatively analyze the performance of different algorithms in this series. The results are shown in Table 1. Table 1 shows that, at different energy ranges, the PSNR and SSIM of the images reconstructed by the DCSR-Net network of this invention are higher than those of the comparative methods, and the values ​​are even higher at the 120 keV energy range.

[0082] Table 1

[0083]

[0084] Therefore, in summary, the method of the present invention can effectively suppress artifacts in sparse angle CT images and obtain higher quality CT images, and has certain application prospects.

Claims

1. A spectral CT imaging method based on a deep convolutional sparse representation reconstruction network, characterized in that, Includes the following steps: Step 1: Construct the training dataset and test set; Step 2: Design and construct neural network priors for energy spectrum CT reconstruction Sparse representation prior ; Step 3: Incorporate the prior knowledge into the reconstruction model and decompose and solve it to obtain the decomposed energy spectrum CT reconstruction formula; Specifically, the reconstruction model incorporating prior knowledge is as follows: In the above formula The energy spectrum CT image to be reconstructed. For the scanned and preprocessed projection data, A For the projection matrix, As a priori for neural networks, To represent sparsity priors and For regularization parameters; The decomposed energy spectrum CT reconstruction formula is as follows: In the above formula The energy spectrum CT image to be reconstructed. For the scanned and preprocessed projection data, A Let be the projection matrix. As a priori for neural networks, and For regularization parameters, For Lagrange multipliers, To reconstruct the number of iterations, The number of sparse convolution iterations. For convolution operators, for In matrix form, for In matrix form, , , Less than The largest eigenvalue constant, I Let be the identity matrix, and z be the dual variable of u. This is a soft threshold operation. For the threshold; Step 4: Network the decomposed energy spectrum CT reconstruction to form a deep convolutional sparse representation reconstruction network, and train the network using the training dataset to obtain the trained network parameters. Step 5: Input the energy spectrum CT projection data from the test set into the trained deep convolutional sparse representation reconstruction network, output the reconstructed energy spectrum CT image, and verify the reconstruction performance.

2. The energy spectrum CT imaging method based on a deep convolutional sparse representation reconstruction network according to claim 1, characterized in that: The training dataset in step 1 is constructed by collecting multi-channel energy-spectral CT images and processing them to obtain high-quality energy-spectral CT images. ; High-quality energy-spectral CT images were reprojected and noise was added to obtain multi-channel energy-spectral CT projection data with noise intensity close to that of the test set data. To form training data pairs The test set was constructed by collecting multi-channel energy spectral CT projection data. This is used to test the algorithm.

3. The energy spectrum CT imaging method based on a deep convolutional sparse representation reconstruction network according to claim 1, characterized in that: Neural network priors in step 2 It is a three-layer lightweight residual network consisting of two consecutive convolutions with ReLU operations, one convolution operation, and a skip summation connection.

4. The energy spectrum CT imaging method based on a deep convolutional sparse representation reconstruction network according to claim 1, characterized in that: In step 2, sparse representation priors Using a convolutional sparse coding model, it can be represented as: In the above formula The energy spectrum CT image to be reconstructed. For convolution operators, For convolution kernel groups, These are feature map sets corresponding to different convolution kernels. This is the regularization parameter.

5. The energy spectrum CT imaging method based on a deep convolutional sparse representation reconstruction network according to claim 1, characterized in that: Step 4 involves a deep convolutional sparse representation reconstruction network comprising four sub-modules: a parametric network module, a restoration neural network module, a convolutional sparse representation network module, and an image update module. Through the restoration neural network module, the convolutional sparse representation network module, and the image update module... T The secondary concatenation yields a deep convolutional sparse representation reconstruction network; the parameter network module is used to learn the hyperparameters in the reconstruction. The network structure consists of two consecutive convolutional layers and one Softplus operation, with a 1×1 kernel size. The input is the initial reconstructed image. The restoration neural network module is used to learn the prior information of the neural network. Used to solve equation (3) and return The value; the convolutional sparse representation network module is used to learn the sparse representation prior. Used to solve equation (2) and return The value, the network structure order is first a convolutional layer, then a skip summation connection, followed by... K The sequence consists of one ReLU operation and one convolutional layer, followed by one convolutional layer, one ReLU operation, and finally one convolutional layer. All convolutional kernels are 3×3 in size and have 32 channels. The image update module is used to calculate the reconstructed energy spectrum CT image to solve equation (1) and return... The value of is expressed as follows: In the above formula The energy spectrum CT image to be reconstructed. For the scanned and preprocessed projection data, A Let be the projection matrix. The back projection matrix, For regularization parameters, For Lagrange multipliers, For convolution operators, To reconstruct the number of iterations, The number of sparse convolution iterations. for In matrix form, for In matrix form, z for u dual variables, I The vector is composed entirely of 1s, and the initial reconstructed image is obtained by the FBP reconstruction algorithm.

6. The energy spectrum CT imaging method based on a deep convolutional sparse representation reconstruction network according to claim 1, characterized in that: In step 4, the loss function definition for reconstructing the network during training is given by depthwise convolutional sparse representation. for: In the above formula, For high-quality energy spectrum CT images, To output the reconstructed energy spectrum CT image, This is the function for calculating the mean square error. The function for calculating perception. This is the loss weight.

7. A spectral CT imaging method based on a deep convolutional sparse representation reconstruction network according to any one of claims 1-6, characterized in that: Step 4 iteratively updates the network model parameters using the mini-batch stochastic gradient descent algorithm. Training stops when the change in the loss function value before and after the training cycle is within 5%, thus obtaining the network model parameters. Step 5 uses the trained network to implement the test set energy spectrum CT projection data. The reconstruction process is as follows: the projection data in the test set are fed into the trained deep convolutional sparse representation reconstruction network one by one, and the network is used to calculate and output the reconstructed high-quality energy spectrum CT image.

Images