A metal artifact correction method for limited-angle CT based on cartoon texture dictionary network
Through the alternating iteration optimization algorithm of the cartoon texture dictionary network, the image is decomposed into cartoon and texture components, solving the problems of high-quality reconstruction and metal artifact correction of metal implants under limited angle CT, and achieving high-precision CT image reconstruction and pseudo-sham Movie correction, with model interpretability.
Patent Information
- Application Number
- CN202311747475.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-19
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2043-12-19
AI Technical Summary
The prior art is difficult to reconstruct CT images of metal implants at high quality at finite angle CT, and the existing metal artifact correction methods have problems of shadow artifacts and secondary artifacts, and it is impossible to solve the finite angle CT reconstruction and metal. Artifact correction problem.
Using the alternating iteration optimization algorithm of the cartoon texture dictionary network, the image is decomposed into cartoon components and texture components by simulating finite angle CT projection data, the optimization model is established and the alternating iteration is minimized. High-precision CT images are output, and cartoon and texture knowledge is used. Perform high-precision artifact correction.
High-precision CT image reconstruction and metal artifact correction under finite angle conditions are realized, which improves the reconstruction quality, and the designed cartoon-texture depth dictionary network has model interpretability.
Smart Images

Figure CN117726704B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of medical imaging, in particular to the problem of high-quality reconstruction of limited-angle CT under the presence of metal implants, and can be used in a CT imaging system. Background Art
[0002] Computed tomography (CT) is a non-invasive imaging technique that can visualize the internal structure of an object and has been widely used in medical diagnosis, safety inspection, and industrial non-destructive testing. In order to reduce radiation dose and reduce the risk of cancer and scanning time, limited angle CT (LACT) scanning is usually used in clinical practice to obtain measurement data, that is, the scanning angle is smaller than the normal scanning angle. However, due to incomplete projection data, the reconstruction results will produce serious shadow artifacts. For patients with metal implants, such as spinal and dental implants, due to the lack of additional projection data in the metal traces, new streak artifacts will be generated on the basis of shadow artifacts, thus affecting subsequent clinical diagnosis and treatment. Therefore, how to reconstruct high-precision CT images for patients with metal implants under limited angle conditions is an important issue in the field of medical imaging.
[0003] At present, many algorithms have been developed for LACT reconstruction. These can be divided into the following categories: analytical methods, model-based iterative methods, and deep learning-based methods. Filtered back projection (FBP) is a commonly used analytical method that can reconstruct high-quality images under the condition that the projection data is complete. However, when the projection data is incomplete, the reconstruction results will produce serious shadow artifacts. The model-based iterative method reconstructs the image from the undersampled projection data by iteratively minimizing the predefined image domain regularizer and data fidelity terms. This method is usually very time-consuming and the parameter adjustment is very cumbersome. According to the different domain knowledge used, deep learning-based methods can be mainly divided into three categories: image domain methods, sine domain methods, and dual domain methods. The image domain network regards the reconstruction of limited-angle images as an image restoration problem, directly removes artifacts without using sine domain data, but often ignores the consistency constraints of the measured data. The sine domain method uses the network to repair the missing view data in the sine map, and then obtains the reconstructed image through the FBP operation. However, due to the limited ability of the network to repair the projection domain data, processing the projection data alone often reduces the resolution of the reconstructed image. In order to solve the problems of single-domain methods, dual-domain methods have emerged, but the current dual-domain methods generally lack model interpretability. Existing CT metal artifact reduction (MAR) technologies are mainly divided into traditional metal artifact correction methods and deep learning-based metal artifact correction methods. Traditional methods mainly use various interpolation techniques to replace the metal-contaminated parts of the original projection data. However, this method can easily introduce secondary artifacts, causing serious structural deformation of the parts with severe metal artifacts in the reconstructed image. Similar to the LACT reconstruction algorithm based on deep learning, the MAR reconstruction algorithm based on deep learning is also mainly divided into three categories: image domain method, sine domain method and dual-domain method. At present, there is no corresponding method for simultaneously solving the limited angle CT reconstruction and metal artifact reduction (LAMAR) problems. It is not feasible to directly apply the previous LACT and MAR methods to LAMAR, which may produce unsatisfactory reconstruction quality. Therefore, based on the existing limited angle CT reconstruction and metal artifact correction methods, a high-quality reconstruction correction method for limited angle CT in the presence of metal implants is proposed. Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide a limited-angle CT metal artifact correction method based on a cartoon texture dictionary network, which is used to solve the limited-angle CT reconstruction problem in the presence of metal implants. The structure of the algorithm is optimized through alternating iterations to fully utilize cartoon and texture knowledge for high-precision artifact correction, and the iterative process is expanded into a network; the corresponding cartoon texture network is designed according to the different characteristics of the cartoon texture components, which can reconstruct the detailed structure of the image, improve the reconstruction quality, and achieve high-precision metal artifact correction. The cartoon-texture deep dictionary network designed by the present invention has model interpretability.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a limited angle CT metal artifact correction method based on a cartoon texture dictionary network, characterized in that it includes the following steps:
[0006] Step S1, simulating CT projection data affected by metal under limited angle sampling, that is, obtaining a sine diagram with limited angle sampling and containing metal traces, and performing a filtered back projection FBP operation on the projection data to obtain an initial estimate of the CT image to be reconstructed;
[0007] Step S2, decomposing the CT image to be reconstructed to obtain cartoon component images and texture component images;
[0008] Step S3, establishing an optimization model based on cartoon components and texture components;
[0009] Step S4, performing alternating iterative minimization on the constructed optimization model, and iteratively updating the cartoon component, texture component and CT image respectively;
[0010] Step S5: output the reconstructed high-precision CT image.
[0011] A further improvement of the technical solution of the present invention is that: in the step S1, limited-angle CT data acquisition is simulated, firstly, an X-ray beam is simulated to scan the object to be tested containing metal implants; secondly, the detector receives and obtains projection data; the size of the fully sampled projection data is specified to be 360×641, wherein 360 and 641 represent the number of projection views and detectors, respectively, i.e., 360 projection views are uniformly sampled between 0 and 360 degrees; and limited-angle projection data containing metal traces of a size of 120×641 is obtained by downsampling, i.e., the limited angle is 0 to 120 degrees.
[0012] A further improvement of the technical solution of the present invention is that: in step S1, the finite angle sampling sinusoidal graph s containing metal traces is obtained by downsampling lama , the CT image x containing a large number of artifacts is obtained by the FBP algorithm lama As the initial value of the CT image to be reconstructed, denoted as x lama =FBP(s lama ).
[0013] A further improvement of the technical solution of the present invention is that in step S2, the CT image x to be reconstructed is decomposed into a cartoon component u and a texture component v, that is, x=u+v, and at the same time, the texture component is modeled as follows using a comprehensive sparse representation: in, represents the convolution filter, is a sparse feature map.
[0014] A further improvement of the technical solution of the present invention is that the process of establishing the optimization model based on the cartoon component and the texture component in step S3 is:
[0015] First, establish the LAMAR optimization model, the formula is as follows:
[0016]
[0017] Among them, * represents the convolution operation, ||·||1 represents the l1 norm, λ1 and λ2 represent the trade-off parameters, and the tight frame W satisfies the tight property e is a unit vector, through the interleaved sparse feature map To construct the global representation coefficient vector
[0018] Secondly, design a local dictionary D L ∈R n×m The global convolution dictionary composed of N shifts The local dictionary D L The column vector of The optimization model based on cartoon component and texture component is:
[0019]
[0020] A further improvement of the technical solution of the present invention is that step S4 uses an alternating iteration method to solve the constructed optimization model:
[0021]
[0022]
[0023] At the tth iteration, u, D c and α are updated alternately as follows:
[0024] Update u, given the noisy image The image after tight frame filtering can be expressed as:
[0025]
[0026] Among them, soft(u,ε)=sign(u)max(|u|-ε,0) represents the soft threshold function, which uses a deep threshold network DTN to extract the noise image The adaptive learning threshold is f(·) represents the proposed DTN,
[0027] Update α,D c ,Mode <4> The coefficient α and the convolution dictionary D can be solved iteratively c There are two sub-problems to solve; using the neighboring gradient descent operator to update the representation coefficient α, the k+1th iteration is:
[0028]
[0029] Among them, k represents the threshold; L is the convolution dictionary D c (k) The maximum eigenvalue of k and B k To parameterize K neighboring gradient descent operators to construct a deep dictionary network, where A k and B k They are respectively learnable parsing and synthesis convolution operators:
[0030]
[0031] but:
[0032]
[0033] in Using DTN network from Medium adaptive learning threshold; Convolution dictionary D c t In the training phase, learning is done through back propagation.
[0034] Update x:
[0035] x t =u t +v t <9> .
[0036] A further improvement of the technical solution of the present invention is that the output result is constrained by using image loss, cartoon component loss and texture component loss, and the loss function is:
[0037]
[0038] Among them, x gt 、u gt 、v gtThey correspond to the real CT image and its cartoon component and texture component respectively; ω1 and ω2 are weight parameters for balancing different loss terms; mask is a binary non-metal mask, where the metal area is equal to 1 and the non-metal area is 0, and ⊙ represents element-by-element multiplication; x T 、u T and v T They represent the CT image, cartoon component and texture component reconstructed in the last stage respectively.
[0039] The further improvement of the technical solution of the present invention is that: in step S4, the optimization model is solved by using an alternating iterative algorithm, and the iterative process is expanded into a deep neural network, the deep expanded network consists of T stages, corresponding to T iterations of the iterative algorithm, each stage includes U-Net and V-Net, U-Net is composed of a tight standard framework with an adaptive threshold, and V-Net adopts a deep dictionary network architecture; wherein x 0 =x lama ,u t and v t Based on the formula <5> and <8> Updated by U-Net and V-Net, x t By <9> renew.
[0040] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:
[0041] 1. Based on the principle of limited angle CT imaging, the present invention designs a cartoon-texture deep dictionary network for LAMAR system. The existing methods of limited angle CT reconstruction and metal artifact correction have poor quality for limited angle CT reconstruction in the presence of metal implants. The present invention can reconstruct CT images with better accuracy and richer details from incomplete sinusoidal images by using cartoon and texture knowledge;
[0042] 2. The present invention is used to simultaneously solve the problems of limited-angle CT reconstruction and metal artifact correction. The structure of the alternating iterative optimization algorithm fully utilizes cartoon and texture knowledge to perform high-precision artifact correction, and the iterative process is expanded into a network; the corresponding cartoon texture network is designed according to the different characteristics of the cartoon texture component, which can reconstruct the detailed structure of the image, improve the reconstruction quality, and achieve high-precision metal artifact correction. The cartoon-texture deep dictionary network designed by the present invention has model interpretability. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is the overall flow chart of the present invention;
[0044] Figure 2 It is the overall architecture of the cartoon-texture deep dictionary network based on the present invention;
[0045] Figure 3Schematic diagram of the reconstruction results of LAMAR using different methods. DETAILED DESCRIPTION
[0046] The present invention is further described in detail below in conjunction with embodiments:
[0047] like Figure 1 As shown, a limited angle CT metal artifact correction method based on a cartoon texture dictionary network includes the following steps:
[0048] Step S1, simulating CT projection data affected by metal under limited angle sampling, simulating limited angle CT data acquisition, first simulating an X-ray beam to scan the object to be tested containing metal implants; secondly, the detector receives and obtains projection data; the size of the fully sampled projection data is specified to be 360×641, where 360 and 641 represent the number of projection views and detectors, respectively, that is, 360 projection views are uniformly sampled between 0 and 360 degrees; and limited angle projection data containing metal traces of a size of 120×641 is obtained by downsampling, that is, the limited angle is 0 to 360 degrees.
[0049] Obtain a sine graph with limited angle sampling and containing metal traces, and perform a filtered back projection FBP operation on the projection data, that is, downsampling to obtain a sine graph with limited angle sampling and containing metal traces s lama , the CT image x containing a large number of artifacts is obtained by the FBP algorithm lama As the initial value of the CT image to be reconstructed, denoted as x lama =FBP(s lama ) to obtain an initial estimate of the CT image to be reconstructed.
[0050] Step S2, decomposing the CT image to be reconstructed to obtain cartoon component images and texture component images;
[0051] The CT image x to be reconstructed is decomposed into a cartoon component u and a texture component v, that is, x = u + v. At the same time, the texture component is modeled using a comprehensive sparse representation as: in, represents the convolution filter, is a sparse feature map.
[0052] Step S3: Establish an optimization model based on cartoon components and texture components. The specific process is as follows:
[0053] First, establish the LAMAR optimization model, the formula is as follows:
[0054]
[0055] Among them, * represents the convolution operation, ||·||1 represents the l1 norm, λ1 and λ2 represent the trade-off parameters, and the tight frame W satisfies the tight property e is a unit vector. For convenience, we interleave the sparse feature map. To construct the global representation coefficient vector
[0056] Secondly, design a local dictionary D L ∈R n×m The global convolution dictionary composed of N shifts The local dictionary D L The column vector of The optimization model based on cartoon component and texture component is:
[0057]
[0058] Step S4: performing alternating iterative minimization on the constructed optimization model, and iteratively updating the cartoon component, texture component and CT image respectively.
[0059] Use the alternating iterative method to solve the constructed optimization model:
[0060]
[0061]
[0062] At the tth iteration, u, D c and α are updated alternately as follows:
[0063] Update u, given the noisy image The image after tight frame filtering can be expressed as:
[0064]
[0065] Among them, soft(u,ε)=sign(u)max(|u|-ε,0) represents the soft threshold function, which uses a deep threshold network DTN to extract the noise image The adaptive learning threshold is f(·) represents the proposed DTN,
[0066] Update α,D c ,Mode <4> The coefficient α and the convolution dictionary D can be solved iteratively c There are two sub-problems to solve; using the neighboring gradient descent operator to update the representation coefficient α, the k+1th iteration is:
[0067]
[0068] Among them, k represents the threshold; L is the convolution dictionary D c (k)For the convolutional dictionary update subproblem, ADMM generally involves multiple internal iterations, which makes the convolutional dictionary learning (CDL) process time-consuming. In order to effectively solve the CDL problem, A k and B k To parameterize K neighboring gradient descent operators to construct a deep dictionary network, where A k and B k They are respectively learnable parsing and synthesis convolution operators:
[0069]
[0070] but:
[0071]
[0072] in Using DTN network from Medium adaptive learning threshold; Convolution dictionary D c t In the training phase, learning is done through back propagation.
[0073] Update x:
[0074] x t =u t +v t <9> .
[0075] Image loss, cartoon component loss, and texture component loss are used to constrain the output results. The loss function is:
[0076]
[0077] Among them, x gt 、u gt 、v gt They correspond to the real CT image and its cartoon component and texture component respectively; ω1 and ω2 are weight parameters for balancing different loss terms; mask is a binary non-metal mask, where the metal area is equal to 1 and the non-metal area is 0, and ⊙ represents element-by-element multiplication; x T 、u T and v T They represent the CT image, cartoon component and texture component reconstructed in the last stage respectively.
[0078] The optimization model is solved using an alternating iterative algorithm, and the iterative process is expanded into a deep neural network. The deep expanded network consists of T stages, corresponding to T iterations of the iterative algorithm. Each stage contains U-Net and V-Net. Figure 2The network architecture diagram of the method for simultaneous limited angle CT reconstruction and high-precision correction of metal artifacts based on cartoon-texture deep dictionary network proposed in this invention is given. Among them, U-Net is composed of a tight standard frame with adaptive threshold, and V-Net adopts a deep dictionary network architecture; 0 =x lama ,u t and v t Based on the formula <5> and <8> Updated by U-Net and V-Net, x t By <9> renew.
[0079] Step S5: output the reconstructed high-precision CT image.
[0080] The training parameter settings of the above embodiment are: Adam optimizer is used to train the network, and the learning rate is set to 1×10 -4 , ω1=ω2=1, K=4, T=7; the network is implemented using the deep learning framework Pytorch, using NVIDIA RTX 3090GPU and i9-10850K CPU (3.60Hz).
[0081] In order to accurately analyze the effectiveness of the method proposed in the present invention, limited angle CT in the presence of metal implants was reconstructed based on the above method, and the currently advanced algorithms: FBP, FBPCovNet, DD-Net, ACDNet, InDuDoNet, and EPNet were compared; and the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) were used to qualitatively evaluate the method proposed in the present invention.
[0082] Figure 3 The images in the figure are, in order, the reference image, the FBP reconstructed image (network input image), the ACDNet reconstructed corrected image, the DD-Net reconstructed corrected image, the FBPCovNet reconstructed corrected image, the InDuDoNet reconstructed corrected image, the EPNet reconstructed corrected image and the reconstructed corrected image by the method of the present invention.
[0083] From the qualitative and quantitative results, it can be seen that the present invention can not only achieve high-quality reconstruction of limited angle CT images, but also effectively remove complex metal artifacts. In addition, it can be observed that the cartoon-texture depth dictionary method for limited angle CT reconstruction in the presence of metal implants is significantly better than the separate limited angle CT reconstruction and metal artifact correction methods.
[0084] Compared with the prior art, the present invention has the following advantages:
[0085] 1) Based on the principle of limited angle CT imaging, the present invention designs a cartoon-texture deep dictionary network for LAMAR system. The existing methods of limited angle CT reconstruction and metal artifact correction have poor quality for limited angle CT reconstruction in the presence of metal implants. The present invention can reconstruct CT images with better accuracy and richer details from incomplete sinusoidal images using cartoon and texture knowledge. It should be emphasized that the present invention is the first method for high-quality reconstruction of limited angle CT in the presence of metal implants.
[0086] 2) The present invention is used to simultaneously solve the problems of limited-angle CT reconstruction and metal artifact correction. The structure of the alternating iterative optimization algorithm fully utilizes cartoon and texture knowledge to perform high-precision artifact correction, and the iterative process is expanded into a network; the corresponding cartoon texture network is designed according to the different characteristics of the cartoon texture component, which can reconstruct the detailed structure of the image, improve the reconstruction quality, and achieve high-precision metal artifact correction. The cartoon-texture deep dictionary network designed by the present invention has model interpretability.
[0087] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. Obviously, for professionals in this field, after understanding the content and principle of the present invention, it is possible to make various modifications and changes in form and details without departing from the principle and structure of the present invention, but these modifications and changes based on the idea of the present invention are still within the scope of protection of the claims of the present invention.
Claims
1. A limited-angle CT metal artifact correction method based on a cartoon texture dictionary network, characterized by: The steps include: Step S1, simulating CT projection data affected by metal under limited angle sampling, that is, obtaining a sine diagram with limited angle sampling and containing metal traces, and performing a filtered back projection FBP operation on the projection data to obtain an initial estimate of the CT image to be reconstructed; Step S2, decomposing the CT image to be reconstructed to obtain cartoon component images and texture component images; Step S3, establishing an optimization model based on cartoon components and texture components; The process of establishing an optimization model based on cartoon components and texture components is: First, establish the LAMAR optimization model, the formula is as follows: Among them, * represents the convolution operation, ||·||1 represents the l1 norm, λ1 and λ2 represent the trade-off parameters, and the tight frame W satisfies the tight property e is a unit vector, through the interleaved sparse feature map To construct the global representation coefficient vector α=(α1,α1,…,α Nm )∈R Nm ; Secondly, design a local dictionary D L ∈R n×m The global convolution dictionary composed of N shifts The local dictionary D L The column vector of The optimization model based on cartoon component and texture component is: Step S4, performing alternating iterative minimization on the constructed optimization model, and iteratively updating the cartoon component, texture component and CT image respectively; Use the alternating iterative method to solve the constructed optimization model: At the tth iteration, u, D c and α are updated alternately as follows: Update u, given the noisy image The image after tight frame filtering can be expressed as: Among them, soft(u,ε)=sign(u)max(|u|-ε,0) represents the soft threshold function, which uses a deep threshold network DTN to extract the noise image The adaptive learning threshold is f(·) represents the proposed DTN, Update α,D c ,Mode <4> The coefficient α and the convolution dictionary D can be solved iteratively c There are two sub-problems to solve; using the neighboring gradient descent operator to update the representation coefficient α, the k+1th iteration is: Among them, k represents the threshold; L is the convolution dictionary D c (k) The maximum eigenvalue of k and B k To parameterize K neighboring gradient descent operators to construct a deep dictionary network, where A k and B k They are respectively learnable parsing and synthesis convolution operators: but: in Using DTN network from Medium adaptive learning threshold; Convolutional dictionary D c t In the training phase, learning is done through back propagation. Update x: x t =u t +v t <9>; The optimization model is solved by using an alternating iterative algorithm, and the iterative process is expanded into a deep neural network. The deep expanded network consists of T stages, corresponding to T iterations of the iterative algorithm. Each stage contains U-Net and V-Net. U-Net is composed of a tight standard framework with adaptive thresholds, and V-Net adopts a deep dictionary network architecture; where x 0 =x lama ,u t and v t Based on the formula <5> and <8> Updated by U-Net and V-Net, x t By <9> renew; Step S5: output the reconstructed high-precision CT image.
2. The method for correcting metal artifacts in limited-angle CT using a cartoon texture dictionary network according to claim 1, characterized in that: In the step S1, limited-angle CT data acquisition is simulated, firstly, an X-ray beam is simulated to scan the object to be tested containing metal implants; secondly, the detector receives and obtains projection data; the size of the fully sampled projection data is specified to be 360×641, wherein 360 and 641 represent the number of projection views and the number of detectors, respectively, i.e., 360 projection views are uniformly sampled between 0 and 360 degrees; and limited-angle projection data containing metal traces of a size of 120×641 is obtained by downsampling, i.e., the limited angle is 0 to 120 degrees.
3. The method for correcting metal artifacts in limited angle CT using a cartoon texture dictionary network according to claim 2, characterized in that: In step S1, the down-sampling is performed to obtain a finite angle sampling sinusoidal graph containing metal traces s lama , the CT image x containing a large number of artifacts is obtained by the FBP algorithm lama As the initial value of the CT image to be reconstructed, denoted as x lama =FBP(s lama ).
4. The method for correcting metal artifacts in limited-angle CT using a cartoon texture dictionary network according to claim 1, characterized in that: In step S2, the CT image x to be reconstructed is decomposed into a cartoon component u and a texture component v, that is, x=u+v. At the same time, the texture component is modeled using a comprehensive sparse representation as: in, represents the convolution filter, is a sparse feature map.
5. The method for correcting metal artifacts in limited-angle CT using a cartoon texture dictionary network according to claim 1, characterized in that: Image loss, cartoon component loss, and texture component loss are used to constrain the output results. The loss function is: Among them, x gt 、u gt 、v gt They correspond to the real CT image and its cartoon component and texture component respectively; ω1 and ω2 are weight parameters for balancing different loss terms; mask is a binary non-metal mask, where the metal area is equal to 1 and the non-metal area is 0, and ⊙ represents element-by-element multiplication; x T 、u T and v T They represent the CT image, cartoon component and texture component reconstructed in the last stage respectively.
